Upload folder using huggingface_hub

icml26/c39d243f-0c59-4f70-8ee6-d9e742174491/appendix_chunks.jsonl

@@ -0,0 +1,91 @@
+{"paper_id": "c39d243f-0c59-4f70-8ee6-d9e742174491", "chunk_id": "c39d243f-0c59-4f70-8ee6-d9e742174491:0092", "section": "A. VLM-RB", "page_start": 12, "page_end": 12, "type": "Text", "text": "632633634", "source": "marker_v2", "marker_block_id": "/page/11/Text/13"}
+{"paper_id": "c39d243f-0c59-4f70-8ee6-d9e742174491", "chunk_id": "c39d243f-0c59-4f70-8ee6-d9e742174491:0093", "section": "A. VLM-RB", "page_start": 12, "page_end": 12, "type": "Text", "text": "638639", "source": "marker_v2", "marker_block_id": "/page/11/Text/16"}
+{"paper_id": "c39d243f-0c59-4f70-8ee6-d9e742174491", "chunk_id": "c39d243f-0c59-4f70-8ee6-d9e742174491:0094", "section": "A. VLM-RB", "page_start": 12, "page_end": 12, "type": "Text", "text": "648649", "source": "marker_v2", "marker_block_id": "/page/11/Text/24"}
+{"paper_id": "c39d243f-0c59-4f70-8ee6-d9e742174491", "chunk_id": "c39d243f-0c59-4f70-8ee6-d9e742174491:0095", "section": "A. VLM-RB", "page_start": 12, "page_end": 12, "type": "Text", "text": "653654", "source": "marker_v2", "marker_block_id": "/page/11/Text/28"}
+{"paper_id": "c39d243f-0c59-4f70-8ee6-d9e742174491", "chunk_id": "c39d243f-0c59-4f70-8ee6-d9e742174491:0096", "section": "Algorithm 1 VLM-Prioritized Replay Buffer", "page_start": 12, "page_end": 12, "type": "Code", "text": "608 Require: prompt P; clip length L. 609 replay buffer \\mathcal{D}; clip buffer \\mathcal{C}; queues \\mathcal{Q}_{\\mathrm{in}} and \\mathcal{Q}_{\\mathrm{out}}. 610 1: function VLMWORKER 611 2: while true do (idxs, \\tau^O) \\leftarrow pop(Q_{in}) 3: 612 \\mathbf{p^{VLM}} = f_{\\mathrm{VLM}}(\\tau^O, \\mathsf{P}) \\in \\mathbb{R} 4: 613 (Equation 1) \\operatorname{push}(\\mathcal{Q}_{\\mathrm{out}},(\\operatorname{idxs},\\operatorname{\\mathbf{p}^{\\acute{\\mathbf{VLM}}}})) 5: 614 615 6: Init: launch VLMWORKER asynchronously. 7: while Training do 616 8: ⊳(1) Env step 617 9: Step env to obtain (s, a, r, s') and visual observation o = \\psi(s) 618 10: idx \\leftarrow INSERT(\\mathcal{D}, (s, a, r, s'; \\bar{p})) 619 \\mathcal{C} \\leftarrow \\mathcal{C} \\cup (\\text{idx}, o) 11: 620 12: ⊳ (2) Enqueue clips (streaming) if |\\mathcal{C}| = L or terminated or truncated then 621 13: idxs, \\tau^O \\leftarrow \\{(idx_i, o_i)\\}_{i=0}^{|C|-1} 622 14: \\operatorname{push}(\\mathcal{Q}_{\\operatorname{in}},(\\operatorname{idxs},\\tau^O)) 623 15: \\mathcal{C} \\leftarrow \\emptyset 16: 624 ▷ (3) Apply VLM scores (drain output queue) 17: 625 while \\mathcal{Q}_{\\mathrm{out}} not empty do 18: (\\texttt{idxs}, \\mathbf{p^{VLM}}) \\leftarrow \\mathbf{pop}(\\mathcal{Q}_{\\text{out}}) 626 19: 627 20: for all idx \\in idxs do \\texttt{SETPRIORITY}(\\mathcal{D}, \\texttt{idx}, \\mathbf{p^{VLM}}) 628 21: \\bar{p} \\leftarrow \\text{CMA}(\\bar{p}, \\mathbf{p^{VLM}}) 22: 629 \\triangleright (4) Sample and learn 23: 630 \\mathcal{B} \\sim \\lambda_t q_t^{\\hat{P}} + (1 - \\lambda_t) q_t^{U} (Equation 2) 631 25: UPDATELEARNER(\\mathcal{B}); update \\lambda_t", "source": "marker_v2", "marker_block_id": "/page/11/Code/3"}
+{"paper_id": "c39d243f-0c59-4f70-8ee6-d9e742174491", "chunk_id": "c39d243f-0c59-4f70-8ee6-d9e742174491:0097", "section": "Algorithm 1 VLM-Prioritized Replay Buffer", "page_start": 12, "page_end": 12, "type": "Text", "text": "In Algorithm 1, C denotes a temporary clip buffer, Q_{in} and Q_{out} are asynchronous communication queues, and CMA refers to the Cumulative Moving Average update rule for the global priority statistics.", "source": "marker_v2", "marker_block_id": "/page/11/Text/4"}
+{"paper_id": "c39d243f-0c59-4f70-8ee6-d9e742174491", "chunk_id": "c39d243f-0c59-4f70-8ee6-d9e742174491:0098", "section": "B. Environment Details", "page_start": 12, "page_end": 12, "type": "Text", "text": "In this section, we describe the environments used in our experiments.", "source": "marker_v2", "marker_block_id": "/page/11/Text/6"}
+{"paper_id": "c39d243f-0c59-4f70-8ee6-d9e742174491", "chunk_id": "c39d243f-0c59-4f70-8ee6-d9e742174491:0099", "section": "B. Environment Details", "page_start": 12, "page_end": 12, "type": "Text", "text": "MiniGrid / DoorKey. We use the <code>DoorKey</code> tasks from the MiniGrid suite (Chevalier-Boisvert et al., 2023) with sizes 8x8, 12x12, and 16x16. Each episode requires the agent to complete a sequence of subtasks: (i) navigate to and pick up the key, (ii) reach the locked door and open it using the <code>toggle</code> action (which is only possible while holding the key), and (iii) proceed to the goal tile. This structure enforces temporal dependencies and compositional reasoning. The reward function is sparse: the agent receives zero reward at all intermediate steps and a positive reward only upon reaching the goal, with the reward magnitude linearly decaying with the number of steps taken, following the standard MiniGrid protocol. Episodes terminate either upon successful completion or when the maximum step limit is reached.", "source": "marker_v2", "marker_block_id": "/page/11/Text/7"}
+{"paper_id": "c39d243f-0c59-4f70-8ee6-d9e742174491", "chunk_id": "c39d243f-0c59-4f70-8ee6-d9e742174491:0100", "section": "B. Environment Details", "page_start": 12, "page_end": 12, "type": "Text", "text": "State observations are provided in a symbolic format, where each s_t \\in \\{0, 1, \\ldots\\}^{N \\times N \\times 3} encodes the full grid. The three channels correspond to (i) object indices, (ii) color indices, and (iii) a state channel that includes both the door state and the agent's orientation. The action space is discrete with |\\mathcal{A}| = 5 , consisting of turn left, turn right, move forward, pick up, and toggle actions.", "source": "marker_v2", "marker_block_id": "/page/11/Text/8"}
+{"paper_id": "c39d243f-0c59-4f70-8ee6-d9e742174491", "chunk_id": "c39d243f-0c59-4f70-8ee6-d9e742174491:0101", "section": "B. Environment Details", "page_start": 12, "page_end": 12, "type": "Text", "text": "OGBench / Scene. We use the scene-play manipulation environment from OGBench (Park et al., 2024), which is a MuJoCo-based tabletop task involving a UR5e arm with a parallel gripper. The agent interacts with a set of objects: a cube, a sliding drawer, a sliding window, and two toggle buttons. Each button controls the lock state of either the drawer or the window, introducing dependencies that require the agent to execute explicit unlock, manipulate, and (re-)lock sequences to achieve many goals. Episodes are capped at a horizon of 750 steps. To facilitate learning, we employ potential-based reward", "source": "marker_v2", "marker_block_id": "/page/11/Text/9"}
+{"paper_id": "c39d243f-0c59-4f70-8ee6-d9e742174491", "chunk_id": "c39d243f-0c59-4f70-8ee6-d9e742174491:0102", "section": "B. Environment Details", "page_start": 13, "page_end": 13, "type": "Text", "text": "shaping with a per-step living cost, defined as:", "source": "marker_v2", "marker_block_id": "/page/12/Text/1"}
+{"paper_id": "c39d243f-0c59-4f70-8ee6-d9e742174491", "chunk_id": "c39d243f-0c59-4f70-8ee6-d9e742174491:0103", "section": "B. Environment Details", "page_start": 13, "page_end": 13, "type": "Equation", "text": "r_t = -1 \\ + \\ \\gamma \\, \\phi(s_{t+1}) - \\phi(s_t), \\qquad \\phi(s) = \\frac{1}{K} \\sum_{i=1}^K \\mathbb{I} \\left[ \\mathrm{subgoal}_i(s) \\right].", "source": "marker_v2", "marker_block_id": "/page/12/Equation/2"}
+{"paper_id": "c39d243f-0c59-4f70-8ee6-d9e742174491", "chunk_id": "c39d243f-0c59-4f70-8ee6-d9e742174491:0104", "section": "B. Environment Details", "page_start": 13, "page_end": 13, "type": "Text", "text": "Here, \\phi(s) denotes the fraction of satisfied subgoals, where each subgoal is a task-specific predicate over cube placement, button states, and the positions of the drawer and window. An episode terminates either when all subgoals are satisfied (\\phi(s_t) = 1) or when the maximum horizon is reached.", "source": "marker_v2", "marker_block_id": "/page/12/Text/3"}
+{"paper_id": "c39d243f-0c59-4f70-8ee6-d9e742174491", "chunk_id": "c39d243f-0c59-4f70-8ee6-d9e742174491:0105", "section": "B. Environment Details", "page_start": 13, "page_end": 13, "type": "Text", "text": "Observations are provided as full state vectors, s_t \\in \\mathbb{R}^{40} , and actions are specified as continuous end-effector controls, a_t \\in \\mathbb{R}^5 .", "source": "marker_v2", "marker_block_id": "/page/12/Text/4"}
+{"paper_id": "c39d243f-0c59-4f70-8ee6-d9e742174491", "chunk_id": "c39d243f-0c59-4f70-8ee6-d9e742174491:0106", "section": "B. Environment Details", "page_start": 13, "page_end": 13, "type": "Text", "text": "To address the exploration challenges inherent in these long-horizon tasks, we initialize the replay buffer with a lightweight warm-start of 10 demonstration episodes, which are held fixed across all methods. This small set of demonstrations provides minimal task-relevant coverage, while leaving the remainder of the training protocol unchanged.", "source": "marker_v2", "marker_block_id": "/page/12/Text/5"}
+{"paper_id": "c39d243f-0c59-4f70-8ee6-d9e742174491", "chunk_id": "c39d243f-0c59-4f70-8ee6-d9e742174491:0107", "section": "B. Environment Details", "page_start": 13, "page_end": 13, "type": "Text", "text": "We evaluate on a set of predefined Scene goals (tasks 3–5), which are designed to progressively increase the degree of temporal composition required for successful completion.", "source": "marker_v2", "marker_block_id": "/page/12/Text/6"}
+{"paper_id": "c39d243f-0c59-4f70-8ee6-d9e742174491", "chunk_id": "c39d243f-0c59-4f70-8ee6-d9e742174491:0108", "section": "B. Environment Details", "page_start": 13, "page_end": 13, "type": "ListGroup", "text": "Task 3 (rearrange-medium). The agent must move the cube to a specified tabletop location, open the drawer, close the window, and terminate with both the drawer and window unlocked. Notably, the window begins in a locked and open state, so the policy must first unlock it before closing, while simultaneously coordinating the manipulation of the drawer and the relocation of the cube. Task 4 (put-in-drawer). The agent is required to place the cube inside the drawer and terminate with the drawer closed and unlocked, while ensuring the window remains locked. This sequence involves unlocking the drawer, opening it, inserting the cube, and then closing the drawer. Task 5 ( rearrange-hard ). The agent must place the cube inside the (closed) drawer and leave the window open, while ensuring that both the drawer and window are locked at the end of the episode. Achieving this goal requires the agent to execute both unlock and relock sequences, coordinate drawer opening and closing, and correctly position the cube.", "source": "marker_v2", "marker_block_id": "/page/12/ListGroup/241"}
+{"paper_id": "c39d243f-0c59-4f70-8ee6-d9e742174491", "chunk_id": "c39d243f-0c59-4f70-8ee6-d9e742174491:0109", "section": "C. Prompts", "page_start": 14, "page_end": 14, "type": "Text", "text": "Task prompts used in our experiments. We use a binary \"success visible\" query with a strict output format.", "source": "marker_v2", "marker_block_id": "/page/13/Text/2"}
+{"paper_id": "c39d243f-0c59-4f70-8ee6-d9e742174491", "chunk_id": "c39d243f-0c59-4f70-8ee6-d9e742174491:0110", "section": "C. Prompts", "page_start": 14, "page_end": 14, "type": "ListGroup", "text": "1. MiniGrid/DoorKey: \"Does this clip contain a clear instance of goal satisfaction anywhere in it? If no visible success occurs, answer No. Do not guess. Output exactly Answer: Yes or Answer: No.\" 2. OGBench/Scene: \"Is there at least one clear instance of goal satisfaction in these frames? Look for contact + displacement consistent with the goal (lift off surface, place into receptacle, open/close articulation, move to target zone). Do not guess. If not visible, answer No. Output exactly Answer: Yes or Answer: No.\"", "source": "marker_v2", "marker_block_id": "/page/13/ListGroup/314"}
+{"paper_id": "c39d243f-0c59-4f70-8ee6-d9e742174491", "chunk_id": "c39d243f-0c59-4f70-8ee6-d9e742174491:0111", "section": "C. Prompts", "page_start": 14, "page_end": 14, "type": "Text", "text": "Meta-prompt for generating task prompts. To facilitate applying our procedure to new environments, we use the following meta-prompt to generate a task prompt from a task identifier and optional human context. The meta-prompt encourages generic, visually grounded success criteria and enforces a strict output schema.", "source": "marker_v2", "marker_block_id": "/page/13/Text/5"}
+{"paper_id": "c39d243f-0c59-4f70-8ee6-d9e742174491", "chunk_id": "c39d243f-0c59-4f70-8ee6-d9e742174491:0112", "section": "Meta-prompt template for generating task prompts", "page_start": 14, "page_end": 14, "type": "Text", "text": "User: You are an expert in Reinforcement Learning and Visual Language Models. I will provide a short clip from an agent rollout. Your job is to write a single text prompt that asks a VLM to output a binary judgment about whether goal satisfaction / competent task progress is clearly visible in the clip. Requirements:", "source": "marker_v2", "marker_block_id": "/page/13/Text/7"}
+{"paper_id": "c39d243f-0c59-4f70-8ee6-d9e742174491", "chunk_id": "c39d243f-0c59-4f70-8ee6-d9e742174491:0113", "section": "Meta-prompt template for generating task prompts", "page_start": 14, "page_end": 14, "type": "ListGroup", "text": "Keep the prompt environment-agnostic: rely on visible physics and outcomes rather than simulator-specific rules. Specify what to look for in broad categories when helpful (e.g., contact + displacement for manipulation; reaching a target region for navigation). Explicitly instruct: do not guess; if success is not clearly visible, answer No. End the prompt with the string: Output exactly Answer: Yes or Answer: No.", "source": "marker_v2", "marker_block_id": "/page/13/ListGroup/315"}
+{"paper_id": "c39d243f-0c59-4f70-8ee6-d9e742174491", "chunk_id": "c39d243f-0c59-4f70-8ee6-d9e742174491:0114", "section": "Examples (few-shot)", "page_start": 14, "page_end": 14, "type": "Text", "text": "• Task: Minigrid-doorkey", "source": "marker_v2", "marker_block_id": "/page/13/Text/13"}
+{"paper_id": "c39d243f-0c59-4f70-8ee6-d9e742174491", "chunk_id": "c39d243f-0c59-4f70-8ee6-d9e742174491:0115", "section": "Examples (few-shot)", "page_start": 14, "page_end": 14, "type": "Text", "text": "Prompt: \"Does this clip contain a clear instance of goal satisfaction anywhere in it? If no visible success occurs, answer No. Do not guess. Output exactly Answer: Yes or Answer: No.\"", "source": "marker_v2", "marker_block_id": "/page/13/Text/14"}
+{"paper_id": "c39d243f-0c59-4f70-8ee6-d9e742174491", "chunk_id": "c39d243f-0c59-4f70-8ee6-d9e742174491:0116", "section": "Examples (few-shot)", "page_start": 14, "page_end": 14, "type": "Text", "text": "• Task: OGBench-scene-task-4", "source": "marker_v2", "marker_block_id": "/page/13/Text/15"}
+{"paper_id": "c39d243f-0c59-4f70-8ee6-d9e742174491", "chunk_id": "c39d243f-0c59-4f70-8ee6-d9e742174491:0117", "section": "Examples (few-shot)", "page_start": 14, "page_end": 14, "type": "Text", "text": "Prompt: \"Is there at least one clear instance of goal satisfaction in these frames? Look for contact + displacement consistent with the goal (lift off surface, place into receptacle, open/close articulation, move to target zone). Do not guess. If not visible, answer No. Output exactly Answer: Yes or Answer: No.\"", "source": "marker_v2", "marker_block_id": "/page/13/Text/16"}
+{"paper_id": "c39d243f-0c59-4f70-8ee6-d9e742174491", "chunk_id": "c39d243f-0c59-4f70-8ee6-d9e742174491:0118", "section": "Current request", "page_start": 14, "page_end": 14, "type": "Text", "text": "Task Name: {{TASK ID}}", "source": "marker_v2", "marker_block_id": "/page/13/Text/18"}
+{"paper_id": "c39d243f-0c59-4f70-8ee6-d9e742174491", "chunk_id": "c39d243f-0c59-4f70-8ee6-d9e742174491:0119", "section": "Current request", "page_start": 14, "page_end": 14, "type": "Text", "text": "Optional Human Context: {{HUMAN CONTEXT}}", "source": "marker_v2", "marker_block_id": "/page/13/Text/19"}
+{"paper_id": "c39d243f-0c59-4f70-8ee6-d9e742174491", "chunk_id": "c39d243f-0c59-4f70-8ee6-d9e742174491:0120", "section": "Current request", "page_start": 15, "page_end": 15, "type": "Text", "text": "Example: prompt families for a new benchmark. As an illustration, we can apply the meta-prompt to a suite such as DeepMind Control by grouping tasks by their visible success patterns (locomotion / balance / interaction) and instantiating one generic prompt per group.", "source": "marker_v2", "marker_block_id": "/page/14/Text/2"}
+{"paper_id": "c39d243f-0c59-4f70-8ee6-d9e742174491", "chunk_id": "c39d243f-0c59-4f70-8ee6-d9e742174491:0121", "section": "Example: prompt families for DeepMind Control Suite", "page_start": 15, "page_end": 15, "type": "Text", "text": "User: Group DeepMind Control tasks by shared visible success patterns and propose one generic success-detection prompt per group (not task-specific). Each prompt must end with: \"Output exactly Answer: Yes or Answer: No.\"", "source": "marker_v2", "marker_block_id": "/page/14/Text/4"}
+{"paper_id": "c39d243f-0c59-4f70-8ee6-d9e742174491", "chunk_id": "c39d243f-0c59-4f70-8ee6-d9e742174491:0122", "section": "Example: prompt families for DeepMind Control Suite", "page_start": 15, "page_end": 15, "type": "Text", "text": "Model: We used Gemini 3 Pro to generate these prompts. Here is the partition into three semantic groups based on the nature of the movement, with a generalized \"vibe-check\" prompt for each.", "source": "marker_v2", "marker_block_id": "/page/14/Text/5"}
+{"paper_id": "c39d243f-0c59-4f70-8ee6-d9e742174491", "chunk_id": "c39d243f-0c59-4f70-8ee6-d9e742174491:0123", "section": "Group 1: Going Somewhere (Locomotion)", "page_start": 15, "page_end": 15, "type": "Text", "text": "The Semantic Goal: The agent must make progress across the environment. The VLM needs to distinguish between an agent that is struggling/falling (chaos) and one that is moving purposefully (control).", "source": "marker_v2", "marker_block_id": "/page/14/Text/7"}
+{"paper_id": "c39d243f-0c59-4f70-8ee6-d9e742174491", "chunk_id": "c39d243f-0c59-4f70-8ee6-d9e742174491:0124", "section": "Group 1: Going Somewhere (Locomotion)", "page_start": 15, "page_end": 15, "type": "Text", "text": "The Group: cheetah-run, dog-walk, dog-trot, dog-run, fish-swim, hopper-hop, humanoid-walk, humanoid-run, quadruped-walk, quadruped-run, walker-walk, walker-run.", "source": "marker_v2", "marker_block_id": "/page/14/Text/8"}
+{"paper_id": "c39d243f-0c59-4f70-8ee6-d9e742174491", "chunk_id": "c39d243f-0c59-4f70-8ee6-d9e742174491:0125", "section": "The Prompt", "page_start": 15, "page_end": 15, "type": "Text", "text": "\"Does the agent demonstrate controlled, purposeful movement through the environment? Look for the agent making progress in a specific direction (locomotion) rather than just flailing in place, falling over, or crashing. The movement should look stable and competent. If the agent falls, flips, or fails to move effectively, answer No. Output exactly Answer: Yes or Answer: No.\"", "source": "marker_v2", "marker_block_id": "/page/14/Text/10"}
+{"paper_id": "c39d243f-0c59-4f70-8ee6-d9e742174491", "chunk_id": "c39d243f-0c59-4f70-8ee6-d9e742174491:0126", "section": "Group 2: Defying Gravity (Stabilization)", "page_start": 15, "page_end": 15, "type": "Text", "text": "The Semantic Goal: The agent is fighting gravity. Success is defined by stillness or verticality. Whether it is a pole swinging up or a dog standing still, the \"vibe\" of success is reaching a stable, upright state and holding it.", "source": "marker_v2", "marker_block_id": "/page/14/Text/12"}
+{"paper_id": "c39d243f-0c59-4f70-8ee6-d9e742174491", "chunk_id": "c39d243f-0c59-4f70-8ee6-d9e742174491:0127", "section": "Group 2: Defying Gravity (Stabilization)", "page_start": 15, "page_end": 15, "type": "Text", "text": "The Group: acrobot-swingup, cartpole-balance, cartpole-balance sparse, cartpole-swingup, cartpole-swingup sparse, dog-stand, hopper-stand, humanoid-stand, walker-stand, pendulum-swingup.", "source": "marker_v2", "marker_block_id": "/page/14/Text/13"}
+{"paper_id": "c39d243f-0c59-4f70-8ee6-d9e742174491", "chunk_id": "c39d243f-0c59-4f70-8ee6-d9e742174491:0128", "section": "The Prompt", "page_start": 15, "page_end": 15, "type": "Text", "text": "\"Does the agent successfully achieve and maintain a stable, upright posture against gravity? Look for the agent balancing itself or an object vertically. Success means the agent (or the object it controls) is standing up or swinging up to a balanced position and staying there without collapsing, spinning out of control, or falling over. If it looks unstable or falls, answer No. Output exactly Answer: Yes or Answer: No.\"", "source": "marker_v2", "marker_block_id": "/page/14/Text/15"}
+{"paper_id": "c39d243f-0c59-4f70-8ee6-d9e742174491", "chunk_id": "c39d243f-0c59-4f70-8ee6-d9e742174491:0129", "section": "Group 3: Touching the Thing (Interaction)", "page_start": 15, "page_end": 15, "type": "Text", "text": "The Semantic Goal: This is about precise manipulation. The agent is not trying to go somewhere or just stand up; it is trying to force an interaction between its end-effector and a target object.", "source": "marker_v2", "marker_block_id": "/page/14/Text/17"}
+{"paper_id": "c39d243f-0c59-4f70-8ee6-d9e742174491", "chunk_id": "c39d243f-0c59-4f70-8ee6-d9e742174491:0130", "section": "Group 3: Touching the Thing (Interaction)", "page_start": 15, "page_end": 15, "type": "Text", "text": "The Group: ball in cup-catch, finger-spin, finger-turn easy, finger-turn hard, reacher-easy, reacher-hard.", "source": "marker_v2", "marker_block_id": "/page/14/Text/18"}
+{"paper_id": "c39d243f-0c59-4f70-8ee6-d9e742174491", "chunk_id": "c39d243f-0c59-4f70-8ee6-d9e742174491:0131", "section": "The Prompt", "page_start": 15, "page_end": 15, "type": "Text", "text": "\"Does the agent successfully interact with the target object in this scene? Look for the agent reaching a target, catching an object, or spinning a mechanism as intended. There should be clear, controlled contact or manipulation consistent with solving a task. If the agent misses the target, drops the object, or flails without making meaningful contact, answer No. Output exactly Answer: Yes or Answer: No.\"", "source": "marker_v2", "marker_block_id": "/page/14/Text/20"}
+{"paper_id": "c39d243f-0c59-4f70-8ee6-d9e742174491", "chunk_id": "c39d243f-0c59-4f70-8ee6-d9e742174491:0132", "section": "D. Baselines", "page_start": 16, "page_end": 16, "type": "Text", "text": "DoorKey presents a particularly challenging regime, characterized by extremely sparse rewards (identically zero until termination) and a long-horizon, sequential dependency structure (key to door to goal). In such settings, replay schemes relying on dense feedback, informative TD statistics, or a well-defined similarity metric are fundamentally misaligned with the task: they fail to prioritize transitions relevant to solving the task, particularly throughout the extended pre-success phase.", "source": "marker_v2", "marker_block_id": "/page/15/Text/2"}
+{"paper_id": "c39d243f-0c59-4f70-8ee6-d9e742174491", "chunk_id": "c39d243f-0c59-4f70-8ee6-d9e742174491:0133", "section": "D. Baselines", "page_start": 16, "page_end": 16, "type": "Text", "text": "Experience Replay Optimization (ERO) (Zha et al., 2019) ERO optimizes a learned rejection sampler using a REINFORCE-style update, where the scalar reward is defined as the change in evaluation return. In DoorKey, evaluation returns remain identically zero for a substantial portion of training (reflecting the absence of successes), which in turn implies rreplay ≈ 0 and results in near-zero gradients for the replay policy over extended periods. Furthermore, the replay-policy features used here, (r i , |δ i |, ti/Tmax), are weakly informative under sparse rewards: r i = 0 for almost all transitions and |δ i | is largely determined by bootstrap noise in the early stages of learning. As a result, the rejection policy effectively acts as an untrained stochastic filter for much of the training. When successes eventually occur, rreplay becomes highly variable, leading to unstable and non-stationary updates which can bias acceptance toward incidental correlates, such as late timesteps. The top-up rule, which fills with the highest-P i candidates when acceptance is low, further concentrates replay on a narrow subset of transitions. This reduces diversity and fails to provide a consistent signal for solving the task.", "source": "marker_v2", "marker_block_id": "/page/15/Text/3"}
+{"paper_id": "c39d243f-0c59-4f70-8ee6-d9e742174491", "chunk_id": "c39d243f-0c59-4f70-8ee6-d9e742174491:0134", "section": "D. Baselines", "page_start": 16, "page_end": 16, "type": "Text", "text": "ReLo (Reducible Loss Prioritization) (Sujit et al., 2023) ReLo prioritizes transitions based on the difference between the magnitudes of online and target TD residuals, pReLo = max(0, |δθ| − |δθ − |) +ϵ. In the sparse-reward DoorKey setting, both residuals are typically dominated by bootstrapping error rather than meaningful reward propagation throughout the prolonged pre-success phase. Moreover, the use of hard target updates periodically synchronizes the online and target networks, further diminishing any systematic separation between |δθ| and |δθ − |. Consequently, |δθ| − |δθ − | is often small or negative and is clipped to approximately ϵ, so the sampling distribution effectively reverts toward uniform, yet still incurs the variance and bias tradeoffs associated with prioritized replay.", "source": "marker_v2", "marker_block_id": "/page/15/Text/4"}
+{"paper_id": "c39d243f-0c59-4f70-8ee6-d9e742174491", "chunk_id": "c39d243f-0c59-4f70-8ee6-d9e742174491:0135", "section": "D. Baselines", "page_start": 16, "page_end": 16, "type": "Text", "text": "Attentive Experience Replay (AER) (Sun et al., 2020) AER selects samples closest (in a frozen embedding space) to the agent's current state, thereby inducing a strongly local replay distribution. In DoorKey this notion of locality is fundamentally misaligned with the task structure. Once trajectories reach later phases (near the door or goal), nearestneighbor replay disproportionately samples those regions and neglects earlier structural dependencies, such as key acquisition, thereby impeding temporal credit assignment across the long horizon. This effect is further amplified by the use of a randomly-MLinitialized frozen encoder: the resulting metric primarily captures superficial spatial proximity in the grid encoding, rather than functional or task-relevant similarity. As a result, the selection rule acts as a myopic location-based filter, rather than an attentive semantic sampler.", "source": "marker_v2", "marker_block_id": "/page/15/Text/5"}
+{"paper_id": "c39d243f-0c59-4f70-8ee6-d9e742174491", "chunk_id": "c39d243f-0c59-4f70-8ee6-d9e742174491:0136", "section": "E. Implementation Details", "page_start": 17, "page_end": 17, "type": "Text", "text": "To ensure a fair comparison, all baselines are implemented with the same backbone architecture, optimizer, and training schedule as VLM-RB . Unless otherwise specified, each method employs a PER -style replay buffer parameterized by exponents \\alpha and \\beta .", "source": "marker_v2", "marker_block_id": "/page/16/Text/2"}
+{"paper_id": "c39d243f-0c59-4f70-8ee6-d9e742174491", "chunk_id": "c39d243f-0c59-4f70-8ee6-d9e742174491:0137", "section": "E. Implementation Details", "page_start": 17, "page_end": 17, "type": "Text", "text": "Uniform Experience Replay In this setting, each transition i in the buffer is sampled with equal probability, p_i = 1/N_{curr} , where N_{curr} denotes the current buffer size.", "source": "marker_v2", "marker_block_id": "/page/16/Text/3"}
+{"paper_id": "c39d243f-0c59-4f70-8ee6-d9e742174491", "chunk_id": "c39d243f-0c59-4f70-8ee6-d9e742174491:0138", "section": "E. Implementation Details", "page_start": 17, "page_end": 17, "type": "Text", "text": "PER ( Prioritized Experience Replay ) Here, the priority assigned to each transition i is given by p_i = |\\delta_i| + \\epsilon , where \\delta_i denotes the most recent TD-error for that transition, and \\epsilon = 10^{-6} guarantees that every transition can be sampled. The probability of sampling transition i is then defined as", "source": "marker_v2", "marker_block_id": "/page/16/Text/4"}
+{"paper_id": "c39d243f-0c59-4f70-8ee6-d9e742174491", "chunk_id": "c39d243f-0c59-4f70-8ee6-d9e742174491:0139", "section": "E. Implementation Details", "page_start": 17, "page_end": 17, "type": "Equation", "text": "\\mathbf{p}_i = \\frac{p_i^{\\alpha}}{\\sum_k p_k^{\\alpha}},", "source": "marker_v2", "marker_block_id": "/page/16/Equation/5"}
+{"paper_id": "c39d243f-0c59-4f70-8ee6-d9e742174491", "chunk_id": "c39d243f-0c59-4f70-8ee6-d9e742174491:0140", "section": "E. Implementation Details", "page_start": 17, "page_end": 17, "type": "Text", "text": "with \\alpha determining the extent to which prioritization influences sampling. To account for the bias from prioritized sampling, importance-sampling (IS) weights are computed as", "source": "marker_v2", "marker_block_id": "/page/16/Text/6"}
+{"paper_id": "c39d243f-0c59-4f70-8ee6-d9e742174491", "chunk_id": "c39d243f-0c59-4f70-8ee6-d9e742174491:0141", "section": "E. Implementation Details", "page_start": 17, "page_end": 17, "type": "Equation", "text": "w_i = \\left(\\frac{1}{N} \\cdot \\frac{1}{p_i}\\right)^{\\beta}.", "source": "marker_v2", "marker_block_id": "/page/16/Equation/7"}
+{"paper_id": "c39d243f-0c59-4f70-8ee6-d9e742174491", "chunk_id": "c39d243f-0c59-4f70-8ee6-d9e742174491:0142", "section": "E. Implementation Details", "page_start": 17, "page_end": 17, "type": "Text", "text": "These weights are normalized by 1/\\max_i(w_i) prior to being used in the loss.", "source": "marker_v2", "marker_block_id": "/page/16/Text/8"}
+{"paper_id": "c39d243f-0c59-4f70-8ee6-d9e742174491", "chunk_id": "c39d243f-0c59-4f70-8ee6-d9e742174491:0143", "section": "E. Implementation Details", "page_start": 17, "page_end": 17, "type": "Text", "text": "Experience Replay Optimization (ERO) ERO (Zha et al., 2019) replaces standard replay sampling with a learned rejection policy applied to uniformly drawn candidates. The replay policy is parameterized as an MLP \\phi_{\\psi} with two hidden layers of 64 units each (ReLU activations, sigmoid output), mapping transition features to a retention probability.", "source": "marker_v2", "marker_block_id": "/page/16/Text/9"}
+{"paper_id": "c39d243f-0c59-4f70-8ee6-d9e742174491", "chunk_id": "c39d243f-0c59-4f70-8ee6-d9e742174491:0144", "section": "E. Implementation Details", "page_start": 17, "page_end": 17, "type": "Equation", "text": "p_i = \\phi_{\\psi}(r_i, |\\delta_i|, t_i/T_{\\text{max}}) \\in (0, 1),", "source": "marker_v2", "marker_block_id": "/page/16/Equation/10"}
+{"paper_id": "c39d243f-0c59-4f70-8ee6-d9e742174491", "chunk_id": "c39d243f-0c59-4f70-8ee6-d9e742174491:0145", "section": "E. Implementation Details", "page_start": 17, "page_end": 17, "type": "Equation", "text": "\\mathcal{L}_{ERO} = -r_{replay} \\sum_{i \\in \\mathcal{B}} \\log p_i,", "source": "marker_v2", "marker_block_id": "/page/16/Equation/12"}
+{"paper_id": "c39d243f-0c59-4f70-8ee6-d9e742174491", "chunk_id": "c39d243f-0c59-4f70-8ee6-d9e742174491:0146", "section": "E. Implementation Details", "page_start": 17, "page_end": 17, "type": "Text", "text": "where gradients are applied only to \\psi .", "source": "marker_v2", "marker_block_id": "/page/16/Text/13"}
+{"paper_id": "c39d243f-0c59-4f70-8ee6-d9e742174491", "chunk_id": "c39d243f-0c59-4f70-8ee6-d9e742174491:0147", "section": "E. Implementation Details", "page_start": 17, "page_end": 17, "type": "Text", "text": "ReLo (Reducible Loss Prioritization) ReLo (Sujit et al., 2023) assigns priority to each transition based on a reducible-loss score, which is computed from the online and target Q-networks. Given a transition (s,a,r,s'), the target value is defined as y=r+\\gamma\\max_{a'}Q_{\\theta^-}(s',a') . The online and target TD-errors are then \\delta_{\\text{online}}(s,a)=Q_{\\theta}(s,a)-y and \\delta_{\\text{target}}(s,a)=Q_{\\theta^-}(s,a)-y , respectively. The sampling priority is", "source": "marker_v2", "marker_block_id": "/page/16/Text/14"}
+{"paper_id": "c39d243f-0c59-4f70-8ee6-d9e742174491", "chunk_id": "c39d243f-0c59-4f70-8ee6-d9e742174491:0148", "section": "E. Implementation Details", "page_start": 17, "page_end": 17, "type": "Equation", "text": "p_{\\text{ReLo}}(s, a) = \\max(0, |\\delta_{\\text{online}}(s, a)| - |\\delta_{\\text{target}}(s, a)|) + \\epsilon, \\qquad \\epsilon = 10^{-6}.", "source": "marker_v2", "marker_block_id": "/page/16/Equation/15"}
+{"paper_id": "c39d243f-0c59-4f70-8ee6-d9e742174491", "chunk_id": "c39d243f-0c59-4f70-8ee6-d9e742174491:0149", "section": "E. Implementation Details", "page_start": 17, "page_end": 17, "type": "Text", "text": "Each new transition is initialized with the maximum priority value among all transitions currently in the buffer. The method sets \\alpha = 0.6 and initializes \\beta = 0.4 , annealing \\beta linearly to 1.0. Standard importance-sampling weights are applied.", "source": "marker_v2", "marker_block_id": "/page/16/Text/16"}
+{"paper_id": "c39d243f-0c59-4f70-8ee6-d9e742174491", "chunk_id": "c39d243f-0c59-4f70-8ee6-d9e742174491:0150", "section": "E. Implementation Details", "page_start": 18, "page_end": 18, "type": "Text", "text": "Attentive Experience Replay (AER) AER (Sun et al., 2020) selects transitions using an attentive sampling mechanism. A dedicated encoder \\phi is constructed with the same backbone architecture as the Q-network. This encoder is randomly initialized and remains fixed throughout training. At each training step t, a candidate pool of size N_{\\text{cand}} = \\lfloor \\lambda_t B \\rfloor is sampled uniformly from the buffer, where \\lambda_t decays linearly from \\lambda_0 = 4 to 1 over the course of training. If the candidate pool size N_{\\text{cand}} is less than or equal to the batch size B, the method reverts to the default sampling strategy. Given the current state s_{\\text{curr}} and a set of candidate states \\{s_i\\} , distances are computed in the frozen embedding space as d(s_{\\text{curr}}, s_i) = \\|\\phi(s_{\\text{curr}}) - \\phi(s_i)\\|_2^2 . The B candidates with the smallest distances are then selected deterministically to form the training batch. Following Sun et al. (2020), importance sampling is disabled.", "source": "marker_v2", "marker_block_id": "/page/17/Text/1"}
+{"paper_id": "c39d243f-0c59-4f70-8ee6-d9e742174491", "chunk_id": "c39d243f-0c59-4f70-8ee6-d9e742174491:0151", "section": "E.1. Architectural and Implementation Details", "page_start": 18, "page_end": 18, "type": "Text", "text": "MiniGrid agents (DQN / IQN). Both DQN and IQN are trained on symbolic Minigrid observations s_t \\in \\{0, \\dots\\}^{N \\times N \\times 3} , which encode object identities, colors, and state information such as door status and agent orientation. To process these inputs, we employ a shared encoder which first embeds each channel, then applies a residual dilated CNN with three blocks. The resulting features are aggregated using both average and max global pooling, yielding a 256-dimensional representation. This design aims to capture both local and global spatial structure in the environment. For DQN, the 256-dimensional feature vector is passed through a small MLP, which outputs scalar Q-values for each discrete action. In IQN, we replace the scalar output head with an implicit quantile head. Specifically, we embed sampled quantile fractions using cosine functions, project these embeddings to match the feature dimension, and combine them element-wise with the state features. The resulting representations are mapped to per-action quantile values. We optimize using the quantile Huber loss and select actions using either Double-DQN-style (Van Hasselt et al., 2016) or target-network-based strategies, as described in Appendix E. This approach allows IQN to model the full distribution over returns, rather than just the mean.", "source": "marker_v2", "marker_block_id": "/page/17/Text/3"}
+{"paper_id": "c39d243f-0c59-4f70-8ee6-d9e742174491", "chunk_id": "c39d243f-0c59-4f70-8ee6-d9e742174491:0152", "section": "E.1. Architectural and Implementation Details", "page_start": 18, "page_end": 18, "type": "Text", "text": "OGBench Scene agents (SAC / TD3). For OGBench/Scene, both SAC and TD3 operate on flattened state-based observations s_t \\in \\mathbb{R}^{40} . We adopt a unified architecture across tasks: the actor is a fully-connected network with three hidden layers of width 512 and ReLU activations. This network outputs either a tanh-squashed Gaussian policy (for SAC) or a deterministic tanh policy (for TD3), enabling flexible action selection in continuous spaces. The critic is implemented as an ensemble of N=10 Q-functions, each parameterized by a 3-layer MLP that receives the concatenated state and action as input and outputs a scalar value. For target computation, we randomly select M=2 ensemble members and aggregate their predictions using either a mean (in all main SAC/TD3 runs) or a min operator. This ensemble approach is intended to improve stability and reduce overestimation bias. For actor updates, we use the same Q-function ensemble, aggregating Q-values according to the actor reduction rule. In all reported experiments, we use the mean over ensemble members. This ensures consistency between actor and critic updates.", "source": "marker_v2", "marker_block_id": "/page/17/Text/4"}
+{"paper_id": "c39d243f-0c59-4f70-8ee6-d9e742174491", "chunk_id": "c39d243f-0c59-4f70-8ee6-d9e742174491:0153", "section": "E.1. Architectural and Implementation Details", "page_start": 18, "page_end": 18, "type": "Text", "text": "Replay buffer and prioritization. All methods are implemented with a two-branch replay ensemble: one prioritized branch and one uniform branch which share an underlying storage, implemented using ReplayBufferEnsemble (Bou et al., 2023) with mixture weights controlled by the sampling ratio schedule in Tables 2–3. For the PER baseline, we assign priorities using the standard PER formula p_i = |\\delta_i| + \\epsilon , where \\delta_i is the TD error. When importance-sampling corrections are enabled, we apply them as usual. New transitions are initially assigned a small default priority, and we update their TD-errors after each gradient step to ensure accurate prioritization. In VLM-RB, we augment each transition with both a TD-based metric and a VLM-based score. For MiniGrid/DoorKey, replay priorities are determined solely by the VLM score. In contrast, for OGBench/Scene, we set priorities as the product of the VLM score and the TD-based metric (see Section 3.3). This design allows us to tailor prioritization to the characteristics of each environment.", "source": "marker_v2", "marker_block_id": "/page/17/Text/5"}
+{"paper_id": "c39d243f-0c59-4f70-8ee6-d9e742174491", "chunk_id": "c39d243f-0c59-4f70-8ee6-d9e742174491:0154", "section": "E.1. Architectural and Implementation Details", "page_start": 18, "page_end": 18, "type": "Text", "text": "VLM worker. The VLM worker uses Perception-LM-1B (Cho et al., 2025) with a fixed prompt to detect binary success. For each transition, it processes a clip of L=32 frames, applying left padding if the clip is shorter. To compute a scalar priority, we sum the probabilities assigned to all \"Yes\" and \"No\" token variants in the first generated token's logits, and calculate their ratio to obtain a probability-style score. We hard-threshold these scores at 0.5, assigning 1 or 0 accordingly. The resulting priorities are asynchronously streamed back to update the replay buffer, ensuring the main learner is not blocked during this process.", "source": "marker_v2", "marker_block_id": "/page/17/Text/6"}
+{"paper_id": "c39d243f-0c59-4f70-8ee6-d9e742174491", "chunk_id": "c39d243f-0c59-4f70-8ee6-d9e742174491:0155", "section": "E.1. Architectural and Implementation Details", "page_start": 19, "page_end": 19, "type": "Caption", "text": "Table 2. Hyperparameters and implementation details for OGBench experiments. Shared parameters are listed once; divergent parameters are compared side-by-side.", "source": "marker_v2", "marker_block_id": "/page/18/Caption/1"}
+{"paper_id": "c39d243f-0c59-4f70-8ee6-d9e742174491", "chunk_id": "c39d243f-0c59-4f70-8ee6-d9e742174491:0156", "section": "E.1. Architectural and Implementation Details", "page_start": 19, "page_end": 19, "type": "Table", "text": "Hyperparameter TD3 SAC Network Architecture Hidden Dimensions [512, 512, 512] Q-Network Layer Norm False True Optimization Critic Learning Rate 3 × 10−4 Actor Learning Rate 3 × 10−4 1 × 10−3 Alpha Learning Rate – 3 × 10−4 Batch Size 256 Discount Factor (γ) 0.99 Target Smoothing (τ ) 0.005 Max Grad Norm 10.0 Learning Starts 10,000 steps Ensemble & Update Ratios Ensemble Size (N) 10 Subsampled Q-Networks (M) 2 Critic UTD Ratio 4 Actor UTD Ratio 2 Target Q Reduction min mean Actor Q Reduction min mean Data & Replay Buffer Replay Buffer Size 1 × 106 Expert Demos 10 Algorithm Specifics Policy Update Frequency 2 1 Exploration Noise (std) 0.1 – Target Policy Noise (std) 0.2 – Noise Clip 0.5 – Entropy Coeff. (α) – 1.0 (Initial) Target Entropy Scale – 0.5 Auto Entropy Tuning – True PER Baseline Settings PER Alpha (α) 0.7 PER Beta (β) 1.0 Importance Sampling False VLM-RB Settings VLM Model Facebook Perception-LM-1B Priority Mode Hard Threshold (> 0.5) Trajectory Length (L) 32 Frames Importance Sampling False Prioritized Sampling Ratio (λ0 and λmax) Annealed 0.0 → 0.5 Annealing Schedule (Tschedule) Linear over 500k steps", "source": "marker_v2", "marker_block_id": "/page/18/Table/2"}
+{"paper_id": "c39d243f-0c59-4f70-8ee6-d9e742174491", "chunk_id": "c39d243f-0c59-4f70-8ee6-d9e742174491:0157", "section": "E.1. Architectural and Implementation Details", "page_start": 20, "page_end": 20, "type": "Caption", "text": "Table 3. Hyperparameters and implementation details for MiniGrid experiments. Shared parameters are listed once; divergent parameters are compared side-by-side.", "source": "marker_v2", "marker_block_id": "/page/19/Caption/1"}
+{"paper_id": "c39d243f-0c59-4f70-8ee6-d9e742174491", "chunk_id": "c39d243f-0c59-4f70-8ee6-d9e742174491:0158", "section": "E.1. Architectural and Implementation Details", "page_start": 20, "page_end": 20, "type": "Table", "text": "Hyperparameter DQN IQN Optimization Learning Rate 4 × 10−5 Batch Size 128 Discount Factor (γ) 0.95 Target Update Frequency 1,000 steps Target Update Rate (τ ) 1.0 (Hard Update) Learning Starts 500 steps Train Frequency 4 steps Max Grad Norm 1.0 Exploration (Epsilon-Greedy) epsstart 1.0 epsend 0.05 Exploration Fraction 0.5 (First 50% of training) Algorithm Specifics Double DQN Enabled Noisy Nets – Disabled Num Quantiles (Policy) – 32 Num Quantiles (Train/Target) – 64 Num Cosine Basis Functions – 64 Huber Kappa (κ) – 1.0 PER Baseline Settings PER Alpha (α) 0.7 PER Beta (β) 1.0 Importance Sampling True Prioritized Sampling Ratio 1.0 (Always Prioritized) VLM-RB Settings VLM Model Facebook Perception-LM-1B Priority Mode Hard Threshold (> 0.5) Trajectory Length (L) 32 Frames Importance Sampling False Prioritized Sampling Ratio (λ0 and λmax) Annealed 0.0 → 0.5 Annealing Schedule (Tschedule) Linear over 500k steps", "source": "marker_v2", "marker_block_id": "/page/19/Table/2"}
+{"paper_id": "c39d243f-0c59-4f70-8ee6-d9e742174491", "chunk_id": "c39d243f-0c59-4f70-8ee6-d9e742174491:0159", "section": "F.1. Mixing Schedule", "page_start": 21, "page_end": 21, "type": "TableGroup", "text": "Table 4. Ablation of mixing schedule (λmax) on DoorKey-16x16. Performance denotes the highest Average Success Rate (M = 5 seeds). Sample Efficiency tracks the steps required to reach the baseline's best performance. Improvement λmax Base. Perf. (↑) Sample Eff. (↓) None UER PER N/A (0.00/0.24) N/A (0.00/0.80) N/A (1000K/916K) N/A (1000K/592K) PER +0.0% (0.80/0.80) 0.25 UER +233.3% (0.80/0.24) +40.6% (544K/916K) -0.3% (594K/592K) 0.50 UER +316.7% (1.00/0.24) +62.9% (340K/916K) PER +25.0% (1.00/0.80) +35.5% (382K/592K) 0.75 UER +266.7% (0.88/0.24) +57.0% (394K/916K) PER +10.0% (0.88/0.80) +17.2% (490K/592K) 1.00 UER +316.7% (1.00/0.24) +69.0% (284K/916K) PER +25.0% (1.00/0.80) +40.2% (354K/592K)", "source": "marker_v2", "marker_block_id": "/page/20/TableGroup/523"}
+{"paper_id": "c39d243f-0c59-4f70-8ee6-d9e742174491", "chunk_id": "c39d243f-0c59-4f70-8ee6-d9e742174491:0160", "section": "F.1. Mixing Schedule", "page_start": 21, "page_end": 21, "type": "Text", "text": "To understand the effect of the final mixing coefficient λmax, we conduct an ablation on the MiniGrid/DoorKey-16x16 task (Fig .6) . We fix Tschedule = 5 · 10 5 , ensuring that λ t is annealed linearly from 0 (corresponding to purely uniform sampling) to λmax over the first half of training (right panel). We sweep λmax ∈ {0.25, 0.5, 0.75, 1.0} and include a None baseline, which disables the schedule and relies entirely on VLM-prioritized sampling. We observe that larger λmax values (0.75, 1.0) reach 50% success marginally earlier than smaller values, but their final success rates remain lower within the fixed training budget. In contrast, λmax = 0.5 emerges as the most reliable choice in this environment: it consistently achieves 100% success within the budget and attains the highest performance at the 90% success threshold among all options considered. The smallest value (λmax = 0.25) results in slower learning, and the fully prioritized variant ( None ) fails to solve the task under our setup, with success rates remaining near zero. This suggests that maintaining a non-trivial fraction of uniform sampling is essential for effective learning in this setting. Based on these findings, we fix λmax = 0.5", "source": "marker_v2", "marker_block_id": "/page/20/Text/6"}
+{"paper_id": "c39d243f-0c59-4f70-8ee6-d9e742174491", "chunk_id": "c39d243f-0c59-4f70-8ee6-d9e742174491:0161", "section": "F.1. Mixing Schedule", "page_start": 21, "page_end": 21, "type": "Text", "text": "for all main experiments without further per-task tuning.", "source": "marker_v2", "marker_block_id": "/page/20/Text/16"}
+{"paper_id": "c39d243f-0c59-4f70-8ee6-d9e742174491", "chunk_id": "c39d243f-0c59-4f70-8ee6-d9e742174491:0162", "section": "F.1. Mixing Schedule", "page_start": 21, "page_end": 21, "type": "Caption", "text": "Figure 6. MiniGrid/DoorKey-16x16: λmax Ablation (5 seeds), \"None\" corresponds to no scheduling, i.e., only VLM-prioritized sampling. Left: performance for different λmax values. Right: λmax value evolution.", "source": "marker_v2", "marker_block_id": "/page/20/Caption/10"}
+{"paper_id": "c39d243f-0c59-4f70-8ee6-d9e742174491", "chunk_id": "c39d243f-0c59-4f70-8ee6-d9e742174491:0163", "section": "F.2. VLM Size", "page_start": 22, "page_end": 22, "type": "TableGroup", "text": "Table 5. Performance comparison of Perception-LM (Cho et al., 2025) models. Values in parentheses denote the relative percentage change compared to 1B. We use an Nvidia RTX 4090 GPU, and run 100 batches of 32 frames. Model Load (GiB) Peak (GiB) Time (s) FPS 1B 2.86 3.77 0.46 69.27 3B 6.56 (+130%) 8.16 (+116%) 0.77 (+66%) 41.75 (-40%) 8B 18.25 (+539%) 20.34 (+439%) 2.15 (+366%) 14.88 (-79%)", "source": "marker_v2", "marker_block_id": "/page/21/TableGroup/379"}
+{"paper_id": "c39d243f-0c59-4f70-8ee6-d9e742174491", "chunk_id": "c39d243f-0c59-4f70-8ee6-d9e742174491:0164", "section": "F.2. VLM Size", "page_start": 22, "page_end": 22, "type": "Text", "text": "We investigate how scaling the VLM affects both inference overhead and downstream RL performance. Table 5 quantifies the resource requirements of Perception-LM (Cho et al., 2025) variants for clip scoring (Nvidia RTX 4090 GPU; 100 batches of 32-frame clips). We observe that increasing the VLM size leads to a substantial increase in memory footprint and a corresponding reduction in throughput. Relative to the 1B model, the 3B variant increases peak memory by +116% and reduces FPS by 40%; the 8B variant increases peak memory by +439% and reduces FPS by 79%.", "source": "marker_v2", "marker_block_id": "/page/21/Text/4"}
+{"paper_id": "c39d243f-0c59-4f70-8ee6-d9e742174491", "chunk_id": "c39d243f-0c59-4f70-8ee6-d9e742174491:0165", "section": "F.2. VLM Size", "page_start": 22, "page_end": 22, "type": "Text", "text": "To assess the impact of VLM size on downstream performance, we compare the 1B, 3B, and 8B models on MiniGrid/DoorKey-16x16 (Fig. 7) . Notably, despite the increased inference cost of larger models, we do not observe consistent improvements in RL performance relative to the 1B configuration. These results suggest that in this setting, the clip-scoring signal saturates once the VLM is sufficiently reliable at separating task-relevant from irrelevant segments. Given this trade-off, we select the 1B Perception-LM as the default backbone for all main experiments.", "source": "marker_v2", "marker_block_id": "/page/21/Text/5"}
+{"paper_id": "c39d243f-0c59-4f70-8ee6-d9e742174491", "chunk_id": "c39d243f-0c59-4f70-8ee6-d9e742174491:0166", "section": "F.2. VLM Size", "page_start": 22, "page_end": 22, "type": "FigureGroup", "text": "Figure 7. MiniGrid/DoorKey-16x16: VLM Size Ablation (5 seeds)", "source": "marker_v2", "marker_block_id": "/page/21/FigureGroup/380"}
+{"paper_id": "c39d243f-0c59-4f70-8ee6-d9e742174491", "chunk_id": "c39d243f-0c59-4f70-8ee6-d9e742174491:0167", "section": "F.3. VLM Prior", "page_start": 22, "page_end": 22, "type": "Text", "text": "Motivated by the \"Modified Game\" paradigm of Dubey et al. (2018) , we probe whether VLM-RB leverages the VLM's pre-trained visual semantics to improve performance. To isolate the effect of visual semantics, we modify only the rendered frames provided to the VLM for scoring, while leaving both the underlying MDP and the agent's observations unchanged. If VLM-RB relies on semantic cues such as identifying keys, doors, or goal-relevant interactions, we expect its performance gains to diminish when these cues are distorted or removed.", "source": "marker_v2", "marker_block_id": "/page/21/Text/9"}
+{"paper_id": "c39d243f-0c59-4f70-8ee6-d9e742174491", "chunk_id": "c39d243f-0c59-4f70-8ee6-d9e742174491:0168", "section": "F.3. VLM Prior", "page_start": 22, "page_end": 22, "type": "Text", "text": "To test this, we introduce two renderer perturbations (Fig. 8) . First, Sprite Swap replaces object sprites with semantically conflicting alternatives, such as rendering keys as lava or doors as boxes, thereby introducing misleading visual priors. Second, Texture replaces all object appearances with abstract high-contrast patterns, removing naturalistic semantics entirely.", "source": "marker_v2", "marker_block_id": "/page/21/Text/10"}
+{"paper_id": "c39d243f-0c59-4f70-8ee6-d9e742174491", "chunk_id": "c39d243f-0c59-4f70-8ee6-d9e742174491:0169", "section": "F.3. VLM Prior", "page_start": 22, "page_end": 22, "type": "Text", "text": "Empirically, we observe that the unmodified setting achieves near-perfect success and converges the fastest (Fig. 4) . In contrast, both Sprite Swap and Texture slow learning and reduce final success rates, with Texture also leading to the largest variance across seeds. Because the control problem and agent inputs remain fixed, this degradation suggests that the VLM produces less informative priorities when visual evidence for goal-relevant events is either misleading or lacks", "source": "marker_v2", "marker_block_id": "/page/21/Text/11"}
+{"paper_id": "c39d243f-0c59-4f70-8ee6-d9e742174491", "chunk_id": "c39d243f-0c59-4f70-8ee6-d9e742174491:0170", "section": "F.3. VLM Prior", "page_start": 23, "page_end": 23, "type": "Text", "text": "semantic content.", "source": "marker_v2", "marker_block_id": "/page/22/Text/1"}
+{"paper_id": "c39d243f-0c59-4f70-8ee6-d9e742174491", "chunk_id": "c39d243f-0c59-4f70-8ee6-d9e742174491:0171", "section": "F.3. VLM Prior", "page_start": 23, "page_end": 23, "type": "FigureGroup", "text": "(c) Abstract patterns that remove naturalistic cues", "source": "marker_v2", "marker_block_id": "/page/22/FigureGroup/317"}
+{"paper_id": "c39d243f-0c59-4f70-8ee6-d9e742174491", "chunk_id": "c39d243f-0c59-4f70-8ee6-d9e742174491:0172", "section": "F.3. VLM Prior", "page_start": 23, "page_end": 23, "type": "Caption", "text": "Figure 8. Samples of the modified visuals. We modify only the frames passed to the VLM for scoring (agent observations and environment dynamics unchanged).", "source": "marker_v2", "marker_block_id": "/page/22/Caption/7"}
+{"paper_id": "c39d243f-0c59-4f70-8ee6-d9e742174491", "chunk_id": "c39d243f-0c59-4f70-8ee6-d9e742174491:0173", "section": "F.3. VLM Prior", "page_start": 23, "page_end": 23, "type": "Text", "text": "misleading sprites", "source": "marker_v2", "marker_block_id": "/page/22/Text/23"}
+{"paper_id": "c39d243f-0c59-4f70-8ee6-d9e742174491", "chunk_id": "c39d243f-0c59-4f70-8ee6-d9e742174491:0174", "section": "F.4. Computational Overhead", "page_start": 23, "page_end": 23, "type": "TableGroup", "text": "Table 6. Training Throughput (steps/second) comparison on MiniGrid/DoorKey-16x16. Higher is better. Hardware PER VLM-RB Rel. Speed NVIDIA A100 111 97 87% NVIDIA A40 92 81 88% NVIDIA A4000 76 67 88%", "source": "marker_v2", "marker_block_id": "/page/22/TableGroup/318"}
+{"paper_id": "c39d243f-0c59-4f70-8ee6-d9e742174491", "chunk_id": "c39d243f-0c59-4f70-8ee6-d9e742174491:0175", "section": "F.4. Computational Overhead", "page_start": 23, "page_end": 23, "type": "Text", "text": "How much does VLM-RB actually slow down training in practice? To answer this, we measure throughput (steps per second) on the MiniGrid/DoorKey-16x16 task using DQN, comparing PER and VLM-RB across three dual-GPU setups (NVIDIA A100, A40, and A4000). In each case, the RL learner and VLM are placed on separate devices. Notably, in this distributed configuration, the main bottleneck is inter-process communication (IPC) and data transfer, rather than competition for computational resources.", "source": "marker_v2", "marker_block_id": "/page/22/Text/12"}
+{"paper_id": "c39d243f-0c59-4f70-8ee6-d9e742174491", "chunk_id": "c39d243f-0c59-4f70-8ee6-d9e742174491:0176", "section": "F.4. Computational Overhead", "page_start": 23, "page_end": 23, "type": "Text", "text": "The results, summarized in Table 6, reveal a consistent and modest throughput reduction of about 12% across all hardware types. This suggests that, even when VLM inference is separated from the RL learner, the method maintains efficient scaling. In other words, the additional overhead introduced by VLM-RB remains limited in practical distributed settings.", "source": "marker_v2", "marker_block_id": "/page/22/Text/13"}
+{"paper_id": "c39d243f-0c59-4f70-8ee6-d9e742174491", "chunk_id": "c39d243f-0c59-4f70-8ee6-d9e742174491:0177", "section": "F.4. Computational Overhead", "page_start": 23, "page_end": 23, "type": "Text", "text": "It is important to note that these measurements use a standard inference setup, without any aggressive optimizations such as TensorRT or quantization (e.g., INT8 or FP4). We anticipate that employing a dedicated serving stack, such as vLLM or TGI, would further close the speed gap between VLM-RB and the baseline.", "source": "marker_v2", "marker_block_id": "/page/22/Text/14"}
+{"paper_id": "c39d243f-0c59-4f70-8ee6-d9e742174491", "chunk_id": "c39d243f-0c59-4f70-8ee6-d9e742174491:0178", "section": "G. Experiments", "page_start": 24, "page_end": 24, "type": "Text", "text": "In this section, we present the full results of the baselines on all environments.", "source": "marker_v2", "marker_block_id": "/page/23/Text/2"}
+{"paper_id": "c39d243f-0c59-4f70-8ee6-d9e742174491", "chunk_id": "c39d243f-0c59-4f70-8ee6-d9e742174491:0179", "section": "G. Experiments", "page_start": 24, "page_end": 24, "type": "TableGroup", "text": "Table 7. Performance denotes the highest Average Success Rate (M = 5 seeds). Sample Efficiency tracks the steps required to reach the baseline's best performance. Wall-Clock Saving estimates the reduction in real training time, accounting for a 12% inference overhead per step. Alg. Task Baseline Performance (↑) Best ASR Sample Efficiency (↓) Steps to Base. Best Wall-Clock Saving (↑) Time vs. Baseline Scene-3 UER PER +0.0% (1.00/1.00) +0.0% (1.00/1.00) +54.6% (206K/454K) +25.4% (206K/276K) +49.2% +16.4% SAC Scene-4 UER PER +4.2% (1.00/0.96) +4.2% (1.00/0.96) +28.3% (592K/826K) +26.0% (592K/800K) +19.7% +17.1% Scene-5 UER PER +100.0% (0.88/0.44) +41.9% (0.88/0.62) +43.8% (514K/914K) +6.3% (708K/756K) +37.0% -4.8% Scene-3 UER PER +0.0% (1.00/1.00) +0.0% (1.00/1.00) +15.7% (214K/254K) +16.4% (214K/256K) +5.6% +6.4% TD3 Scene-4 UER PER +47.1% (1.00/0.68) +0.0% (1.00/1.00) +63.1% (268K/726K) +9.0% (426K/468K) +58.7% -1.9% Scene-5 UER PER +150.0% (0.70/0.28) +59.1% (0.70/0.44) +49.4% (366K/724K) +28.1% (614K/854K) +43.4% +19.5% DoorKey-8x8 UER PER +0.0% (1.00/1.00) +0.0% (1.00/1.00) +10.7% (150K/168K) +27.9% (150K/208K) +0.0% +19.2% DQN DoorKey-12x12 UER PER +66.7% (1.00/0.60) +0.0% (1.00/1.00) +62.6% (216K/578K) +6.1% (246K/262K) +58.1% -5.2% DoorKey-16x16 UER PER +316.7% (1.00/0.24) +25.0% (1.00/0.80) +62.9% (340K/916K) +35.5% (382K/592K) +58.4% +27.7% DoorKey-8x8 UER PER +0.0% (1.00/1.00) +0.0% (1.00/1.00) +26.6% (138K/188K) +16.9% (138K/166K) +17.8% +6.9% IQN DoorKey-12x12 UER PER +56.2% (1.00/0.64) +56.2% (1.00/0.64) +44.7% (388K/702K) +42.3% (388K/672K) +38.1% +35.3% DoorKey-16x16 UER PER +166.7% (0.64/0.24) +300.0% (0.64/0.16) +11.0% (774K/870K) +13.8% (562K/652K) +0.4% +3.5%", "source": "marker_v2", "marker_block_id": "/page/23/TableGroup/676"}
+{"paper_id": "c39d243f-0c59-4f70-8ee6-d9e742174491", "chunk_id": "c39d243f-0c59-4f70-8ee6-d9e742174491:0180", "section": "G. Experiments", "page_start": 24, "page_end": 24, "type": "Caption", "text": "Table 8. Performance denotes the highest Average Success Rate (M = 5 seeds). Sample Efficiency tracks the steps required to reach the baseline's best performance. Both metrics are averaged across the aggregated algorithms and tasks.", "source": "marker_v2", "marker_block_id": "/page/23/Caption/7"}
+{"paper_id": "c39d243f-0c59-4f70-8ee6-d9e742174491", "chunk_id": "c39d243f-0c59-4f70-8ee6-d9e742174491:0181", "section": "G. Experiments", "page_start": 24, "page_end": 24, "type": "Table", "text": "Env Type Agg. Algorithms Baseline Performance (↑) Mean Best ASR Sample Efficiency (↓) Mean Steps to Base Peak Scene (SAC + TD3) UER PER +28.0% (0.93/0.73) +11.2% (0.93/0.84) +44.6% (360K/650K) +19.1% (460K/568K) DoorKey (DQN + IQN) UER PER +51.6% (0.94/0.62) +22.6% (0.94/0.77) +41.4% (334K/570K) +26.9% (311K/425K)", "source": "marker_v2", "marker_block_id": "/page/23/Table/9"}
+{"paper_id": "c39d243f-0c59-4f70-8ee6-d9e742174491", "chunk_id": "c39d243f-0c59-4f70-8ee6-d9e742174491:0182", "section": "G. Experiments", "page_start": 25, "page_end": 25, "type": "FigureGroup", "text": "Figure 9. VLM-RB consistently outperforms baselines across continuous and discrete tasks. The plots show aggregated success rates for four algorithms (DQN, IQN, SAC, TD3) on MiniGrid and OGBench domains. Annotations highlight the relative improvement in sample efficiency (horizontal arrows, reaching peak performance faster) and the best success rate (vertical arrows). Shaded regions indicate standard deviation across seeds.", "source": "marker_v2", "marker_block_id": "/page/24/FigureGroup/210"}

icml26/c39d243f-0c59-4f70-8ee6-d9e742174491/appendix_text_v3.txt

@@ -0,0 +1,272 @@
+[p. 12 | section: A. VLM-RB | type: Text]
+632633634
+
+[p. 12 | section: A. VLM-RB | type: Text]
+638639
+
+[p. 12 | section: A. VLM-RB | type: Text]
+648649
+
+[p. 12 | section: A. VLM-RB | type: Text]
+653654
+
+[p. 12 | section: Algorithm 1 VLM-Prioritized Replay Buffer | type: Code]
+608 Require: prompt P; clip length L. 609 replay buffer \mathcal{D}; clip buffer \mathcal{C}; queues \mathcal{Q}_{\mathrm{in}} and \mathcal{Q}_{\mathrm{out}}. 610 1: function VLMWORKER 611 2: while true do (idxs, \tau^O) \leftarrow pop(Q_{in}) 3: 612 \mathbf{p^{VLM}} = f_{\mathrm{VLM}}(\tau^O, \mathsf{P}) \in \mathbb{R} 4: 613 (Equation 1) \operatorname{push}(\mathcal{Q}_{\mathrm{out}},(\operatorname{idxs},\operatorname{\mathbf{p}^{\acute{\mathbf{VLM}}}})) 5: 614 615 6: Init: launch VLMWORKER asynchronously. 7: while Training do 616 8: ⊳(1) Env step 617 9: Step env to obtain (s, a, r, s') and visual observation o = \psi(s) 618 10: idx \leftarrow INSERT(\mathcal{D}, (s, a, r, s'; \bar{p})) 619 \mathcal{C} \leftarrow \mathcal{C} \cup (\text{idx}, o) 11: 620 12: ⊳ (2) Enqueue clips (streaming) if |\mathcal{C}| = L or terminated or truncated then 621 13: idxs, \tau^O \leftarrow \{(idx_i, o_i)\}_{i=0}^{|C|-1} 622 14: \operatorname{push}(\mathcal{Q}_{\operatorname{in}},(\operatorname{idxs},\tau^O)) 623 15: \mathcal{C} \leftarrow \emptyset 16: 624 ▷ (3) Apply VLM scores (drain output queue) 17: 625 while \mathcal{Q}_{\mathrm{out}} not empty do 18: (\texttt{idxs}, \mathbf{p^{VLM}}) \leftarrow \mathbf{pop}(\mathcal{Q}_{\text{out}}) 626 19: 627 20: for all idx \in idxs do \texttt{SETPRIORITY}(\mathcal{D}, \texttt{idx}, \mathbf{p^{VLM}}) 628 21: \bar{p} \leftarrow \text{CMA}(\bar{p}, \mathbf{p^{VLM}}) 22: 629 \triangleright (4) Sample and learn 23: 630 \mathcal{B} \sim \lambda_t q_t^{\hat{P}} + (1 - \lambda_t) q_t^{U} (Equation 2) 631 25: UPDATELEARNER(\mathcal{B}); update \lambda_t
+
+[p. 12 | section: Algorithm 1 VLM-Prioritized Replay Buffer | type: Text]
+In Algorithm 1, C denotes a temporary clip buffer, Q_{in} and Q_{out} are asynchronous communication queues, and CMA refers to the Cumulative Moving Average update rule for the global priority statistics.
+
+[p. 12 | section: B. Environment Details | type: Text]
+In this section, we describe the environments used in our experiments.
+
+[p. 12 | section: B. Environment Details | type: Text]
+MiniGrid / DoorKey. We use the <code>DoorKey</code> tasks from the MiniGrid suite (Chevalier-Boisvert et al., 2023) with sizes 8x8, 12x12, and 16x16. Each episode requires the agent to complete a sequence of subtasks: (i) navigate to and pick up the key, (ii) reach the locked door and open it using the <code>toggle</code> action (which is only possible while holding the key), and (iii) proceed to the goal tile. This structure enforces temporal dependencies and compositional reasoning. The reward function is sparse: the agent receives zero reward at all intermediate steps and a positive reward only upon reaching the goal, with the reward magnitude linearly decaying with the number of steps taken, following the standard MiniGrid protocol. Episodes terminate either upon successful completion or when the maximum step limit is reached.
+
+[p. 12 | section: B. Environment Details | type: Text]
+State observations are provided in a symbolic format, where each s_t \in \{0, 1, \ldots\}^{N \times N \times 3} encodes the full grid. The three channels correspond to (i) object indices, (ii) color indices, and (iii) a state channel that includes both the door state and the agent's orientation. The action space is discrete with |\mathcal{A}| = 5 , consisting of turn left, turn right, move forward, pick up, and toggle actions.
+
+[p. 12 | section: B. Environment Details | type: Text]
+OGBench / Scene. We use the scene-play manipulation environment from OGBench (Park et al., 2024), which is a MuJoCo-based tabletop task involving a UR5e arm with a parallel gripper. The agent interacts with a set of objects: a cube, a sliding drawer, a sliding window, and two toggle buttons. Each button controls the lock state of either the drawer or the window, introducing dependencies that require the agent to execute explicit unlock, manipulate, and (re-)lock sequences to achieve many goals. Episodes are capped at a horizon of 750 steps. To facilitate learning, we employ potential-based reward
+
+[p. 13 | section: B. Environment Details | type: Text]
+shaping with a per-step living cost, defined as:
+
+[p. 13 | section: B. Environment Details | type: Equation]
+r_t = -1 \ + \ \gamma \, \phi(s_{t+1}) - \phi(s_t), \qquad \phi(s) = \frac{1}{K} \sum_{i=1}^K \mathbb{I} \left[ \mathrm{subgoal}_i(s) \right].
+
+[p. 13 | section: B. Environment Details | type: Text]
+Here, \phi(s) denotes the fraction of satisfied subgoals, where each subgoal is a task-specific predicate over cube placement, button states, and the positions of the drawer and window. An episode terminates either when all subgoals are satisfied (\phi(s_t) = 1) or when the maximum horizon is reached.
+
+[p. 13 | section: B. Environment Details | type: Text]
+Observations are provided as full state vectors, s_t \in \mathbb{R}^{40} , and actions are specified as continuous end-effector controls, a_t \in \mathbb{R}^5 .
+
+[p. 13 | section: B. Environment Details | type: Text]
+To address the exploration challenges inherent in these long-horizon tasks, we initialize the replay buffer with a lightweight warm-start of 10 demonstration episodes, which are held fixed across all methods. This small set of demonstrations provides minimal task-relevant coverage, while leaving the remainder of the training protocol unchanged.
+
+[p. 13 | section: B. Environment Details | type: Text]
+We evaluate on a set of predefined Scene goals (tasks 3–5), which are designed to progressively increase the degree of temporal composition required for successful completion.
+
+[p. 13 | section: B. Environment Details | type: ListGroup]
+Task 3 (rearrange-medium). The agent must move the cube to a specified tabletop location, open the drawer, close the window, and terminate with both the drawer and window unlocked. Notably, the window begins in a locked and open state, so the policy must first unlock it before closing, while simultaneously coordinating the manipulation of the drawer and the relocation of the cube. Task 4 (put-in-drawer). The agent is required to place the cube inside the drawer and terminate with the drawer closed and unlocked, while ensuring the window remains locked. This sequence involves unlocking the drawer, opening it, inserting the cube, and then closing the drawer. Task 5 ( rearrange-hard ). The agent must place the cube inside the (closed) drawer and leave the window open, while ensuring that both the drawer and window are locked at the end of the episode. Achieving this goal requires the agent to execute both unlock and relock sequences, coordinate drawer opening and closing, and correctly position the cube.
+
+[p. 14 | section: C. Prompts | type: Text]
+Task prompts used in our experiments. We use a binary "success visible" query with a strict output format.
+
+[p. 14 | section: C. Prompts | type: ListGroup]
+1. MiniGrid/DoorKey: "Does this clip contain a clear instance of goal satisfaction anywhere in it? If no visible success occurs, answer No. Do not guess. Output exactly Answer: Yes or Answer: No." 2. OGBench/Scene: "Is there at least one clear instance of goal satisfaction in these frames? Look for contact + displacement consistent with the goal (lift off surface, place into receptacle, open/close articulation, move to target zone). Do not guess. If not visible, answer No. Output exactly Answer: Yes or Answer: No."
+
+[p. 14 | section: C. Prompts | type: Text]
+Meta-prompt for generating task prompts. To facilitate applying our procedure to new environments, we use the following meta-prompt to generate a task prompt from a task identifier and optional human context. The meta-prompt encourages generic, visually grounded success criteria and enforces a strict output schema.
+
+[p. 14 | section: Meta-prompt template for generating task prompts | type: Text]
+User: You are an expert in Reinforcement Learning and Visual Language Models. I will provide a short clip from an agent rollout. Your job is to write a single text prompt that asks a VLM to output a binary judgment about whether goal satisfaction / competent task progress is clearly visible in the clip. Requirements:
+
+[p. 14 | section: Meta-prompt template for generating task prompts | type: ListGroup]
+Keep the prompt environment-agnostic: rely on visible physics and outcomes rather than simulator-specific rules. Specify what to look for in broad categories when helpful (e.g., contact + displacement for manipulation; reaching a target region for navigation). Explicitly instruct: do not guess; if success is not clearly visible, answer No. End the prompt with the string: Output exactly Answer: Yes or Answer: No.
+
+[p. 14 | section: Examples (few-shot) | type: Text]
+• Task: Minigrid-doorkey
+
+[p. 14 | section: Examples (few-shot) | type: Text]
+Prompt: "Does this clip contain a clear instance of goal satisfaction anywhere in it? If no visible success occurs, answer No. Do not guess. Output exactly Answer: Yes or Answer: No."
+
+[p. 14 | section: Examples (few-shot) | type: Text]
+• Task: OGBench-scene-task-4
+
+[p. 14 | section: Examples (few-shot) | type: Text]
+Prompt: "Is there at least one clear instance of goal satisfaction in these frames? Look for contact + displacement consistent with the goal (lift off surface, place into receptacle, open/close articulation, move to target zone). Do not guess. If not visible, answer No. Output exactly Answer: Yes or Answer: No."
+
+[p. 14 | section: Current request | type: Text]
+Task Name: {{TASK ID}}
+
+[p. 14 | section: Current request | type: Text]
+Optional Human Context: {{HUMAN CONTEXT}}
+
+[p. 15 | section: Current request | type: Text]
+Example: prompt families for a new benchmark. As an illustration, we can apply the meta-prompt to a suite such as DeepMind Control by grouping tasks by their visible success patterns (locomotion / balance / interaction) and instantiating one generic prompt per group.
+
+[p. 15 | section: Example: prompt families for DeepMind Control Suite | type: Text]
+User: Group DeepMind Control tasks by shared visible success patterns and propose one generic success-detection prompt per group (not task-specific). Each prompt must end with: "Output exactly Answer: Yes or Answer: No."
+
+[p. 15 | section: Example: prompt families for DeepMind Control Suite | type: Text]
+Model: We used Gemini 3 Pro to generate these prompts. Here is the partition into three semantic groups based on the nature of the movement, with a generalized "vibe-check" prompt for each.
+
+[p. 15 | section: Group 1: Going Somewhere (Locomotion) | type: Text]
+The Semantic Goal: The agent must make progress across the environment. The VLM needs to distinguish between an agent that is struggling/falling (chaos) and one that is moving purposefully (control).
+
+[p. 15 | section: Group 1: Going Somewhere (Locomotion) | type: Text]
+The Group: cheetah-run, dog-walk, dog-trot, dog-run, fish-swim, hopper-hop, humanoid-walk, humanoid-run, quadruped-walk, quadruped-run, walker-walk, walker-run.
+
+[p. 15 | section: The Prompt | type: Text]
+"Does the agent demonstrate controlled, purposeful movement through the environment? Look for the agent making progress in a specific direction (locomotion) rather than just flailing in place, falling over, or crashing. The movement should look stable and competent. If the agent falls, flips, or fails to move effectively, answer No. Output exactly Answer: Yes or Answer: No."
+
+[p. 15 | section: Group 2: Defying Gravity (Stabilization) | type: Text]
+The Semantic Goal: The agent is fighting gravity. Success is defined by stillness or verticality. Whether it is a pole swinging up or a dog standing still, the "vibe" of success is reaching a stable, upright state and holding it.
+
+[p. 15 | section: Group 2: Defying Gravity (Stabilization) | type: Text]
+The Group: acrobot-swingup, cartpole-balance, cartpole-balance sparse, cartpole-swingup, cartpole-swingup sparse, dog-stand, hopper-stand, humanoid-stand, walker-stand, pendulum-swingup.
+
+[p. 15 | section: The Prompt | type: Text]
+"Does the agent successfully achieve and maintain a stable, upright posture against gravity? Look for the agent balancing itself or an object vertically. Success means the agent (or the object it controls) is standing up or swinging up to a balanced position and staying there without collapsing, spinning out of control, or falling over. If it looks unstable or falls, answer No. Output exactly Answer: Yes or Answer: No."
+
+[p. 15 | section: Group 3: Touching the Thing (Interaction) | type: Text]
+The Semantic Goal: This is about precise manipulation. The agent is not trying to go somewhere or just stand up; it is trying to force an interaction between its end-effector and a target object.
+
+[p. 15 | section: Group 3: Touching the Thing (Interaction) | type: Text]
+The Group: ball in cup-catch, finger-spin, finger-turn easy, finger-turn hard, reacher-easy, reacher-hard.
+
+[p. 15 | section: The Prompt | type: Text]
+"Does the agent successfully interact with the target object in this scene? Look for the agent reaching a target, catching an object, or spinning a mechanism as intended. There should be clear, controlled contact or manipulation consistent with solving a task. If the agent misses the target, drops the object, or flails without making meaningful contact, answer No. Output exactly Answer: Yes or Answer: No."
+
+[p. 16 | section: D. Baselines | type: Text]
+DoorKey presents a particularly challenging regime, characterized by extremely sparse rewards (identically zero until termination) and a long-horizon, sequential dependency structure (key to door to goal). In such settings, replay schemes relying on dense feedback, informative TD statistics, or a well-defined similarity metric are fundamentally misaligned with the task: they fail to prioritize transitions relevant to solving the task, particularly throughout the extended pre-success phase.
+
+[p. 16 | section: D. Baselines | type: Text]
+Experience Replay Optimization (ERO) (Zha et al., 2019) ERO optimizes a learned rejection sampler using a REINFORCE-style update, where the scalar reward is defined as the change in evaluation return. In DoorKey, evaluation returns remain identically zero for a substantial portion of training (reflecting the absence of successes), which in turn implies rreplay ≈ 0 and results in near-zero gradients for the replay policy over extended periods. Furthermore, the replay-policy features used here, (r i , |δ i |, ti/Tmax), are weakly informative under sparse rewards: r i = 0 for almost all transitions and |δ i | is largely determined by bootstrap noise in the early stages of learning. As a result, the rejection policy effectively acts as an untrained stochastic filter for much of the training. When successes eventually occur, rreplay becomes highly variable, leading to unstable and non-stationary updates which can bias acceptance toward incidental correlates, such as late timesteps. The top-up rule, which fills with the highest-P i candidates when acceptance is low, further concentrates replay on a narrow subset of transitions. This reduces diversity and fails to provide a consistent signal for solving the task.
+
+[p. 16 | section: D. Baselines | type: Text]
+ReLo (Reducible Loss Prioritization) (Sujit et al., 2023) ReLo prioritizes transitions based on the difference between the magnitudes of online and target TD residuals, pReLo = max(0, |δθ| − |δθ − |) +ϵ. In the sparse-reward DoorKey setting, both residuals are typically dominated by bootstrapping error rather than meaningful reward propagation throughout the prolonged pre-success phase. Moreover, the use of hard target updates periodically synchronizes the online and target networks, further diminishing any systematic separation between |δθ| and |δθ − |. Consequently, |δθ| − |δθ − | is often small or negative and is clipped to approximately ϵ, so the sampling distribution effectively reverts toward uniform, yet still incurs the variance and bias tradeoffs associated with prioritized replay.
+
+[p. 16 | section: D. Baselines | type: Text]
+Attentive Experience Replay (AER) (Sun et al., 2020) AER selects samples closest (in a frozen embedding space) to the agent's current state, thereby inducing a strongly local replay distribution. In DoorKey this notion of locality is fundamentally misaligned with the task structure. Once trajectories reach later phases (near the door or goal), nearestneighbor replay disproportionately samples those regions and neglects earlier structural dependencies, such as key acquisition, thereby impeding temporal credit assignment across the long horizon. This effect is further amplified by the use of a randomly-MLinitialized frozen encoder: the resulting metric primarily captures superficial spatial proximity in the grid encoding, rather than functional or task-relevant similarity. As a result, the selection rule acts as a myopic location-based filter, rather than an attentive semantic sampler.
+
+[p. 17 | section: E. Implementation Details | type: Text]
+To ensure a fair comparison, all baselines are implemented with the same backbone architecture, optimizer, and training schedule as VLM-RB . Unless otherwise specified, each method employs a PER -style replay buffer parameterized by exponents \alpha and \beta .
+
+[p. 17 | section: E. Implementation Details | type: Text]
+Uniform Experience Replay In this setting, each transition i in the buffer is sampled with equal probability, p_i = 1/N_{curr} , where N_{curr} denotes the current buffer size.
+
+[p. 17 | section: E. Implementation Details | type: Text]
+PER ( Prioritized Experience Replay ) Here, the priority assigned to each transition i is given by p_i = |\delta_i| + \epsilon , where \delta_i denotes the most recent TD-error for that transition, and \epsilon = 10^{-6} guarantees that every transition can be sampled. The probability of sampling transition i is then defined as
+
+[p. 17 | section: E. Implementation Details | type: Equation]
+\mathbf{p}_i = \frac{p_i^{\alpha}}{\sum_k p_k^{\alpha}},
+
+[p. 17 | section: E. Implementation Details | type: Text]
+with \alpha determining the extent to which prioritization influences sampling. To account for the bias from prioritized sampling, importance-sampling (IS) weights are computed as
+
+[p. 17 | section: E. Implementation Details | type: Equation]
+w_i = \left(\frac{1}{N} \cdot \frac{1}{p_i}\right)^{\beta}.
+
+[p. 17 | section: E. Implementation Details | type: Text]
+These weights are normalized by 1/\max_i(w_i) prior to being used in the loss.
+
+[p. 17 | section: E. Implementation Details | type: Text]
+Experience Replay Optimization (ERO) ERO (Zha et al., 2019) replaces standard replay sampling with a learned rejection policy applied to uniformly drawn candidates. The replay policy is parameterized as an MLP \phi_{\psi} with two hidden layers of 64 units each (ReLU activations, sigmoid output), mapping transition features to a retention probability.
+
+[p. 17 | section: E. Implementation Details | type: Equation]
+p_i = \phi_{\psi}(r_i, |\delta_i|, t_i/T_{\text{max}}) \in (0, 1),
+
+[p. 17 | section: E. Implementation Details | type: Equation]
+\mathcal{L}_{ERO} = -r_{replay} \sum_{i \in \mathcal{B}} \log p_i,
+
+[p. 17 | section: E. Implementation Details | type: Text]
+where gradients are applied only to \psi .
+
+[p. 17 | section: E. Implementation Details | type: Text]
+ReLo (Reducible Loss Prioritization) ReLo (Sujit et al., 2023) assigns priority to each transition based on a reducible-loss score, which is computed from the online and target Q-networks. Given a transition (s,a,r,s'), the target value is defined as y=r+\gamma\max_{a'}Q_{\theta^-}(s',a') . The online and target TD-errors are then \delta_{\text{online}}(s,a)=Q_{\theta}(s,a)-y and \delta_{\text{target}}(s,a)=Q_{\theta^-}(s,a)-y , respectively. The sampling priority is
+
+[p. 17 | section: E. Implementation Details | type: Equation]
+p_{\text{ReLo}}(s, a) = \max(0, |\delta_{\text{online}}(s, a)| - |\delta_{\text{target}}(s, a)|) + \epsilon, \qquad \epsilon = 10^{-6}.
+
+[p. 17 | section: E. Implementation Details | type: Text]
+Each new transition is initialized with the maximum priority value among all transitions currently in the buffer. The method sets \alpha = 0.6 and initializes \beta = 0.4 , annealing \beta linearly to 1.0. Standard importance-sampling weights are applied.
+
+[p. 18 | section: E. Implementation Details | type: Text]
+Attentive Experience Replay (AER) AER (Sun et al., 2020) selects transitions using an attentive sampling mechanism. A dedicated encoder \phi is constructed with the same backbone architecture as the Q-network. This encoder is randomly initialized and remains fixed throughout training. At each training step t, a candidate pool of size N_{\text{cand}} = \lfloor \lambda_t B \rfloor is sampled uniformly from the buffer, where \lambda_t decays linearly from \lambda_0 = 4 to 1 over the course of training. If the candidate pool size N_{\text{cand}} is less than or equal to the batch size B, the method reverts to the default sampling strategy. Given the current state s_{\text{curr}} and a set of candidate states \{s_i\} , distances are computed in the frozen embedding space as d(s_{\text{curr}}, s_i) = \|\phi(s_{\text{curr}}) - \phi(s_i)\|_2^2 . The B candidates with the smallest distances are then selected deterministically to form the training batch. Following Sun et al. (2020), importance sampling is disabled.
+
+[p. 18 | section: E.1. Architectural and Implementation Details | type: Text]
+MiniGrid agents (DQN / IQN). Both DQN and IQN are trained on symbolic Minigrid observations s_t \in \{0, \dots\}^{N \times N \times 3} , which encode object identities, colors, and state information such as door status and agent orientation. To process these inputs, we employ a shared encoder which first embeds each channel, then applies a residual dilated CNN with three blocks. The resulting features are aggregated using both average and max global pooling, yielding a 256-dimensional representation. This design aims to capture both local and global spatial structure in the environment. For DQN, the 256-dimensional feature vector is passed through a small MLP, which outputs scalar Q-values for each discrete action. In IQN, we replace the scalar output head with an implicit quantile head. Specifically, we embed sampled quantile fractions using cosine functions, project these embeddings to match the feature dimension, and combine them element-wise with the state features. The resulting representations are mapped to per-action quantile values. We optimize using the quantile Huber loss and select actions using either Double-DQN-style (Van Hasselt et al., 2016) or target-network-based strategies, as described in Appendix E. This approach allows IQN to model the full distribution over returns, rather than just the mean.
+
+[p. 18 | section: E.1. Architectural and Implementation Details | type: Text]
+OGBench Scene agents (SAC / TD3). For OGBench/Scene, both SAC and TD3 operate on flattened state-based observations s_t \in \mathbb{R}^{40} . We adopt a unified architecture across tasks: the actor is a fully-connected network with three hidden layers of width 512 and ReLU activations. This network outputs either a tanh-squashed Gaussian policy (for SAC) or a deterministic tanh policy (for TD3), enabling flexible action selection in continuous spaces. The critic is implemented as an ensemble of N=10 Q-functions, each parameterized by a 3-layer MLP that receives the concatenated state and action as input and outputs a scalar value. For target computation, we randomly select M=2 ensemble members and aggregate their predictions using either a mean (in all main SAC/TD3 runs) or a min operator. This ensemble approach is intended to improve stability and reduce overestimation bias. For actor updates, we use the same Q-function ensemble, aggregating Q-values according to the actor reduction rule. In all reported experiments, we use the mean over ensemble members. This ensures consistency between actor and critic updates.
+
+[p. 18 | section: E.1. Architectural and Implementation Details | type: Text]
+Replay buffer and prioritization. All methods are implemented with a two-branch replay ensemble: one prioritized branch and one uniform branch which share an underlying storage, implemented using ReplayBufferEnsemble (Bou et al., 2023) with mixture weights controlled by the sampling ratio schedule in Tables 2–3. For the PER baseline, we assign priorities using the standard PER formula p_i = |\delta_i| + \epsilon , where \delta_i is the TD error. When importance-sampling corrections are enabled, we apply them as usual. New transitions are initially assigned a small default priority, and we update their TD-errors after each gradient step to ensure accurate prioritization. In VLM-RB, we augment each transition with both a TD-based metric and a VLM-based score. For MiniGrid/DoorKey, replay priorities are determined solely by the VLM score. In contrast, for OGBench/Scene, we set priorities as the product of the VLM score and the TD-based metric (see Section 3.3). This design allows us to tailor prioritization to the characteristics of each environment.
+
+[p. 18 | section: E.1. Architectural and Implementation Details | type: Text]
+VLM worker. The VLM worker uses Perception-LM-1B (Cho et al., 2025) with a fixed prompt to detect binary success. For each transition, it processes a clip of L=32 frames, applying left padding if the clip is shorter. To compute a scalar priority, we sum the probabilities assigned to all "Yes" and "No" token variants in the first generated token's logits, and calculate their ratio to obtain a probability-style score. We hard-threshold these scores at 0.5, assigning 1 or 0 accordingly. The resulting priorities are asynchronously streamed back to update the replay buffer, ensuring the main learner is not blocked during this process.
+
+[p. 19 | section: E.1. Architectural and Implementation Details | type: Caption]
+Table 2. Hyperparameters and implementation details for OGBench experiments. Shared parameters are listed once; divergent parameters are compared side-by-side.
+
+[p. 19 | section: E.1. Architectural and Implementation Details | type: Table]
+Hyperparameter TD3 SAC Network Architecture Hidden Dimensions [512, 512, 512] Q-Network Layer Norm False True Optimization Critic Learning Rate 3 × 10−4 Actor Learning Rate 3 × 10−4 1 × 10−3 Alpha Learning Rate – 3 × 10−4 Batch Size 256 Discount Factor (γ) 0.99 Target Smoothing (τ ) 0.005 Max Grad Norm 10.0 Learning Starts 10,000 steps Ensemble & Update Ratios Ensemble Size (N) 10 Subsampled Q-Networks (M) 2 Critic UTD Ratio 4 Actor UTD Ratio 2 Target Q Reduction min mean Actor Q Reduction min mean Data & Replay Buffer Replay Buffer Size 1 × 106 Expert Demos 10 Algorithm Specifics Policy Update Frequency 2 1 Exploration Noise (std) 0.1 – Target Policy Noise (std) 0.2 – Noise Clip 0.5 – Entropy Coeff. (α) – 1.0 (Initial) Target Entropy Scale – 0.5 Auto Entropy Tuning – True PER Baseline Settings PER Alpha (α) 0.7 PER Beta (β) 1.0 Importance Sampling False VLM-RB Settings VLM Model Facebook Perception-LM-1B Priority Mode Hard Threshold (> 0.5) Trajectory Length (L) 32 Frames Importance Sampling False Prioritized Sampling Ratio (λ0 and λmax) Annealed 0.0 → 0.5 Annealing Schedule (Tschedule) Linear over 500k steps
+
+[p. 20 | section: E.1. Architectural and Implementation Details | type: Caption]
+Table 3. Hyperparameters and implementation details for MiniGrid experiments. Shared parameters are listed once; divergent parameters are compared side-by-side.
+
+[p. 20 | section: E.1. Architectural and Implementation Details | type: Table]
+Hyperparameter DQN IQN Optimization Learning Rate 4 × 10−5 Batch Size 128 Discount Factor (γ) 0.95 Target Update Frequency 1,000 steps Target Update Rate (τ ) 1.0 (Hard Update) Learning Starts 500 steps Train Frequency 4 steps Max Grad Norm 1.0 Exploration (Epsilon-Greedy) epsstart 1.0 epsend 0.05 Exploration Fraction 0.5 (First 50% of training) Algorithm Specifics Double DQN Enabled Noisy Nets – Disabled Num Quantiles (Policy) – 32 Num Quantiles (Train/Target) – 64 Num Cosine Basis Functions – 64 Huber Kappa (κ) – 1.0 PER Baseline Settings PER Alpha (α) 0.7 PER Beta (β) 1.0 Importance Sampling True Prioritized Sampling Ratio 1.0 (Always Prioritized) VLM-RB Settings VLM Model Facebook Perception-LM-1B Priority Mode Hard Threshold (> 0.5) Trajectory Length (L) 32 Frames Importance Sampling False Prioritized Sampling Ratio (λ0 and λmax) Annealed 0.0 → 0.5 Annealing Schedule (Tschedule) Linear over 500k steps
+
+[p. 21 | section: F.1. Mixing Schedule | type: TableGroup]
+Table 4. Ablation of mixing schedule (λmax) on DoorKey-16x16. Performance denotes the highest Average Success Rate (M = 5 seeds). Sample Efficiency tracks the steps required to reach the baseline's best performance. Improvement λmax Base. Perf. (↑) Sample Eff. (↓) None UER PER N/A (0.00/0.24) N/A (0.00/0.80) N/A (1000K/916K) N/A (1000K/592K) PER +0.0% (0.80/0.80) 0.25 UER +233.3% (0.80/0.24) +40.6% (544K/916K) -0.3% (594K/592K) 0.50 UER +316.7% (1.00/0.24) +62.9% (340K/916K) PER +25.0% (1.00/0.80) +35.5% (382K/592K) 0.75 UER +266.7% (0.88/0.24) +57.0% (394K/916K) PER +10.0% (0.88/0.80) +17.2% (490K/592K) 1.00 UER +316.7% (1.00/0.24) +69.0% (284K/916K) PER +25.0% (1.00/0.80) +40.2% (354K/592K)
+
+[p. 21 | section: F.1. Mixing Schedule | type: Text]
+To understand the effect of the final mixing coefficient λmax, we conduct an ablation on the MiniGrid/DoorKey-16x16 task (Fig .6) . We fix Tschedule = 5 · 10 5 , ensuring that λ t is annealed linearly from 0 (corresponding to purely uniform sampling) to λmax over the first half of training (right panel). We sweep λmax ∈ {0.25, 0.5, 0.75, 1.0} and include a None baseline, which disables the schedule and relies entirely on VLM-prioritized sampling. We observe that larger λmax values (0.75, 1.0) reach 50% success marginally earlier than smaller values, but their final success rates remain lower within the fixed training budget. In contrast, λmax = 0.5 emerges as the most reliable choice in this environment: it consistently achieves 100% success within the budget and attains the highest performance at the 90% success threshold among all options considered. The smallest value (λmax = 0.25) results in slower learning, and the fully prioritized variant ( None ) fails to solve the task under our setup, with success rates remaining near zero. This suggests that maintaining a non-trivial fraction of uniform sampling is essential for effective learning in this setting. Based on these findings, we fix λmax = 0.5
+
+[p. 21 | section: F.1. Mixing Schedule | type: Text]
+for all main experiments without further per-task tuning.
+
+[p. 21 | section: F.1. Mixing Schedule | type: Caption]
+Figure 6. MiniGrid/DoorKey-16x16: λmax Ablation (5 seeds), "None" corresponds to no scheduling, i.e., only VLM-prioritized sampling. Left: performance for different λmax values. Right: λmax value evolution.
+
+[p. 22 | section: F.2. VLM Size | type: TableGroup]
+Table 5. Performance comparison of Perception-LM (Cho et al., 2025) models. Values in parentheses denote the relative percentage change compared to 1B. We use an Nvidia RTX 4090 GPU, and run 100 batches of 32 frames. Model Load (GiB) Peak (GiB) Time (s) FPS 1B 2.86 3.77 0.46 69.27 3B 6.56 (+130%) 8.16 (+116%) 0.77 (+66%) 41.75 (-40%) 8B 18.25 (+539%) 20.34 (+439%) 2.15 (+366%) 14.88 (-79%)
+
+[p. 22 | section: F.2. VLM Size | type: Text]
+We investigate how scaling the VLM affects both inference overhead and downstream RL performance. Table 5 quantifies the resource requirements of Perception-LM (Cho et al., 2025) variants for clip scoring (Nvidia RTX 4090 GPU; 100 batches of 32-frame clips). We observe that increasing the VLM size leads to a substantial increase in memory footprint and a corresponding reduction in throughput. Relative to the 1B model, the 3B variant increases peak memory by +116% and reduces FPS by 40%; the 8B variant increases peak memory by +439% and reduces FPS by 79%.
+
+[p. 22 | section: F.2. VLM Size | type: Text]
+To assess the impact of VLM size on downstream performance, we compare the 1B, 3B, and 8B models on MiniGrid/DoorKey-16x16 (Fig. 7) . Notably, despite the increased inference cost of larger models, we do not observe consistent improvements in RL performance relative to the 1B configuration. These results suggest that in this setting, the clip-scoring signal saturates once the VLM is sufficiently reliable at separating task-relevant from irrelevant segments. Given this trade-off, we select the 1B Perception-LM as the default backbone for all main experiments.
+
+[p. 22 | section: F.2. VLM Size | type: FigureGroup]
+Figure 7. MiniGrid/DoorKey-16x16: VLM Size Ablation (5 seeds)
+
+[p. 22 | section: F.3. VLM Prior | type: Text]
+Motivated by the "Modified Game" paradigm of Dubey et al. (2018) , we probe whether VLM-RB leverages the VLM's pre-trained visual semantics to improve performance. To isolate the effect of visual semantics, we modify only the rendered frames provided to the VLM for scoring, while leaving both the underlying MDP and the agent's observations unchanged. If VLM-RB relies on semantic cues such as identifying keys, doors, or goal-relevant interactions, we expect its performance gains to diminish when these cues are distorted or removed.
+
+[p. 22 | section: F.3. VLM Prior | type: Text]
+To test this, we introduce two renderer perturbations (Fig. 8) . First, Sprite Swap replaces object sprites with semantically conflicting alternatives, such as rendering keys as lava or doors as boxes, thereby introducing misleading visual priors. Second, Texture replaces all object appearances with abstract high-contrast patterns, removing naturalistic semantics entirely.
+
+[p. 22 | section: F.3. VLM Prior | type: Text]
+Empirically, we observe that the unmodified setting achieves near-perfect success and converges the fastest (Fig. 4) . In contrast, both Sprite Swap and Texture slow learning and reduce final success rates, with Texture also leading to the largest variance across seeds. Because the control problem and agent inputs remain fixed, this degradation suggests that the VLM produces less informative priorities when visual evidence for goal-relevant events is either misleading or lacks
+
+[p. 23 | section: F.3. VLM Prior | type: Text]
+semantic content.
+
+[p. 23 | section: F.3. VLM Prior | type: FigureGroup]
+(c) Abstract patterns that remove naturalistic cues
+
+[p. 23 | section: F.3. VLM Prior | type: Caption]
+Figure 8. Samples of the modified visuals. We modify only the frames passed to the VLM for scoring (agent observations and environment dynamics unchanged).
+
+[p. 23 | section: F.3. VLM Prior | type: Text]
+misleading sprites
+
+[p. 23 | section: F.4. Computational Overhead | type: TableGroup]
+Table 6. Training Throughput (steps/second) comparison on MiniGrid/DoorKey-16x16. Higher is better. Hardware PER VLM-RB Rel. Speed NVIDIA A100 111 97 87% NVIDIA A40 92 81 88% NVIDIA A4000 76 67 88%
+
+[p. 23 | section: F.4. Computational Overhead | type: Text]
+How much does VLM-RB actually slow down training in practice? To answer this, we measure throughput (steps per second) on the MiniGrid/DoorKey-16x16 task using DQN, comparing PER and VLM-RB across three dual-GPU setups (NVIDIA A100, A40, and A4000). In each case, the RL learner and VLM are placed on separate devices. Notably, in this distributed configuration, the main bottleneck is inter-process communication (IPC) and data transfer, rather than competition for computational resources.
+
+[p. 23 | section: F.4. Computational Overhead | type: Text]
+The results, summarized in Table 6, reveal a consistent and modest throughput reduction of about 12% across all hardware types. This suggests that, even when VLM inference is separated from the RL learner, the method maintains efficient scaling. In other words, the additional overhead introduced by VLM-RB remains limited in practical distributed settings.
+
+[p. 23 | section: F.4. Computational Overhead | type: Text]
+It is important to note that these measurements use a standard inference setup, without any aggressive optimizations such as TensorRT or quantization (e.g., INT8 or FP4). We anticipate that employing a dedicated serving stack, such as vLLM or TGI, would further close the speed gap between VLM-RB and the baseline.
+
+[p. 24 | section: G. Experiments | type: Text]
+In this section, we present the full results of the baselines on all environments.
+
+[p. 24 | section: G. Experiments | type: TableGroup]
+Table 7. Performance denotes the highest Average Success Rate (M = 5 seeds). Sample Efficiency tracks the steps required to reach the baseline's best performance. Wall-Clock Saving estimates the reduction in real training time, accounting for a 12% inference overhead per step. Alg. Task Baseline Performance (↑) Best ASR Sample Efficiency (↓) Steps to Base. Best Wall-Clock Saving (↑) Time vs. Baseline Scene-3 UER PER +0.0% (1.00/1.00) +0.0% (1.00/1.00) +54.6% (206K/454K) +25.4% (206K/276K) +49.2% +16.4% SAC Scene-4 UER PER +4.2% (1.00/0.96) +4.2% (1.00/0.96) +28.3% (592K/826K) +26.0% (592K/800K) +19.7% +17.1% Scene-5 UER PER +100.0% (0.88/0.44) +41.9% (0.88/0.62) +43.8% (514K/914K) +6.3% (708K/756K) +37.0% -4.8% Scene-3 UER PER +0.0% (1.00/1.00) +0.0% (1.00/1.00) +15.7% (214K/254K) +16.4% (214K/256K) +5.6% +6.4% TD3 Scene-4 UER PER +47.1% (1.00/0.68) +0.0% (1.00/1.00) +63.1% (268K/726K) +9.0% (426K/468K) +58.7% -1.9% Scene-5 UER PER +150.0% (0.70/0.28) +59.1% (0.70/0.44) +49.4% (366K/724K) +28.1% (614K/854K) +43.4% +19.5% DoorKey-8x8 UER PER +0.0% (1.00/1.00) +0.0% (1.00/1.00) +10.7% (150K/168K) +27.9% (150K/208K) +0.0% +19.2% DQN DoorKey-12x12 UER PER +66.7% (1.00/0.60) +0.0% (1.00/1.00) +62.6% (216K/578K) +6.1% (246K/262K) +58.1% -5.2% DoorKey-16x16 UER PER +316.7% (1.00/0.24) +25.0% (1.00/0.80) +62.9% (340K/916K) +35.5% (382K/592K) +58.4% +27.7% DoorKey-8x8 UER PER +0.0% (1.00/1.00) +0.0% (1.00/1.00) +26.6% (138K/188K) +16.9% (138K/166K) +17.8% +6.9% IQN DoorKey-12x12 UER PER +56.2% (1.00/0.64) +56.2% (1.00/0.64) +44.7% (388K/702K) +42.3% (388K/672K) +38.1% +35.3% DoorKey-16x16 UER PER +166.7% (0.64/0.24) +300.0% (0.64/0.16) +11.0% (774K/870K) +13.8% (562K/652K) +0.4% +3.5%
+
+[p. 24 | section: G. Experiments | type: Caption]
+Table 8. Performance denotes the highest Average Success Rate (M = 5 seeds). Sample Efficiency tracks the steps required to reach the baseline's best performance. Both metrics are averaged across the aggregated algorithms and tasks.
+
+[p. 24 | section: G. Experiments | type: Table]
+Env Type Agg. Algorithms Baseline Performance (↑) Mean Best ASR Sample Efficiency (↓) Mean Steps to Base Peak Scene (SAC + TD3) UER PER +28.0% (0.93/0.73) +11.2% (0.93/0.84) +44.6% (360K/650K) +19.1% (460K/568K) DoorKey (DQN + IQN) UER PER +51.6% (0.94/0.62) +22.6% (0.94/0.77) +41.4% (334K/570K) +26.9% (311K/425K)
+
+[p. 25 | section: G. Experiments | type: FigureGroup]
+Figure 9. VLM-RB consistently outperforms baselines across continuous and discrete tasks. The plots show aggregated success rates for four algorithms (DQN, IQN, SAC, TD3) on MiniGrid and OGBench domains. Annotations highlight the relative improvement in sample efficiency (horizontal arrows, reaching peak performance faster) and the best success rate (vertical arrows). Shaded regions indicate standard deviation across seeds.

icml26/c39d243f-0c59-4f70-8ee6-d9e742174491/assets.json

@@ -0,0 +1,134 @@
+[
+  {
+    "filename": "_page_20_Figure_7.jpeg",
+    "path": "data/processed_papers/icml26_20260429_1952_duequeue/marker_raw/c39d243f-0c59-4f70-8ee6-d9e742174491/marker_markdown/c39d243f-0c59-4f70-8ee6-d9e742174491/_page_20_Figure_7.jpeg",
+    "bytes": 1573,
+    "width": 53,
+    "height": 90,
+    "aspect_ratio": 0.5888888888888889,
+    "keep": false,
+    "reject_reason": "too_small",
+    "model_path": null
+  },
+  {
+    "filename": "_page_20_Figure_8.jpeg",
+    "path": "data/processed_papers/icml26_20260429_1952_duequeue/marker_raw/c39d243f-0c59-4f70-8ee6-d9e742174491/marker_markdown/c39d243f-0c59-4f70-8ee6-d9e742174491/_page_20_Figure_8.jpeg",
+    "bytes": 35803,
+    "width": 1146,
+    "height": 344,
+    "aspect_ratio": 3.3313953488372094,
+    "keep": true,
+    "reject_reason": null,
+    "model_path": "assets/_page_20_Figure_8.jpeg"
+  },
+  {
+    "filename": "_page_21_Figure_6.jpeg",
+    "path": "data/processed_papers/icml26_20260429_1952_duequeue/marker_raw/c39d243f-0c59-4f70-8ee6-d9e742174491/marker_markdown/c39d243f-0c59-4f70-8ee6-d9e742174491/_page_21_Figure_6.jpeg",
+    "bytes": 22146,
+    "width": 680,
+    "height": 436,
+    "aspect_ratio": 1.5596330275229358,
+    "keep": true,
+    "reject_reason": null,
+    "model_path": "assets/_page_21_Figure_6.jpeg"
+  },
+  {
+    "filename": "_page_22_Figure_2.jpeg",
+    "path": "data/processed_papers/icml26_20260429_1952_duequeue/marker_raw/c39d243f-0c59-4f70-8ee6-d9e742174491/marker_markdown/c39d243f-0c59-4f70-8ee6-d9e742174491/_page_22_Figure_2.jpeg",
+    "bytes": 3528,
+    "width": 73,
+    "height": 286,
+    "aspect_ratio": 0.25524475524475526,
+    "keep": false,
+    "reject_reason": "too_small",
+    "model_path": null
+  },
+  {
+    "filename": "_page_22_Figure_3.jpeg",
+    "path": "data/processed_papers/icml26_20260429_1952_duequeue/marker_raw/c39d243f-0c59-4f70-8ee6-d9e742174491/marker_markdown/c39d243f-0c59-4f70-8ee6-d9e742174491/_page_22_Figure_3.jpeg",
+    "bytes": 15125,
+    "width": 273,
+    "height": 324,
+    "aspect_ratio": 0.8425925925925926,
+    "keep": true,
+    "reject_reason": null,
+    "model_path": "assets/_page_22_Figure_3.jpeg"
+  },
+  {
+    "filename": "_page_22_Figure_4.jpeg",
+    "path": "data/processed_papers/icml26_20260429_1952_duequeue/marker_raw/c39d243f-0c59-4f70-8ee6-d9e742174491/marker_markdown/c39d243f-0c59-4f70-8ee6-d9e742174491/_page_22_Figure_4.jpeg",
+    "bytes": 27799,
+    "width": 587,
+    "height": 337,
+    "aspect_ratio": 1.741839762611276,
+    "keep": true,
+    "reject_reason": null,
+    "model_path": "assets/_page_22_Figure_4.jpeg"
+  },
+  {
+    "filename": "_page_22_Figure_5.jpeg",
+    "path": "data/processed_papers/icml26_20260429_1952_duequeue/marker_raw/c39d243f-0c59-4f70-8ee6-d9e742174491/marker_markdown/c39d243f-0c59-4f70-8ee6-d9e742174491/_page_22_Figure_5.jpeg",
+    "bytes": 26127,
+    "width": 249,
+    "height": 313,
+    "aspect_ratio": 0.7955271565495208,
+    "keep": true,
+    "reject_reason": null,
+    "model_path": "assets/_page_22_Figure_5.jpeg"
+  },
+  {
+    "filename": "_page_24_Figure_1.jpeg",
+    "path": "data/processed_papers/icml26_20260429_1952_duequeue/marker_raw/c39d243f-0c59-4f70-8ee6-d9e742174491/marker_markdown/c39d243f-0c59-4f70-8ee6-d9e742174491/_page_24_Figure_1.jpeg",
+    "bytes": 74312,
+    "width": 1109,
+    "height": 844,
+    "aspect_ratio": 1.3139810426540284,
+    "keep": true,
+    "reject_reason": null,
+    "model_path": "assets/_page_24_Figure_1.jpeg"
+  },
+  {
+    "filename": "_page_2_Figure_1.jpeg",
+    "path": "data/processed_papers/icml26_20260429_1952_duequeue/marker_raw/c39d243f-0c59-4f70-8ee6-d9e742174491/marker_markdown/c39d243f-0c59-4f70-8ee6-d9e742174491/_page_2_Figure_1.jpeg",
+    "bytes": 30920,
+    "width": 670,
+    "height": 301,
+    "aspect_ratio": 2.2259136212624586,
+    "keep": true,
+    "reject_reason": null,
+    "model_path": "assets/_page_2_Figure_1.jpeg"
+  },
+  {
+    "filename": "_page_4_Figure_2.jpeg",
+    "path": "data/processed_papers/icml26_20260429_1952_duequeue/marker_raw/c39d243f-0c59-4f70-8ee6-d9e742174491/marker_markdown/c39d243f-0c59-4f70-8ee6-d9e742174491/_page_4_Figure_2.jpeg",
+    "bytes": 67022,
+    "width": 1348,
+    "height": 519,
+    "aspect_ratio": 2.597302504816956,
+    "keep": true,
+    "reject_reason": null,
+    "model_path": "assets/_page_4_Figure_2.jpeg"
+  },
+  {
+    "filename": "_page_5_Figure_2.jpeg",
+    "path": "data/processed_papers/icml26_20260429_1952_duequeue/marker_raw/c39d243f-0c59-4f70-8ee6-d9e742174491/marker_markdown/c39d243f-0c59-4f70-8ee6-d9e742174491/_page_5_Figure_2.jpeg",
+    "bytes": 28015,
+    "width": 631,
+    "height": 425,
+    "aspect_ratio": 1.484705882352941,
+    "keep": true,
+    "reject_reason": null,
+    "model_path": "assets/_page_5_Figure_2.jpeg"
+  },
+  {
+    "filename": "_page_6_Figure_1.jpeg",
+    "path": "data/processed_papers/icml26_20260429_1952_duequeue/marker_raw/c39d243f-0c59-4f70-8ee6-d9e742174491/marker_markdown/c39d243f-0c59-4f70-8ee6-d9e742174491/_page_6_Figure_1.jpeg",
+    "bytes": 34562,
+    "width": 635,
+    "height": 516,
+    "aspect_ratio": 1.2306201550387597,
+    "keep": true,
+    "reject_reason": null,
+    "model_path": "assets/_page_6_Figure_1.jpeg"
+  }
+]

icml26/c39d243f-0c59-4f70-8ee6-d9e742174491/dataset_meta.json

@@ -0,0 +1,63 @@
+{
+  "paper_id": "c39d243f-0c59-4f70-8ee6-d9e742174491",
+  "pipeline": "Paper2Markdown-V3",
+  "ok": true,
+  "page_count": 25,
+  "chunk_count": 183,
+  "main_body_chunk_count": 81,
+  "appendix_chunk_count": 91,
+  "reference_chunk_count": 11,
+  "model_text_chars": 41640,
+  "raw_markdown_chars": 104035,
+  "sanitized_chars": 93277,
+  "page_provenance": {
+    "min_page": 1,
+    "max_page": 25,
+    "invalid_count": 0
+  },
+  "marker_block_type_counts": {
+    "Caption": 9,
+    "Code": 1,
+    "Equation": 10,
+    "Figure": 6,
+    "FigureGroup": 6,
+    "Footnote": 5,
+    "ListGroup": 8,
+    "PageFooter": 25,
+    "PageHeader": 27,
+    "SectionHeader": 40,
+    "Table": 3,
+    "TableGroup": 5,
+    "Text": 488
+  },
+  "asset_count_raw": 12,
+  "asset_count_model_kept": 10,
+  "asset_count_rejected": 2,
+  "asset_reject_reasons": {
+    "kept": 10,
+    "too_small": 2
+  },
+  "artifact_leak_audit": {
+    "ok": true,
+    "hits": {
+      "Anonymous Authors": [],
+      "ACKNOWLEDGMENT": [],
+      "OpenReview": [],
+      "\"accept_label\"": [],
+      "\"decision\"": [],
+      "\"decision_tier\"": [],
+      "\"source_status\"": [],
+      "Meta-review": [],
+      "Official Review": [],
+      "official_reviews": [],
+      "meta_reviews": [],
+      "suggested_verdict_score": []
+    },
+    "artifact_count": 2
+  },
+  "default_model_input": "model_text_v3.txt",
+  "appendix_input": "appendix_text_v3.txt",
+  "reference_input": "reference_text_v3.txt",
+  "source": "koala_icml26_due_queue",
+  "run_name": "icml26_20260429_1952_duequeue"
+}

icml26/c39d243f-0c59-4f70-8ee6-d9e742174491/main_body_chunks.jsonl

@@ -0,0 +1,81 @@
+{"paper_id": "c39d243f-0c59-4f70-8ee6-d9e742174491", "chunk_id": "c39d243f-0c59-4f70-8ee6-d9e742174491:0000", "section": "Abstract", "page_start": 1, "page_end": 1, "type": "Text", "text": "Recent advances in Large Language Models (LLMs) and Vision-Language Models (VLMs) have enabled powerful semantic and multimodal reasoning capabilities, creating new opportunities to enhance sample efficiency, high-level planning, and interpretability in reinforcement learning (RL). While prior work has integrated LLMs and VLMs into various components of RL, the replay buffer, a core component for storing and reusing experiences, remains unexplored. We propose addressing this gap by leveraging VLMs to guide the prioritization of experiences in the replay buffer. Our key idea is to use a frozen, pre-trained VLM (requiring no fine-tuning) as an automated evaluator to identify and prioritize promising sub-trajectories from the agent's experiences. Across scenarios, including game-playing and robotics, spanning both discrete and continuous domains, agents trained with our proposed prioritization method achieve 11–52% higher average success rates and improve sample efficiency by 19–45% compared to previous approaches.", "source": "marker_v2", "marker_block_id": "/page/0/Text/3"}
+{"paper_id": "c39d243f-0c59-4f70-8ee6-d9e742174491", "chunk_id": "c39d243f-0c59-4f70-8ee6-d9e742174491:0001", "section": "1. Introduction", "page_start": 1, "page_end": 1, "type": "Text", "text": "Reinforcement learning (RL) has achieved impressive results across domains ranging from robotics to natural language systems (Levine et al., 2016; Ouyang et al., 2022) and logistics management (Tang et al., 2025) . Despite this success, RL remains notoriously sample-inefficient, sensitive to hyperparameters, and heavily dependent on carefully designed reward functions (Kalashnikov et al., 2021; Kroemer et al., 2021; Zhu et al., 2020) .", "source": "marker_v2", "marker_block_id": "/page/0/Text/5"}
+{"paper_id": "c39d243f-0c59-4f70-8ee6-d9e742174491", "chunk_id": "c39d243f-0c59-4f70-8ee6-d9e742174491:0002", "section": "1. Introduction", "page_start": 1, "page_end": 1, "type": "Text", "text": "A well-established line of research—off-policy RL (Lin, 1992; Mnih et al., 2013) —addresses the efficiency bottleneck through experience reuse. During training, transitions", "source": "marker_v2", "marker_block_id": "/page/0/Text/6"}
+{"paper_id": "c39d243f-0c59-4f70-8ee6-d9e742174491", "chunk_id": "c39d243f-0c59-4f70-8ee6-d9e742174491:0003", "section": "1. Introduction", "page_start": 1, "page_end": 1, "type": "Text", "text": "are collected into a dataset known as a replay buffer , and sampled multiple times to update the policy. Ideally, a sampling mechanism would prioritize the most meaningful transitions: those that provide the richest learning signal, irrespective of when they were collected. The core challenge is to identify which experiences are truly meaningful for learning process.", "source": "marker_v2", "marker_block_id": "/page/0/Text/9"}
+{"paper_id": "c39d243f-0c59-4f70-8ee6-d9e742174491", "chunk_id": "c39d243f-0c59-4f70-8ee6-d9e742174491:0004", "section": "1. Introduction", "page_start": 1, "page_end": 1, "type": "Text", "text": "Conventional methods such as Prioritized Experience Replay (PER, Schaul et al. 2015b) approximate this ideal by prioritizing transitions with high temporal-difference (TD) error. The TD error measures the discrepancy between a bootstrapped value estimate and its target (Sutton et al., 1998) . In effect, this approach promotes transitions which are misaligned with the agent's future outcome predictions. While TD error prioritization can correct value estimates, it is fundamentally limited by its lack of semantic awareness. It cannot distinguish transitions which reflect genuine progress toward task completion from those that do not.", "source": "marker_v2", "marker_block_id": "/page/0/Text/10"}
+{"paper_id": "c39d243f-0c59-4f70-8ee6-d9e742174491", "chunk_id": "c39d243f-0c59-4f70-8ee6-d9e742174491:0005", "section": "1. Introduction", "page_start": 1, "page_end": 1, "type": "Text", "text": "This limitation is particularly evident in long-horizon, sparse-reward tasks, such as robotic manipulation in OG-Bench (Park et al., 2024) . For instance, critical transitions—such as unlatching a door or grasping a tool—may yield low TD errors early in training. This is a consequence of delayed credit assignment preventing the reward signal from propagating to these earlier states. Conversely, visually distinct but task-irrelevant motions can yield large TD errors, despite contributing little to actual task completion.", "source": "marker_v2", "marker_block_id": "/page/0/Text/11"}
+{"paper_id": "c39d243f-0c59-4f70-8ee6-d9e742174491", "chunk_id": "c39d243f-0c59-4f70-8ee6-d9e742174491:0006", "section": "1. Introduction", "page_start": 1, "page_end": 1, "type": "Text", "text": "To bridge this gap, we propose replacing heuristic inductive biases with external sources of semantic understanding. Recent literature highlights the potential of integrating largescale Vision-Language Models (VLMs) with RL to leverage pre-trained world knowledge (Liang et al., 2024; Zhang et al., 2024) . Because VLMs combine visual perception with language reasoning, they can interpret complex environments and mitigate RL-specific challenges—such as sparse rewards or inefficient exploration—by providing enriched context (Cao et al., 2024; Schoepp et al., 2025) . This motivates our primary research question:", "source": "marker_v2", "marker_block_id": "/page/0/Text/12"}
+{"paper_id": "c39d243f-0c59-4f70-8ee6-d9e742174491", "chunk_id": "c39d243f-0c59-4f70-8ee6-d9e742174491:0007", "section": "1. Introduction", "page_start": 1, "page_end": 1, "type": "Text", "text": "Can the semantic knowledge of pre-trained VLMs be used to prioritize meaningful experiences in replay buffers?", "source": "marker_v2", "marker_block_id": "/page/0/Text/13"}
+{"paper_id": "c39d243f-0c59-4f70-8ee6-d9e742174491", "chunk_id": "c39d243f-0c59-4f70-8ee6-d9e742174491:0008", "section": "1. Introduction", "page_start": 1, "page_end": 1, "type": "Text", "text": "To answer this, we introduce VLM-RB (Fig .1) , which integrates a pre-trained VLM directly into the experi-", "source": "marker_v2", "marker_block_id": "/page/0/Text/14"}
+{"paper_id": "c39d243f-0c59-4f70-8ee6-d9e742174491", "chunk_id": "c39d243f-0c59-4f70-8ee6-d9e742174491:0009", "section": "1. Introduction", "page_start": 2, "page_end": 2, "type": "Caption", "text": "Figure 1. System Diagram: (1) Data is collected with the current policy \\pi_k interacting with multiple instances of the environment. (2) Transitions (s, a, r, s') are kept in a prioritized replay buffer with a default priority \\bar{\\mathbf{p}} . (3) Asynchronously, after each insertion, a VLM worker scores the corresponding rendered clip \\tau^O under a prompt P and writes the resulting priority \\mathbf{p^{VLM}} back to the replay buffer (see Section 3.1). (4) The learner samples a minibatch using the mixture distribution q_t which interpolates between uniform and VLM-prioritized replay (see Section 3.1). (5) The agent policy is updated to \\pi_{k+1} .", "source": "marker_v2", "marker_block_id": "/page/1/Caption/2"}
+{"paper_id": "c39d243f-0c59-4f70-8ee6-d9e742174491", "chunk_id": "c39d243f-0c59-4f70-8ee6-d9e742174491:0010", "section": "1. Introduction", "page_start": 2, "page_end": 2, "type": "Text", "text": "ence prioritization pipeline. In contrast to heuristics relying on statistical proxies such as TD error, uncertainty (Carrasco-Davis et al., 2025), or density ratios (Zhao & Tresp, 2019), VLM-RB leverages the reasoning capabilities of VLMs to score experiences according to their semantic relevance.", "source": "marker_v2", "marker_block_id": "/page/1/Text/3"}
+{"paper_id": "c39d243f-0c59-4f70-8ee6-d9e742174491", "chunk_id": "c39d243f-0c59-4f70-8ee6-d9e742174491:0011", "section": "1. Introduction", "page_start": 2, "page_end": 2, "type": "Text", "text": "We design this framework as a modular plug-and-play prioritization layer that can be integrated with any off-policy algorithm that uses a replay buffer. By using frozen off-the-shelf VLMs, we leverage strong semantic priors while avoiding the computational cost of fine-tuning or additional inference latency at deployment. Within the training pipeline, VLM-RB promotes efficiency by querying the VLM asynchronously, thereby sidestepping the blocking inference bottleneck that often constrains the throughput of LLM/VLM-enhanced RL methods.", "source": "marker_v2", "marker_block_id": "/page/1/Text/4"}
+{"paper_id": "c39d243f-0c59-4f70-8ee6-d9e742174491", "chunk_id": "c39d243f-0c59-4f70-8ee6-d9e742174491:0012", "section": "1. Introduction", "page_start": 2, "page_end": 2, "type": "Text", "text": "We empirically show that pre-trained VLMs yield semantically grounded prioritization signals which are closely aligned with task objectives. Leveraging this semantic guidance, VLM-RB consistently outperforms existing baselines in both discrete game-playing and continuous robotic manipulation. Importantly, the moderate throughput overhead (about 12%) is more than offset by substantial gains in sample efficiency (19–45%) and average success rates (11–52%).", "source": "marker_v2", "marker_block_id": "/page/1/Text/5"}
+{"paper_id": "c39d243f-0c59-4f70-8ee6-d9e742174491", "chunk_id": "c39d243f-0c59-4f70-8ee6-d9e742174491:0013", "section": "2. Preliminaries", "page_start": 2, "page_end": 2, "type": "Text", "text": "Off-Policy Reinforcement Learning. We consider the standard RL framework, where an agent interacts with an environment modeled as an infinite-horizon Markov Decision Process (MDP) \\mathcal{M} = (\\mathcal{S}, \\mathcal{A}, P, r, \\gamma, \\rho_0) , with state space \\mathcal{S} , action space \\mathcal{A} (either discrete or continuous),", "source": "marker_v2", "marker_block_id": "/page/1/Text/7"}
+{"paper_id": "c39d243f-0c59-4f70-8ee6-d9e742174491", "chunk_id": "c39d243f-0c59-4f70-8ee6-d9e742174491:0014", "section": "2. Preliminaries", "page_start": 2, "page_end": 2, "type": "Text", "text": "transition kernel P: \\mathcal{S} \\times \\mathcal{A} \\to \\Delta(\\mathcal{S}) , reward function r: \\mathcal{S} \\times \\mathcal{A} \\to \\mathbb{R} , discount factor \\gamma \\in (0,1) and initial-state distribution \\rho_0 \\in \\Delta(\\mathcal{S}) . We further assume access to a rendering function \\psi: \\mathcal{S} \\to \\mathcal{O} \\subseteq \\mathbb{R}^{H \\times W \\times 3} which maps states to H \\times W -sized visual observations o_t = \\psi(s_t) . The agent learns a stochastic policy \\pi_\\theta: \\mathcal{S} \\to \\Delta(\\mathcal{A}) to maximize the expected discounted return", "source": "marker_v2", "marker_block_id": "/page/1/Text/8"}
+{"paper_id": "c39d243f-0c59-4f70-8ee6-d9e742174491", "chunk_id": "c39d243f-0c59-4f70-8ee6-d9e742174491:0015", "section": "2. Preliminaries", "page_start": 2, "page_end": 2, "type": "Equation", "text": "J(\\pi) = \\mathbb{E}_{\\tau \\sim p_{\\pi}(\\tau)} \\left[ \\sum_{t=0}^{\\infty} \\gamma^{t} r_{t} \\right],", "source": "marker_v2", "marker_block_id": "/page/1/Equation/9"}
+{"paper_id": "c39d243f-0c59-4f70-8ee6-d9e742174491", "chunk_id": "c39d243f-0c59-4f70-8ee6-d9e742174491:0016", "section": "2. Preliminaries", "page_start": 2, "page_end": 2, "type": "Text", "text": "where \\tau = (s_0, a_0, r_0, s_1, a_1, r_1, ...) denotes a trajectory and p_{\\pi}(\\tau) is the trajectory distribution induced by policy \\pi .", "source": "marker_v2", "marker_block_id": "/page/1/Text/10"}
+{"paper_id": "c39d243f-0c59-4f70-8ee6-d9e742174491", "chunk_id": "c39d243f-0c59-4f70-8ee6-d9e742174491:0017", "section": "2. Preliminaries", "page_start": 2, "page_end": 2, "type": "Text", "text": "To evaluate and improve policies, value functions are used to estimate expected returns. The action-value function (Q-function) Q^{\\pi}: \\mathcal{S} \\times \\mathcal{A} \\to \\mathbb{R} represents the expected discounted return from state s after taking action a and then following policy \\pi :", "source": "marker_v2", "marker_block_id": "/page/1/Text/11"}
+{"paper_id": "c39d243f-0c59-4f70-8ee6-d9e742174491", "chunk_id": "c39d243f-0c59-4f70-8ee6-d9e742174491:0018", "section": "2. Preliminaries", "page_start": 2, "page_end": 2, "type": "Equation", "text": "Q^{\\pi}(s, a) = \\mathbb{E}_{\\tau \\sim p_{\\pi}(\\tau \\mid s_0 = s, a_0 = a)} \\left[ \\sum_{t=0}^{\\infty} \\gamma^t r_t \\right].", "source": "marker_v2", "marker_block_id": "/page/1/Equation/12"}
+{"paper_id": "c39d243f-0c59-4f70-8ee6-d9e742174491", "chunk_id": "c39d243f-0c59-4f70-8ee6-d9e742174491:0019", "section": "2. Preliminaries", "page_start": 2, "page_end": 2, "type": "Text", "text": "In modern RL methods, Q^{\\pi} is typically a learned function (represented by a neural network), denoted the critic .", "source": "marker_v2", "marker_block_id": "/page/1/Text/13"}
+{"paper_id": "c39d243f-0c59-4f70-8ee6-d9e742174491", "chunk_id": "c39d243f-0c59-4f70-8ee6-d9e742174491:0020", "section": "2. Preliminaries", "page_start": 2, "page_end": 2, "type": "Text", "text": "In off-policy RL, rather than exclusively using on-policy data, optimization is performed using previously collected experience stored in a replay buffer (Mnih et al., 2013). Let \\mathcal{D}_t denote the replay buffer at time t. We sample training tuples (s, a, r, s') \\in \\mathcal{D}_t according to a time-dependent distribution q_t \\in \\Delta(\\mathcal{D}_t) . In standard uniform replay (UER), q_t(i) = 1/|\\mathcal{D}_t| for all i \\in \\mathcal{D}_t . This formulation highlights that the optimization dynamics depend on the evolving sampling distribution q_t .", "source": "marker_v2", "marker_block_id": "/page/1/Text/14"}
+{"paper_id": "c39d243f-0c59-4f70-8ee6-d9e742174491", "chunk_id": "c39d243f-0c59-4f70-8ee6-d9e742174491:0021", "section": "2. Preliminaries", "page_start": 3, "page_end": 3, "type": "Caption", "text": "Figure 2. Temporal context resolves visual ambiguity. From the initial state o_i (left), multiple futures are possible. The top sub-trajectory shows a successful grasp, while the bottom shows stagnation. By scoring sub-trajectories rather than single frames, the VLM has sufficient context to distinguish meaningful progress from failure modes.", "source": "marker_v2", "marker_block_id": "/page/2/Caption/2"}
+{"paper_id": "c39d243f-0c59-4f70-8ee6-d9e742174491", "chunk_id": "c39d243f-0c59-4f70-8ee6-d9e742174491:0022", "section": "2. Preliminaries", "page_start": 3, "page_end": 3, "type": "Text", "text": "Experience Replay Prioritization. As opposed to uniform replay, prioritized replay methods bias retrieval toward transitions expected to yield a stronger learning signal. For clarity, we formalize this process as three distinct steps: scoring, prioritization, and sampling. Scoring assigns each transition i \\in \\mathcal{D}_t a scalar value \\mathbf{p}_i \\in \\mathbb{R} reflecting its heuristic utility. Prioritization maps these raw scores into a probability distribution over the buffer, q_t^{\\mathbf{P}} \\in \\Delta(\\mathcal{D}_t) . Sampling then draws a minibatch of indices \\{i_k\\}_{k=1}^B from a target distribution q_t , or a mixture of several such distributions.", "source": "marker_v2", "marker_block_id": "/page/2/Text/3"}
+{"paper_id": "c39d243f-0c59-4f70-8ee6-d9e742174491", "chunk_id": "c39d243f-0c59-4f70-8ee6-d9e742174491:0023", "section": "2. Preliminaries", "page_start": 3, "page_end": 3, "type": "Text", "text": "Prioritized Experience Replay ( PER , Schaul et al. 2015b) is a specific instance of this framework. Its scoring function utilizes the temporal-difference (TD) error magnitude, \\mathbf{p}_i = |\\delta_i| + \\epsilon^1 , where \\delta_i is the discrepancy between the value estimate and its target<sup>2</sup>. The prioritized distribution is defined as q_t^{\\mathbf{P}}(i) \\propto \\mathbf{p}_i^{\\alpha} , where \\alpha \\in [0,1] determines the degree of prioritization. While PER effectively focuses updates on transitions with larger TD errors, it relies on value estimates which may be arbitrary and noisy early in training, particularly in sparse-reward settings.", "source": "marker_v2", "marker_block_id": "/page/2/Text/4"}
+{"paper_id": "c39d243f-0c59-4f70-8ee6-d9e742174491", "chunk_id": "c39d243f-0c59-4f70-8ee6-d9e742174491:0024", "section": "2. Preliminaries", "page_start": 3, "page_end": 3, "type": "Text", "text": "Vision–Language Models. Recent advances in large-scale representation learning have given rise to Vision–Language Models (VLMs), which learn aligned embeddings of visual and textual modalities through joint pre-training on large-scale image–text corpora (Radford et al., 2021; Jia et al., 2021; Zhai et al., 2023). Formally, a VLM defines encoders f_{\\text{img}}: \\mathcal{I} \\to \\mathbb{R}^d and f_{\\text{txt}}: \\mathcal{T} \\to \\mathbb{R}^d mapping images and text to a shared latent space, typically optimized via contrastive objectives which encourage semantic correspondence between paired inputs (Cherti et al., 2023; Zhai et al., 2022; Goel et al., 2022). These models have demonstrated strong generalization across a wide", "source": "marker_v2", "marker_block_id": "/page/2/Text/5"}
+{"paper_id": "c39d243f-0c59-4f70-8ee6-d9e742174491", "chunk_id": "c39d243f-0c59-4f70-8ee6-d9e742174491:0025", "section": "2. Preliminaries", "page_start": 3, "page_end": 3, "type": "Text", "text": "range of downstream tasks, including zero-shot recognition, goal specification, and reward inference (Ahn et al., 2022; Rocamonde et al., 2023). While standard VLMs process static images, Video Question Answering (Video-QA) models extend these capabilities to ingest a temporal sequence of frames (a video clip) along with a natural-language query. By aggregating visual information across time, these models can reason about events, motion, and causal relationships. Concretely, for the rest of this paper, we employ Perception-LM (Cho et al., 2025), a state-of-the-art opensource Video-QA family of models with multiple model sizes (1B/3B/8B).", "source": "marker_v2", "marker_block_id": "/page/2/Text/8"}
+{"paper_id": "c39d243f-0c59-4f70-8ee6-d9e742174491", "chunk_id": "c39d243f-0c59-4f70-8ee6-d9e742174491:0026", "section": "3. VLM-RB", "page_start": 3, "page_end": 3, "type": "Text", "text": "In this section, we present VLM-RB (Fig.1 and Algorithm 1), a plug-and-play framework designed for any off-policy algorithm that leverages a replay buffer. The central idea is to use a VLM to assign semantic scores to subtrajectories, subsequently biasing the sampling distribution toward higher-scoring experiences during policy optimization.", "source": "marker_v2", "marker_block_id": "/page/2/Text/10"}
+{"paper_id": "c39d243f-0c59-4f70-8ee6-d9e742174491", "chunk_id": "c39d243f-0c59-4f70-8ee6-d9e742174491:0027", "section": "3.1. Scoring, Prioritization, and Sampling", "page_start": 3, "page_end": 3, "type": "Text", "text": "We now define how semantic scores are extracted from collected data, how these scores induce a prioritization distribution, and how VLM-RB samples from this distribution.", "source": "marker_v2", "marker_block_id": "/page/2/Text/12"}
+{"paper_id": "c39d243f-0c59-4f70-8ee6-d9e742174491", "chunk_id": "c39d243f-0c59-4f70-8ee6-d9e742174491:0028", "section": "3.1. Scoring, Prioritization, and Sampling", "page_start": 3, "page_end": 3, "type": "Text", "text": "Scoring. VLM-RB uses a pre-trained frozen VLM to extract semantic scores from collected experiences. To enable this, we first construct temporal visual sequences from the agent's history. Formally, let \\tau_i^O = (o_i, o_{i+1}, \\dots, o_{i+L-1}) denote a visual clip comprising L rendered frames. The VLM maps this clip and a text prompt P to a scalar score:", "source": "marker_v2", "marker_block_id": "/page/2/Text/13"}
+{"paper_id": "c39d243f-0c59-4f70-8ee6-d9e742174491", "chunk_id": "c39d243f-0c59-4f70-8ee6-d9e742174491:0029", "section": "3.1. Scoring, Prioritization, and Sampling", "page_start": 3, "page_end": 3, "type": "Equation", "text": "\\mathbf{p}^{\\mathbf{VLM}} = f_{\\mathbf{VLM}}(\\tau^O, \\mathsf{P}) \\in \\mathbb{R}. (1)", "source": "marker_v2", "marker_block_id": "/page/2/Equation/14"}
+{"paper_id": "c39d243f-0c59-4f70-8ee6-d9e742174491", "chunk_id": "c39d243f-0c59-4f70-8ee6-d9e742174491:0030", "section": "3.1. Scoring, Prioritization, and Sampling", "page_start": 3, "page_end": 3, "type": "Text", "text": "The prompt P directs the VLM's scoring mechanism, leveraging the model's inherent world knowledge to identify meaningful behaviors even in the absence of dense reward signals. While this interface supports the injection of detailed and task-specific priors, we find that the VLM's intrinsic scene understanding is sufficient for our purposes. We therefore employ a general task-agnostic prompt (see Appendix C for details).", "source": "marker_v2", "marker_block_id": "/page/2/Text/15"}
+{"paper_id": "c39d243f-0c59-4f70-8ee6-d9e742174491", "chunk_id": "c39d243f-0c59-4f70-8ee6-d9e742174491:0031", "section": "3.1. Scoring, Prioritization, and Sampling", "page_start": 3, "page_end": 3, "type": "Text", "text": "For simplicity, we set f_{\\rm VLM} as a binary indicator, assigning a score of 1 to clips exhibiting seemingly meaningful behavior, and 0 otherwise. This effectively labels useful sequences without the need for hand-crafted task-specific definitions. Crucially, because the VLM is frozen and its evaluation depends only on static visual content, each clip is scored exactly once. This offers a significant efficiency advantage over methods such as PER , which require ongoing priority updates as the value function evolves.", "source": "marker_v2", "marker_block_id": "/page/2/Text/16"}
+{"paper_id": "c39d243f-0c59-4f70-8ee6-d9e742174491", "chunk_id": "c39d243f-0c59-4f70-8ee6-d9e742174491:0032", "section": "3.1. Scoring, Prioritization, and Sampling", "page_start": 3, "page_end": 3, "type": "Footnote", "text": "& lt;sup>1</sup>Here \\epsilon > 0 ensures non-zero probability.", "source": "marker_v2", "marker_block_id": "/page/2/Footnote/6"}
+{"paper_id": "c39d243f-0c59-4f70-8ee6-d9e742174491", "chunk_id": "c39d243f-0c59-4f70-8ee6-d9e742174491:0033", "section": "3.1. Scoring, Prioritization, and Sampling", "page_start": 3, "page_end": 3, "type": "Footnote", "text": "& lt;sup>2</sup>Namely, \\delta_i = Q_{\\theta}(s_t, a_t) - (r + \\gamma Q_{\\theta}(s_{t+1}, a')) where \\theta are the critic parameters and a' is the action predicted by the current policy in s_{t+1} . In Q-learning methods, a' is the argmax action, and in actor-critic methods for continuous control, it is \\pi(s_{t+1}) .", "source": "marker_v2", "marker_block_id": "/page/2/Footnote/7"}
+{"paper_id": "c39d243f-0c59-4f70-8ee6-d9e742174491", "chunk_id": "c39d243f-0c59-4f70-8ee6-d9e742174491:0034", "section": "3.1. Scoring, Prioritization, and Sampling", "page_start": 4, "page_end": 4, "type": "Text", "text": "A natural question is whether scoring individual frames, rather than clips, would suffice. We argue that single frames are fundamentally limited by semantic ambiguity . To illustrate this, consider Fig. 2: a single observation of a robotic gripper hovering above an object. This static frame is visually identical across two vastly different temporal contexts: the onset of a successful grasp or the aftermath of a failed attempt.", "source": "marker_v2", "marker_block_id": "/page/3/Text/1"}
+{"paper_id": "c39d243f-0c59-4f70-8ee6-d9e742174491", "chunk_id": "c39d243f-0c59-4f70-8ee6-d9e742174491:0035", "section": "3.1. Scoring, Prioritization, and Sampling", "page_start": 4, "page_end": 4, "type": "Text", "text": "Without temporal context a VLM cannot disambiguate these scenarios, potentially assigning high scores to frames which actually belong to failure modes. By scoring sub-trajectories instead, we increase the likelihood of the VLM having sufficient temporal information to distinguish meaningful behaviors from failures. While a sliding-window variant could potentially provide denser labels, we leave this (computationally intensive) alternative for future work.", "source": "marker_v2", "marker_block_id": "/page/3/Text/2"}
+{"paper_id": "c39d243f-0c59-4f70-8ee6-d9e742174491", "chunk_id": "c39d243f-0c59-4f70-8ee6-d9e742174491:0036", "section": "3.1. Scoring, Prioritization, and Sampling", "page_start": 4, "page_end": 4, "type": "Text", "text": "Prioritization. We construct a prioritized distribution q P by propagating the VLM score of each clip to all transitions within that clip. Under our binary scoring scheme (p i ∈ {0, 1}), q P becomes uniform over the subset of transitions labeled as semantically meaningful.", "source": "marker_v2", "marker_block_id": "/page/3/Text/3"}
+{"paper_id": "c39d243f-0c59-4f70-8ee6-d9e742174491", "chunk_id": "c39d243f-0c59-4f70-8ee6-d9e742174491:0037", "section": "3.1. Scoring, Prioritization, and Sampling", "page_start": 4, "page_end": 4, "type": "Text", "text": "Sampling. If we were to sample only from q P , we would discard all transitions labeled uninteresting by the VLM. To avoid wasting collected data and to ensure the agent explores the full state space, we instead use a mixture strategy q t interpolating between VLM-guided prioritization and uniform replay:", "source": "marker_v2", "marker_block_id": "/page/3/Text/4"}
+{"paper_id": "c39d243f-0c59-4f70-8ee6-d9e742174491", "chunk_id": "c39d243f-0c59-4f70-8ee6-d9e742174491:0038", "section": "3.1. Scoring, Prioritization, and Sampling", "page_start": 4, "page_end": 4, "type": "Equation", "text": "q_t(i) = \\lambda_t q^{\\mathbf{P}}(i) + (1 - \\lambda_t) q^{\\mathbf{U}}(i), \\tag{2}", "source": "marker_v2", "marker_block_id": "/page/3/Equation/5"}
+{"paper_id": "c39d243f-0c59-4f70-8ee6-d9e742174491", "chunk_id": "c39d243f-0c59-4f70-8ee6-d9e742174491:0039", "section": "3.1. Scoring, Prioritization, and Sampling", "page_start": 4, "page_end": 4, "type": "Text", "text": "where λ t ∈ [0, 1] controls the strength of the VLM guidance. In practice, each batch draws a λ t fraction from q P and the remainder uniformly. We use a linear warm-up schedule: starting with λ 0 = 0 (pure uniform sampling), we gradually anneal to λmax = 0.5 over the first half of training. We hypothesize that this schedule is essential: early in training, broad coverage stabilizes value learning, while later updates bias toward high-utility regions. Our ablation studies (Section 4.4) support this design: purely prioritized sampling is detrimental, but the proposed mixture schedule yields significant efficiency gains.", "source": "marker_v2", "marker_block_id": "/page/3/Text/6"}
+{"paper_id": "c39d243f-0c59-4f70-8ee6-d9e742174491", "chunk_id": "c39d243f-0c59-4f70-8ee6-d9e742174491:0040", "section": "3.2. Efficient Implementation", "page_start": 4, "page_end": 4, "type": "Text", "text": "Incorporating a VLM into the RL loop introduces significant computational overhead relative to standard components. We address this with two key design choices. First, we leverage the decoupling between data collection and policy optimization in off-policy learning to asynchronously score experiences. As illustrated in Fig. 1, the VLM interacts with the replay buffer in the background, ensuring that policy optimization is never blocked by inference latency.", "source": "marker_v2", "marker_block_id": "/page/3/Text/8"}
+{"paper_id": "c39d243f-0c59-4f70-8ee6-d9e742174491", "chunk_id": "c39d243f-0c59-4f70-8ee6-d9e742174491:0041", "section": "3.2. Efficient Implementation", "page_start": 4, "page_end": 4, "type": "Text", "text": "Beyond minimizing latency, this decoupled architecture renders VLM-RB agnostic to the policy input modality: while the VLM requires rendered frames, the policy itself can operate on arbitrary observation spaces such as lowdimensional states, enabling more sample-efficient learning. This architecture further allows a single VLM instance to efficiently serve multiple parallel environments. Second, we use a lightweight 1B parameter model (Cho et al., 2025) . Our ablations show that this model is sufficient to identify meaningful behaviors while maintaining high throughput 3 (see Appendix F.2) .", "source": "marker_v2", "marker_block_id": "/page/3/Text/9"}
+{"paper_id": "c39d243f-0c59-4f70-8ee6-d9e742174491", "chunk_id": "c39d243f-0c59-4f70-8ee6-d9e742174491:0042", "section": "3.3. Boosting with TD-error", "page_start": 4, "page_end": 4, "type": "Text", "text": "We can further refine prioritization by incorporating the TD-error, defining q P (i) ∝ p VLM i · |δ i |, where δ i emphasizes transitions with high prediction error (where the value function is inaccurate), and p VLM i emphasizes semantic relevance. Because the VLM score is a binary indicator, it effectively masks \"irrelevant\" transitions and promotes the remaining transitions based on their TD errors.", "source": "marker_v2", "marker_block_id": "/page/3/Text/11"}
+{"paper_id": "c39d243f-0c59-4f70-8ee6-d9e742174491", "chunk_id": "c39d243f-0c59-4f70-8ee6-d9e742174491:0043", "section": "3.3. Boosting with TD-error", "page_start": 4, "page_end": 4, "type": "Text", "text": "In this scheme, we maintain two scores per transition in the buffer: q P (i) ∝ p VLM i · |δ i | and δ i . The TD error δ i is updated each time the transition is sampled, while q P (i) ∝ p i · |δ i | is updated only once, as described in Section 3.2.", "source": "marker_v2", "marker_block_id": "/page/3/Text/12"}
+{"paper_id": "c39d243f-0c59-4f70-8ee6-d9e742174491", "chunk_id": "c39d243f-0c59-4f70-8ee6-d9e742174491:0044", "section": "4. Experiments", "page_start": 4, "page_end": 4, "type": "Text", "text": "In this section, we evaluate the efficacy of VLM-RB in leveraging semantic priors for efficient exploration. We first ask: does the VLM signal provide meaningful guidance for exploration? To answer this, we analyze its correlation with learned value estimates and its dependence on visual semantics (Section 4.2) . Next, we benchmark VLM-RB against UER, PER, and alternative prioritization schemes, evaluating both performance and sample efficiency across a range of discrete and continuous control tasks (Section 4.3) . Finally, we examine which design choices most affect the performance of VLM-RB, focusing on the sampling mixture and VLM model size (Section 4.4) .", "source": "marker_v2", "marker_block_id": "/page/3/Text/14"}
+{"paper_id": "c39d243f-0c59-4f70-8ee6-d9e742174491", "chunk_id": "c39d243f-0c59-4f70-8ee6-d9e742174491:0045", "section": "4.1. Experimental Setup", "page_start": 4, "page_end": 4, "type": "Text", "text": "Tasks. We evaluate VLM-RB on two domains with discrete and continuous action spaces: (i) DoorKey from MiniGrid (Chevalier-Boisvert et al., 2023) , using grid sizes 8x8, 12x12, and 16x16 to vary exploration difficulty; and (ii) scene from OGBench (Park et al., 2024) , using the predefined tasks 3, 4, and 5, which require increasingly long-horizon compositional manipulation (unlocking/locking and coordinated object placement). In all experiments, agents receive state-based observations (rather", "source": "marker_v2", "marker_block_id": "/page/3/Text/16"}
+{"paper_id": "c39d243f-0c59-4f70-8ee6-d9e742174491", "chunk_id": "c39d243f-0c59-4f70-8ee6-d9e742174491:0046", "section": "4.1. Experimental Setup", "page_start": 4, "page_end": 4, "type": "Footnote", "text": "3 Scaling to larger variants (3B or 8B) yields diminishing returns in downstream RL performance (see Appendix F.2) .", "source": "marker_v2", "marker_block_id": "/page/3/Footnote/17"}
+{"paper_id": "c39d243f-0c59-4f70-8ee6-d9e742174491", "chunk_id": "c39d243f-0c59-4f70-8ee6-d9e742174491:0047", "section": "4.1. Experimental Setup", "page_start": 5, "page_end": 5, "type": "FigureGroup", "text": "Figure 3. Frozen VLM scoring anticipates learned value. We visualize a single reference episode in MiniGrid/DoorKey-16x16 which contains both goal-directed actions and random noise. Left: Visualizations of two 32-frame sub-trajectories. clip A: captures a semantically relevant sequence. clip B: shows a mostly random walk which eventually reaches the goal. Right Top: A timeline of ground-truth events; sparse reward is only received at the final goal. Right Bottom: The rose curve shows the frozen VLM score for L=32-frame clips. The gray curves show the temporal value difference, \\Delta Q(t;L)=Q(t+\\frac{L}{2})-Q(t-\\frac{L}{2}) , calculated from critic checkpoints at increasing training steps (light to dark). Early in training (light gray), the critic is uninformative, while later checkpoints (dark gray) increasingly assign positive value to the same semantic events the VLM identified. This demonstrates that the VLM provides a helpful signal long before the critic successfully converges.", "source": "marker_v2", "marker_block_id": "/page/4/FigureGroup/268"}
+{"paper_id": "c39d243f-0c59-4f70-8ee6-d9e742174491", "chunk_id": "c39d243f-0c59-4f70-8ee6-d9e742174491:0048", "section": "4.1. Experimental Setup", "page_start": 5, "page_end": 5, "type": "Text", "text": "than pixels) to minimize perception-related confounders and isolate the effect of replay prioritization on exploration and sample efficiency. More details about the various tasks are provided in Appendix B.", "source": "marker_v2", "marker_block_id": "/page/4/Text/4"}
+{"paper_id": "c39d243f-0c59-4f70-8ee6-d9e742174491", "chunk_id": "c39d243f-0c59-4f70-8ee6-d9e742174491:0049", "section": "4.1. Experimental Setup", "page_start": 5, "page_end": 5, "type": "Text", "text": "Baselines. We compare VLM-RB to two standard replay sampling methods: Uniform Experience Replay (UER), which samples transitions uniformly from the replay buffer, and PER, which prioritizes sampling by TD-error. We report UER and PER comparisons for four different algorithms: DQN (Van Hasselt et al., 2016), IQN (Dabney et al., 2018), TD3 (Fujimoto et al., 2018), and SAC (Haarnoja et al., 2018). In addition, to compare VLM-RB against alternative replay prioritization methods, we run a focused ablation for DQN comparing to Attentive Experience Replay (AER, Sun et al. 2020), Experience Replay Optimization (ERO, Zha et al. 2019), and Reducible Loss Prioritization (ReLO, Sujit et al. 2023). Implementation details for all baselines are provided in Appendix E.", "source": "marker_v2", "marker_block_id": "/page/4/Text/5"}
+{"paper_id": "c39d243f-0c59-4f70-8ee6-d9e742174491", "chunk_id": "c39d243f-0c59-4f70-8ee6-d9e742174491:0050", "section": "4.1. Experimental Setup", "page_start": 5, "page_end": 5, "type": "Text", "text": "Method Configuration. To adapt to the specific characteristics of the experiment domains, we employ two variants of our prioritization scheme. For the discrete MiniGrid tasks, where semantic progress is binary (e.g., carrying key vs. not), we rely strictly on the binary VLM semantic filter as defined in Section 3.1. Conversely, for the continuous control tasks in OGBench, where fine-grained motion control is required, we utilize the TD-error boosted variant described in Section 3.3. This allows the agent to combine the high-level semantic filtering of the VLM with information", "source": "marker_v2", "marker_block_id": "/page/4/Text/6"}
+{"paper_id": "c39d243f-0c59-4f70-8ee6-d9e742174491", "chunk_id": "c39d243f-0c59-4f70-8ee6-d9e742174491:0051", "section": "4.1. Experimental Setup", "page_start": 5, "page_end": 5, "type": "Text", "text": "from the TD error.", "source": "marker_v2", "marker_block_id": "/page/4/Text/7"}
+{"paper_id": "c39d243f-0c59-4f70-8ee6-d9e742174491", "chunk_id": "c39d243f-0c59-4f70-8ee6-d9e742174491:0052", "section": "4.1. Experimental Setup", "page_start": 5, "page_end": 5, "type": "Text", "text": "Metrics. We report the SUCCESS RATE (SR), defined as the fraction of episodes in which the agent completes the task within the maximum episode length, T_{\\rm max} . The AVERAGE SUCCESS RATE (ASR) is the mean SR over N=32 evaluation episodes for a given seed. Final results are the ASR averaged over M=5 seeds, with the standard error of the mean (SEM) shown as shaded regions in plots.", "source": "marker_v2", "marker_block_id": "/page/4/Text/8"}
+{"paper_id": "c39d243f-0c59-4f70-8ee6-d9e742174491", "chunk_id": "c39d243f-0c59-4f70-8ee6-d9e742174491:0053", "section": "4.2. Do VLMs Contain Human-Like Priors?", "page_start": 5, "page_end": 5, "type": "Text", "text": "Before turning to downstream RL results, we first ask whether the frozen VLM offers a signal that is actually useful for prioritizing data. To understand the semantic grounding of our prioritization, we examine how the frozen VLM scores correlate with the critic value estimates ( \\Delta Q ) as learning progresses. Concretely, we track how the value estimates of the critic evolve on a fixed reference episode, using checkpoints from a successful training run. As shown in Fig. 3, the critic's estimates are initially flat and uninformative (light gray lines). Over the course of training, the critic gradually learns to assign high value to the same semantic events (such as picking up a key or opening a door) that the VLM identified from the outset. This observation suggests that VLM-RB can immediately identify semantically relevant data, enabling faster learning than approaches which must wait for the critic to converge.", "source": "marker_v2", "marker_block_id": "/page/4/Text/10"}
+{"paper_id": "c39d243f-0c59-4f70-8ee6-d9e742174491", "chunk_id": "c39d243f-0c59-4f70-8ee6-d9e742174491:0054", "section": "4.2. Do VLMs Contain Human-Like Priors?", "page_start": 5, "page_end": 5, "type": "Text", "text": "Having demonstrated that VLM scores correlate with well-trained Q-value predictors, we next ask: does this", "source": "marker_v2", "marker_block_id": "/page/4/Text/11"}
+{"paper_id": "c39d243f-0c59-4f70-8ee6-d9e742174491", "chunk_id": "c39d243f-0c59-4f70-8ee6-d9e742174491:0055", "section": "4.2. Do VLMs Contain Human-Like Priors?", "page_start": 6, "page_end": 6, "type": "FigureGroup", "text": "Figure 4. Success depends on alignment with semantic priors. We compare the performance of VLM-RB across frames rendered with Standard visuals (dark circles), Misleading swapped sprites (medium squares), and Abstract textures (light diamonds). The dashed line indicates the ASR of PER. Crucially, the agent's actual observations and the underlying MDP are identical across all settings; only the visual input to the VLM is altered (see Appendix F.3 for visual samples of these modifications). The significant performance drop in the Misleading and Abstract settings confirms that our method relies on the VLM correctly recognizing specific semantic objects (e.g., keys and doors) to accelerate learning.", "source": "marker_v2", "marker_block_id": "/page/5/FigureGroup/711"}
+{"paper_id": "c39d243f-0c59-4f70-8ee6-d9e742174491", "chunk_id": "c39d243f-0c59-4f70-8ee6-d9e742174491:0056", "section": "4.2. Do VLMs Contain Human-Like Priors?", "page_start": 6, "page_end": 6, "type": "Text", "text": "signal reflect interpretable, human-like priors? To test whether VLM-RB specifically leverages the VLM's pretrained visual semantics, we employ the \"Modified Game\" paradigm (Dubey et al., 2018) . We modify only the rendered frames used for VLM scoring, leaving the underlying MDP and agent observations untouched. We consider three settings: (i) the Standard game, with unaltered visuals; (ii) the Misleading game, where key objects are swapped (for example, traps appear as goals); and (iii) the Abstract game, where objects are replaced with random noise patterns (See Appendix F.3 for visual examples).", "source": "marker_v2", "marker_block_id": "/page/5/Text/4"}
+{"paper_id": "c39d243f-0c59-4f70-8ee6-d9e742174491", "chunk_id": "c39d243f-0c59-4f70-8ee6-d9e742174491:0057", "section": "4.2. Do VLMs Contain Human-Like Priors?", "page_start": 6, "page_end": 6, "type": "Text", "text": "As shown in Fig. 4, the Misleading and Abstract variants significantly slow learning and increase variance. Notably, in the Abstract setting, performance degrades to the level of PER (dashed line). Because the underlying MDP remains unchanged, this degradation demonstrates that the VLM's prioritization is only effective when the visual input aligns with natural semantic priors.", "source": "marker_v2", "marker_block_id": "/page/5/Text/5"}
+{"paper_id": "c39d243f-0c59-4f70-8ee6-d9e742174491", "chunk_id": "c39d243f-0c59-4f70-8ee6-d9e742174491:0058", "section": "4.2. Do VLMs Contain Human-Like Priors?", "page_start": 6, "page_end": 6, "type": "Text", "text": "In summary, these results confirm that VLM-RB scores are indicative of meaningful semantic behavior. We now turn to the question: does this semantic guidance actually yield improved sample efficiency and asymptotic performance in off-policy RL?", "source": "marker_v2", "marker_block_id": "/page/5/Text/6"}
+{"paper_id": "c39d243f-0c59-4f70-8ee6-d9e742174491", "chunk_id": "c39d243f-0c59-4f70-8ee6-d9e742174491:0059", "section": "4.3. Main Results: VLMs are Useful for RL Data Prioritization", "page_start": 6, "page_end": 6, "type": "Text", "text": "We evaluate VLM-RB and baselines across both discrete and continuous environments, as detailed in Section 4.1.", "source": "marker_v2", "marker_block_id": "/page/5/Text/8"}
+{"paper_id": "c39d243f-0c59-4f70-8ee6-d9e742174491", "chunk_id": "c39d243f-0c59-4f70-8ee6-d9e742174491:0060", "section": "4.3. Main Results: VLMs are Useful for RL Data Prioritization", "page_start": 6, "page_end": 6, "type": "TableGroup", "text": "Table 1. Improvement with respect to baselines. Performance denotes the best ASR (M = 5 seeds). Sample Efficiency tracks the steps required to reach the baseline's best performance. Values in parentheses indicate the relative improvement over the baseline. Both metrics are averaged across the aggregated algorithms. The corresponding training curves are provided in Fig. 9. Improvement Algorithm Lvl. Base. Perf. (↑) Sample Eff. (↓) 8×8 UER PER +0.0% (1.00/1.00) +0.0% (1.00/1.00) +19.1% (144K/178K) +23.0% (144K/187K) DQN+IQN (DoorKey) 12×12 UER +61.3% (1.00/0.62) +52.8% (302K/640K) PER +22.0% (1.00/0.82) +32.1% (317K/467K) 16×16 UER +241.7% (0.82/0.24) +37.6% (557K/893K) PER +70.8% (0.82/0.48) +24.1% (472K/622K) 3 UER PER +0.0% (1.00/1.00) +0.0% (1.00/1.00) +40.7% (210K/354K) +21.1% (210K/266K) SAC+TD3 (Scene) 4 PER +2.0% (1.00/0.98) UER +22.0% (1.00/0.82) +44.6% (430K/776K) +19.7% (509K/634K) 5 UER +119.4% (0.79/0.36) +46.3% (440K/819K) PER +49.1% (0.79/0.53) +17.9% (661K/805K)", "source": "marker_v2", "marker_block_id": "/page/5/TableGroup/712"}
+{"paper_id": "c39d243f-0c59-4f70-8ee6-d9e742174491", "chunk_id": "c39d243f-0c59-4f70-8ee6-d9e742174491:0061", "section": "4.3. Main Results: VLMs are Useful for RL Data Prioritization", "page_start": 6, "page_end": 6, "type": "Text", "text": "Table 1 summarizes the aggregate improvements in both performance and sample efficiency. VLM-RB consistently outperforms all baselines, regardless of algorithm or task difficulty. The largest gains appear in the most challenging settings: on DoorKey-16x16, VLM-RB achieves a +241.7% improvement over UER and +70.8% over PER; on Scene-5, the improvement over UER is +119.4% and +49.1% over PER. Even in cases where baselines eventually solve the task, VLM-RB achieves substantially better sample efficiency, reducing the number of steps needed to reach baseline performance by as much as 52.8%.", "source": "marker_v2", "marker_block_id": "/page/5/Text/11"}
+{"paper_id": "c39d243f-0c59-4f70-8ee6-d9e742174491", "chunk_id": "c39d243f-0c59-4f70-8ee6-d9e742174491:0062", "section": "4.3. Main Results: VLMs are Useful for RL Data Prioritization", "page_start": 6, "page_end": 6, "type": "Text", "text": "Comparison to Baselines. Figure 5 expands this comparison, showing that VLM-RB is the only method to consistently solve the task when compared to alternative prioritization schemes (AER, ERO, ReLo). The baselines struggle primarily due to the combination of sparse rewards and long sequential dependencies. ERO and ReLo, for example, fail to prioritize task-relevant transitions because their feedback signals (reward gradients and TD-error differences) are either dominated by noise or remain zero throughout the long pre-success phase. On the other hand, similarity-based methods like AER suffer from a structural misalignment: their local similarity metrics lead to the over-sampling of current states while neglecting critical past dependencies (such as key acquisition), which impedes temporal credit assignment over long horizons. Further analysis is provided in Appendix D.", "source": "marker_v2", "marker_block_id": "/page/5/Text/12"}
+{"paper_id": "c39d243f-0c59-4f70-8ee6-d9e742174491", "chunk_id": "c39d243f-0c59-4f70-8ee6-d9e742174491:0063", "section": "4.3. Main Results: VLMs are Useful for RL Data Prioritization", "page_start": 6, "page_end": 6, "type": "Text", "text": "Training Time Finally, we address the computational trade-off. Although VLM inference introduces a moderate throughput overhead (approximately 12% wall-clock", "source": "marker_v2", "marker_block_id": "/page/5/Text/13"}
+{"paper_id": "c39d243f-0c59-4f70-8ee6-d9e742174491", "chunk_id": "c39d243f-0c59-4f70-8ee6-d9e742174491:0064", "section": "4.3. Main Results: VLMs are Useful for RL Data Prioritization", "page_start": 7, "page_end": 7, "type": "Caption", "text": "prioritization in sparse-reward tasks. Aggregated success rates on MiniGrid/DoorKey across grid sizes (8x8, 12x12, 16x16). While alternative methods (AER, ERO, ReLo) fail as they depend on dense rewards or local similarity metrics, VLM-RB successfully bridges the long-horizon dependencies. Curves show the mean across 5 seeds, with shaded standard errors.", "source": "marker_v2", "marker_block_id": "/page/6/Caption/2"}
+{"paper_id": "c39d243f-0c59-4f70-8ee6-d9e742174491", "chunk_id": "c39d243f-0c59-4f70-8ee6-d9e742174491:0065", "section": "4.3. Main Results: VLMs are Useful for RL Data Prioritization", "page_start": 7, "page_end": 7, "type": "Text", "text": "time per step, see Appendix F.4) , this cost is outweighed by the reduction in training steps. In the vast majority of configurations (21/24), VLM-RB reduces the absolute wall-clock time to reach peak performance by up to 58%. In the rare cases where wall-clock time increased, the slowdown was marginal (<5%). By contrast, in more challenging exploration settings, VLM-RB consistently achieved speedups exceeding 20-40% (full results can be seen in Table 7 in Appendix G) . Overall, VLM-RB achieves a favorable tradeoff, substantially accelerating training in difficult sparsereward tasks, while incurring negligible overhead in simpler scenarios.", "source": "marker_v2", "marker_block_id": "/page/6/Text/3"}
+{"paper_id": "c39d243f-0c59-4f70-8ee6-d9e742174491", "chunk_id": "c39d243f-0c59-4f70-8ee6-d9e742174491:0066", "section": "4.4. Which Design Choices Matter? Mixing and Scale", "page_start": 7, "page_end": 7, "type": "Text", "text": "Sampling Mixing Schedule. A central component of our approach is to mix VLM-scored samples with uniformly sampled data (see Eq. 2) . To assess the importance of this feature, we vary only the maximal mixing coefficient, λmax, holding all other parameters fixed. The results shown in Table 4 indicate that retaining a fraction of uniform sampling is necessary to stabilize value learning. See Appendix F.1 for the full analysis of the results.", "source": "marker_v2", "marker_block_id": "/page/6/Text/5"}
+{"paper_id": "c39d243f-0c59-4f70-8ee6-d9e742174491", "chunk_id": "c39d243f-0c59-4f70-8ee6-d9e742174491:0067", "section": "4.4. Which Design Choices Matter? Mixing and Scale", "page_start": 7, "page_end": 7, "type": "Text", "text": "VLM Size. We also ablated the VLM size and found that scaling the model beyond 1B parameters dramatically increases computational cost without yielding consistent improvements in downstream RL performance. Detailed analysis of the trade-offs between throughput, memory, and performance is provided in Appendix F.2.", "source": "marker_v2", "marker_block_id": "/page/6/Text/6"}
+{"paper_id": "c39d243f-0c59-4f70-8ee6-d9e742174491", "chunk_id": "c39d243f-0c59-4f70-8ee6-d9e742174491:0068", "section": "5. Related Work", "page_start": 7, "page_end": 7, "type": "Text", "text": "Prioritized Experience Replay. Extensive research has focused on defining which transitions maximize learning utility. The seminal work, PER (Schaul et al., 2015b) , prioritizes transitions with high TD error, serving as a proxy for \"surprising\" events. Subsequent methods have proposed alternative utility metrics based on state similarity (Sun et al., 2020) , potential loss reduction (Sujit et al., 2023) , target quality/discrepancy criteria (Kumar et al., 2020) , or by explicitly learning a replay policy (Zha et al., 2019) . However, these approaches typically derive priorities from internal, training-dependent signals (e.g., value estimates or TD errors), which can be noisy or undefined early in training. In contrast, VLM-RB leverages a pre-trained VLM to assign semantic priorities, yielding a robust utility score from the very first update.", "source": "marker_v2", "marker_block_id": "/page/6/Text/8"}
+{"paper_id": "c39d243f-0c59-4f70-8ee6-d9e742174491", "chunk_id": "c39d243f-0c59-4f70-8ee6-d9e742174491:0069", "section": "5. Related Work", "page_start": 7, "page_end": 7, "type": "Text", "text": "Orthogonal to the choice of what to prioritize, several works modify how transitions are sampled. Hong et al. (2022) argue for enforcing the ordering of updates within a collected episode, to better align with the Bellman loss that is common in off-policy learning. Another example, Lahire et al. (2021) , focuses on the \"staleness\" of prioritization scores and suggests a double-sampling approach. These directions are largely complementary to VLM-RB and can be combined with VLM-based priorities. Furthermore, because VLM-RB computes priorities from a frozen VLM, it naturally avoids the score staleness issue addressed by the latter.", "source": "marker_v2", "marker_block_id": "/page/6/Text/9"}
+{"paper_id": "c39d243f-0c59-4f70-8ee6-d9e742174491", "chunk_id": "c39d243f-0c59-4f70-8ee6-d9e742174491:0070", "section": "5. Related Work", "page_start": 7, "page_end": 7, "type": "Text", "text": "Finally, rather than treating prioritization as a standalone sampling heuristic, a distinct line of work alters the underlying optimization objective itself. These methods often derive reweighting schemes from first principles, using fdivergence regularization (Li et al., 2024) , trust-region constraints (Novati & Koumoutsakos, 2019) , or density-ratio estimation (Liu et al., 2021; Sinha et al., 2022) to justify non-uniform data usage. Unlike these approaches, which fundamentally change the loss function, VLM-RB preserves the underlying training objective, ensuring it remains modular and compatible with a wide range of existing algorithms.", "source": "marker_v2", "marker_block_id": "/page/6/Text/10"}
+{"paper_id": "c39d243f-0c59-4f70-8ee6-d9e742174491", "chunk_id": "c39d243f-0c59-4f70-8ee6-d9e742174491:0071", "section": "5. Related Work", "page_start": 7, "page_end": 7, "type": "Text", "text": "LLM/VLM-Enhanced RL. Recent work has increasingly used foundation models to improve RL by injecting semantic knowledge into the learning loop. Broadly, these methods leverage LLMs/VLMs to (i) define richer learning signals (reward supervision), (ii) propose plans and exploration targets (high-level guidance), or (iii) assist action selection (policy components). Comprehensive surveys can be found in Cao et al. (2024) and Schoepp et al. (2025) .", "source": "marker_v2", "marker_block_id": "/page/6/Text/11"}
+{"paper_id": "c39d243f-0c59-4f70-8ee6-d9e742174491", "chunk_id": "c39d243f-0c59-4f70-8ee6-d9e742174491:0072", "section": "5. Related Work", "page_start": 7, "page_end": 7, "type": "Text", "text": "A major thread treats LLMs and VLMs as sources of reward. Off-the-shelf VLMs have been used as zero-shot reward scorers to judge state–goal alignment (Rocamonde et al.,", "source": "marker_v2", "marker_block_id": "/page/6/Text/12"}
+{"paper_id": "c39d243f-0c59-4f70-8ee6-d9e742174491", "chunk_id": "c39d243f-0c59-4f70-8ee6-d9e742174491:0073", "section": "5. Related Work", "page_start": 8, "page_end": 8, "type": "Text", "text": "2023) ; LLMs can synthesize dense shaping functions from natural-language task descriptions (Xie et al., 2023) ; and several methods replace or augment human feedback by using VLM/LLM judgments (absolute ratings or pairwise preferences) to train reward models (Fu et al., 2024; Luu et al., 2025; Singh et al., 2025; Ghosh et al., 2025) . Taken together, these approaches reframe reward design as semantic labeling, allowing pre-trained models to provide informative supervision in settings where environment rewards are sparse or absent.", "source": "marker_v2", "marker_block_id": "/page/7/Text/1"}
+{"paper_id": "c39d243f-0c59-4f70-8ee6-d9e742174491", "chunk_id": "c39d243f-0c59-4f70-8ee6-d9e742174491:0074", "section": "5. Related Work", "page_start": 8, "page_end": 8, "type": "Text", "text": "A second strand uses foundation models for planning and exploration. LLMs decompose instructions into temporallyextended plans grounded in affordances (Ahn et al., 2022; Huang et al., 2022) , and guide exploration by proposing semantically diverse subgoals or novelty-oriented objectives using learned embeddings (Ma et al., 2025; Gupta et al., 2022) . This high-level guidance is particularly useful in settings where unguided exploration rarely visits meaningful states.", "source": "marker_v2", "marker_block_id": "/page/7/Text/2"}
+{"paper_id": "c39d243f-0c59-4f70-8ee6-d9e742174491", "chunk_id": "c39d243f-0c59-4f70-8ee6-d9e742174491:0075", "section": "5. Related Work", "page_start": 8, "page_end": 8, "type": "Text", "text": "A third direction integrates LLMs/VLMs more directly into control, either as action advisors or as policy components. Examples range from generating executable policy code (Liang et al., 2022) to large-scale vision–language– action (VLA) models that map multimodal inputs to actions (Zitkovich et al., 2023) .", "source": "marker_v2", "marker_block_id": "/page/7/Text/3"}
+{"paper_id": "c39d243f-0c59-4f70-8ee6-d9e742174491", "chunk_id": "c39d243f-0c59-4f70-8ee6-d9e742174491:0076", "section": "5. Related Work", "page_start": 8, "page_end": 8, "type": "Text", "text": "Despite this breadth, most prior work uses these models to shape what the agent optimizes (rewards), where it searches (plans and exploration), or how it acts (policy), which commonly introduces additional training stages, fine-tuning, or computationally expensive inference. In contrast, VLM-RB targets the replay buffer—the core mechanism enabling sample reuse in off-policy RL—a component which remains underexplored for semantic foundation-model supervision. Moreover, as our experiments show, open-source pre-trained VLMs can already yield consistent gains when used to prioritize replay, without any additional model training or finetuning.", "source": "marker_v2", "marker_block_id": "/page/7/Text/4"}
+{"paper_id": "c39d243f-0c59-4f70-8ee6-d9e742174491", "chunk_id": "c39d243f-0c59-4f70-8ee6-d9e742174491:0077", "section": "6. Conclusions", "page_start": 8, "page_end": 8, "type": "Text", "text": "In this work, we presented VLM-RB, a framework that integrates the semantic reasoning capabilities of pre-trained VLMs directly into the experience replay mechanism. By shifting prioritization from statistical proxies such as TDerror to semantic evaluations of task progress, VLM-RB effectively identifies and promotes meaningful experiences even in sparse-reward regimes where traditional metrics fail. Our empirical results demonstrate that this approach significantly improves both sample efficiency and asymptotic performance across discrete and continuous domains, without requiring fine-tuning or gradient updates to the VLM.", "source": "marker_v2", "marker_block_id": "/page/7/Text/6"}
+{"paper_id": "c39d243f-0c59-4f70-8ee6-d9e742174491", "chunk_id": "c39d243f-0c59-4f70-8ee6-d9e742174491:0078", "section": "6. Conclusions", "page_start": 8, "page_end": 8, "type": "Text", "text": "Limitations. While our approach yields substantial improvements, it is important to clarify the specific assumptions and constraints under which these gains are realized. First, VLM-RB fundamentally relies on the assumption that task progress can be reliably inferred from visual observations. As a result, VLM-RB does not apply to domains where the underlying state lacks a visual representation, such as non-spatial biological systems or abstract network control tasks. In these settings, the VLM cannot ground its reasoning. Second, querying the VLM introduces additional computational overhead. Empirically, we found that this overhead reduced throughput by approximately 12% in our experiments. It is important to note that this computational cost is not intrinsic to the method; the precise overhead depends on hardware configuration and implementation choices, so the trade-off may vary across different setups 4 .", "source": "marker_v2", "marker_block_id": "/page/7/Text/7"}
+{"paper_id": "c39d243f-0c59-4f70-8ee6-d9e742174491", "chunk_id": "c39d243f-0c59-4f70-8ee6-d9e742174491:0079", "section": "6. Conclusions", "page_start": 8, "page_end": 8, "type": "Text", "text": "Future Work. Our findings suggest several concrete directions for future research. A natural extension is the application of VLM-RB to Goal-Conditioned RL (Schaul et al., 2015a) , where the textual description of the current goal can be injected into the VLM prompt to dynamically score alignment, or hindsight textual goals could promote meaningful sub-goals which may have some semantic relation but no direct relevance to the current task (Luu & Yoo, 2021; Jiang et al., 2019) . Another promising direction is to use the replay buffer as a curriculum mechanism, by evolving the prompt over the course of training. For example, the system could prioritize specific skills such as \"open drawer\" in the early stages, and later shift focus to more compositional behaviors. Finally, the prioritization scheme could be extended beyond single VLM scoring, for example, by incorporating more robust inference techniques such as majority voting or LLM-as-a-Judge.", "source": "marker_v2", "marker_block_id": "/page/7/Text/8"}
+{"paper_id": "c39d243f-0c59-4f70-8ee6-d9e742174491", "chunk_id": "c39d243f-0c59-4f70-8ee6-d9e742174491:0080", "section": "6. Conclusions", "page_start": 8, "page_end": 8, "type": "Text", "text": "Impact Statement. This paper presents work aimed at advancing the field of machine learning, specifically in the domain of sample-efficient reinforcement learning. There are many potential societal consequences of our work, none of which we feel must be specifically highlighted here.", "source": "marker_v2", "marker_block_id": "/page/7/Text/9"}

icml26/c39d243f-0c59-4f70-8ee6-d9e742174491/marker_meta.json

@@ -0,0 +1,2189 @@
+{
+  "table_of_contents": [
+    {
+      "title": "VLM-Guided Experience Replay",
+      "heading_level": null,
+      "page_id": 0,
+      "polygon": [
+        [
+          197.47300720214844,
+          89.71875
+        ],
+        [
+          399.9122314453125,
+          89.71875
+        ],
+        [
+          399.9122314453125,
+          104.2581787109375
+        ],
+        [
+          197.47300720214844,
+          104.2581787109375
+        ]
+      ]
+    },
+    {
+      "title": "Anonymous Authors1",
+      "heading_level": null,
+      "page_id": 0,
+      "polygon": [
+        [
+          245.7449951171875,
+          141.74615478515625
+        ],
+        [
+          337.078125,
+          141.74615478515625
+        ],
+        [
+          337.078125,
+          152.96612548828125
+        ],
+        [
+          245.7449951171875,
+          152.96612548828125
+        ]
+      ]
+    },
+    {
+      "title": "Abstract",
+      "heading_level": null,
+      "page_id": 0,
+      "polygon": [
+        [
+          148.517578125,
+          175.18359375
+        ],
+        [
+          194.68325805664062,
+          175.18359375
+        ],
+        [
+          194.68325805664062,
+          188.015380859375
+        ],
+        [
+          148.517578125,
+          188.015380859375
+        ]
+      ]
+    },
+    {
+      "title": "1. Introduction",
+      "heading_level": null,
+      "page_id": 0,
+      "polygon": [
+        [
+          54.087890625,
+          481.46484375
+        ],
+        [
+          132.27606201171875,
+          481.46484375
+        ],
+        [
+          132.27606201171875,
+          494.1533508300781
+        ],
+        [
+          54.087890625,
+          494.1533508300781
+        ]
+      ]
+    },
+    {
+      "title": "2. Preliminaries",
+      "heading_level": null,
+      "page_id": 1,
+      "polygon": [
+        [
+          54.0,
+          629.96484375
+        ],
+        [
+          136.5,
+          629.96484375
+        ],
+        [
+          136.5,
+          639.75
+        ],
+        [
+          54.0,
+          639.75
+        ]
+      ]
+    },
+    {
+      "title": "3. VLM-RB",
+      "heading_level": null,
+      "page_id": 2,
+      "polygon": [
+        [
+          305.103515625,
+          228.0
+        ],
+        [
+          371.25,
+          228.0
+        ],
+        [
+          371.25,
+          237.83203125
+        ],
+        [
+          305.103515625,
+          237.83203125
+        ]
+      ]
+    },
+    {
+      "title": "3.1. Scoring, Prioritization, and Sampling",
+      "heading_level": null,
+      "page_id": 2,
+      "polygon": [
+        [
+          305.40234375,
+          344.953125
+        ],
+        [
+          486.0,
+          344.953125
+        ],
+        [
+          486.0,
+          354.75
+        ],
+        [
+          305.40234375,
+          354.75
+        ]
+      ]
+    },
+    {
+      "title": "3.2. Efficient Implementation",
+      "heading_level": null,
+      "page_id": 3,
+      "polygon": [
+        [
+          52.59375,
+          605.0755157470703
+        ],
+        [
+          179.41456604003906,
+          605.0755157470703
+        ],
+        [
+          179.41456604003906,
+          615.0381164550781
+        ],
+        [
+          52.59375,
+          615.0381164550781
+        ]
+      ]
+    },
+    {
+      "title": "3.3. Boosting with TD-error",
+      "heading_level": null,
+      "page_id": 3,
+      "polygon": [
+        [
+          306.298828125,
+          213.46875
+        ],
+        [
+          425.4171142578125,
+          213.46875
+        ],
+        [
+          425.4171142578125,
+          224.0240478515625
+        ],
+        [
+          306.298828125,
+          224.0240478515625
+        ]
+      ]
+    },
+    {
+      "title": "4. Experiments",
+      "heading_level": null,
+      "page_id": 3,
+      "polygon": [
+        [
+          304.505859375,
+          384.3984375
+        ],
+        [
+          384.47930908203125,
+          384.3984375
+        ],
+        [
+          384.47930908203125,
+          397.2343444824219
+        ],
+        [
+          304.505859375,
+          397.2343444824219
+        ]
+      ]
+    },
+    {
+      "title": "4.1. Experimental Setup",
+      "heading_level": null,
+      "page_id": 3,
+      "polygon": [
+        [
+          306.896484375,
+          562.8215026855469
+        ],
+        [
+          410.291015625,
+          562.8215026855469
+        ],
+        [
+          410.291015625,
+          573.50390625
+        ],
+        [
+          306.896484375,
+          573.50390625
+        ]
+      ]
+    },
+    {
+      "title": "4.2. Do VLMs Contain Human-Like Priors?",
+      "heading_level": null,
+      "page_id": 4,
+      "polygon": [
+        [
+          305.40234375,
+          478.37109375
+        ],
+        [
+          494.25,
+          478.37109375
+        ],
+        [
+          494.25,
+          488.25
+        ],
+        [
+          305.40234375,
+          488.25
+        ]
+      ]
+    },
+    {
+      "title": "4.3. Main Results: VLMs are Useful for RL Data\nPrioritization",
+      "heading_level": null,
+      "page_id": 5,
+      "polygon": [
+        [
+          52.59375,
+          664.3828125
+        ],
+        [
+          262.5823669433594,
+          664.3828125
+        ],
+        [
+          262.5823669433594,
+          686.7690963745117
+        ],
+        [
+          52.59375,
+          686.7690963745117
+        ]
+      ]
+    },
+    {
+      "title": "4.4. Which Design Choices Matter? Mixing and Scale",
+      "heading_level": null,
+      "page_id": 6,
+      "polygon": [
+        [
+          52.59375,
+          504.4615173339844
+        ],
+        [
+          282.12896728515625,
+          504.4615173339844
+        ],
+        [
+          282.12896728515625,
+          514.72265625
+        ],
+        [
+          52.59375,
+          514.72265625
+        ]
+      ]
+    },
+    {
+      "title": "5. Related Work",
+      "heading_level": null,
+      "page_id": 6,
+      "polygon": [
+        [
+          306.59765625,
+          67.2890625
+        ],
+        [
+          391.46484375,
+          67.2890625
+        ],
+        [
+          391.46484375,
+          79.80230712890625
+        ],
+        [
+          306.59765625,
+          79.80230712890625
+        ]
+      ]
+    },
+    {
+      "title": "6. Conclusions",
+      "heading_level": null,
+      "page_id": 7,
+      "polygon": [
+        [
+          53.490234375,
+          543.33984375
+        ],
+        [
+          129.1796875,
+          543.33984375
+        ],
+        [
+          129.1796875,
+          556.6863708496094
+        ],
+        [
+          53.490234375,
+          556.6863708496094
+        ]
+      ]
+    },
+    {
+      "title": "References",
+      "heading_level": null,
+      "page_id": 7,
+      "polygon": [
+        [
+          307.1953125,
+          579.2431640625
+        ],
+        [
+          362.9838562011719,
+          579.2431640625
+        ],
+        [
+          362.9838562011719,
+          591.1983642578125
+        ],
+        [
+          307.1953125,
+          591.1983642578125
+        ]
+      ]
+    },
+    {
+      "title": "A. VLM-RB",
+      "heading_level": null,
+      "page_id": 11,
+      "polygon": [
+        [
+          53.19140625,
+          67.67578125
+        ],
+        [
+          122.25,
+          67.67578125
+        ],
+        [
+          122.25,
+          77.25
+        ],
+        [
+          53.19140625,
+          77.25
+        ]
+      ]
+    },
+    {
+      "title": "Algorithm 1 VLM-Prioritized Replay Buffer",
+      "heading_level": null,
+      "page_id": 11,
+      "polygon": [
+        [
+          53.19140625,
+          94.5
+        ],
+        [
+          244.5,
+          94.5
+        ],
+        [
+          244.5,
+          105.1875
+        ],
+        [
+          53.19140625,
+          105.1875
+        ]
+      ]
+    },
+    {
+      "title": "B. Environment Details",
+      "heading_level": null,
+      "page_id": 11,
+      "polygon": [
+        [
+          53.7890625,
+          459.75
+        ],
+        [
+          175.5,
+          459.75
+        ],
+        [
+          175.5,
+          468.75
+        ],
+        [
+          53.7890625,
+          468.75
+        ]
+      ]
+    },
+    {
+      "title": "C. Prompts",
+      "heading_level": null,
+      "page_id": 13,
+      "polygon": [
+        [
+          53.19140625,
+          67.84710693359375
+        ],
+        [
+          113.66183471679688,
+          67.84710693359375
+        ],
+        [
+          113.66183471679688,
+          79.80230712890625
+        ],
+        [
+          53.19140625,
+          79.80230712890625
+        ]
+      ]
+    },
+    {
+      "title": "Meta-prompt template for generating task prompts",
+      "heading_level": null,
+      "page_id": 13,
+      "polygon": [
+        [
+          66.9375,
+          234.6544189453125
+        ],
+        [
+          268.98974609375,
+          234.6544189453125
+        ],
+        [
+          268.98974609375,
+          244.61700439453125
+        ],
+        [
+          66.9375,
+          244.61700439453125
+        ]
+      ]
+    },
+    {
+      "title": "Examples (few-shot):",
+      "heading_level": null,
+      "page_id": 13,
+      "polygon": [
+        [
+          72.9140625,
+          384.01171875
+        ],
+        [
+          156.7117156982422,
+          384.01171875
+        ],
+        [
+          156.7117156982422,
+          393.7444763183594
+        ],
+        [
+          72.9140625,
+          393.7444763183594
+        ]
+      ]
+    },
+    {
+      "title": "Current request:",
+      "heading_level": null,
+      "page_id": 13,
+      "polygon": [
+        [
+          71.71875,
+          474.4410705566406
+        ],
+        [
+          140.4466552734375,
+          474.4410705566406
+        ],
+        [
+          140.4466552734375,
+          483.78515625
+        ],
+        [
+          71.71875,
+          483.78515625
+        ]
+      ]
+    },
+    {
+      "title": "Example: prompt families for DeepMind Control Suite",
+      "heading_level": null,
+      "page_id": 14,
+      "polygon": [
+        [
+          65.14453125,
+          115.9044189453125
+        ],
+        [
+          286.94232177734375,
+          115.9044189453125
+        ],
+        [
+          286.94232177734375,
+          125.86700439453125
+        ],
+        [
+          65.14453125,
+          125.86700439453125
+        ]
+      ]
+    },
+    {
+      "title": "Group 1: Going Somewhere (Locomotion)",
+      "heading_level": null,
+      "page_id": 14,
+      "polygon": [
+        [
+          74.70703125,
+          220.38818359375
+        ],
+        [
+          237.38241577148438,
+          220.38818359375
+        ],
+        [
+          237.38241577148438,
+          229.3546142578125
+        ],
+        [
+          74.70703125,
+          229.3546142578125
+        ]
+      ]
+    },
+    {
+      "title": "The Prompt:",
+      "heading_level": null,
+      "page_id": 14,
+      "polygon": [
+        [
+          73.51171875,
+          278.82421875
+        ],
+        [
+          124.88088989257812,
+          278.82421875
+        ],
+        [
+          124.88088989257812,
+          288.1344909667969
+        ],
+        [
+          73.51171875,
+          288.1344909667969
+        ]
+      ]
+    },
+    {
+      "title": "Group 2: Defying Gravity (Stabilization)",
+      "heading_level": null,
+      "page_id": 14,
+      "polygon": [
+        [
+          74.109375,
+          363.85009765625
+        ],
+        [
+          231.85015869140625,
+          363.85009765625
+        ],
+        [
+          231.85015869140625,
+          372.8164978027344
+        ],
+        [
+          74.109375,
+          372.8164978027344
+        ]
+      ]
+    },
+    {
+      "title": "The Prompt:",
+      "heading_level": null,
+      "page_id": 14,
+      "polygon": [
+        [
+          73.810546875,
+          431.96484375
+        ],
+        [
+          124.88088989257812,
+          431.96484375
+        ],
+        [
+          124.88088989257812,
+          441.5594787597656
+        ],
+        [
+          73.810546875,
+          441.5594787597656
+        ]
+      ]
+    },
+    {
+      "title": "Group 3: Touching the Thing (Interaction)",
+      "heading_level": null,
+      "page_id": 14,
+      "polygon": [
+        [
+          75.3046875,
+          517.275146484375
+        ],
+        [
+          239.48056030273438,
+          517.275146484375
+        ],
+        [
+          239.48056030273438,
+          527.09765625
+        ],
+        [
+          75.3046875,
+          527.09765625
+        ]
+      ]
+    },
+    {
+      "title": "The Prompt:",
+      "heading_level": null,
+      "page_id": 14,
+      "polygon": [
+        [
+          74.70703125,
+          576.0551300048828
+        ],
+        [
+          124.88088989257812,
+          576.0551300048828
+        ],
+        [
+          124.88088989257812,
+          585.87890625
+        ],
+        [
+          74.70703125,
+          585.87890625
+        ]
+      ]
+    },
+    {
+      "title": "D. Baselines",
+      "heading_level": null,
+      "page_id": 15,
+      "polygon": [
+        [
+          52.892578125,
+          67.67578125
+        ],
+        [
+          116.97341918945312,
+          67.67578125
+        ],
+        [
+          116.97341918945312,
+          79.80230712890625
+        ],
+        [
+          52.892578125,
+          79.80230712890625
+        ]
+      ]
+    },
+    {
+      "title": "E. Implementation Details",
+      "heading_level": null,
+      "page_id": 16,
+      "polygon": [
+        [
+          51.697265625,
+          68.25
+        ],
+        [
+          189.0,
+          68.25
+        ],
+        [
+          189.0,
+          79.6640625
+        ],
+        [
+          51.697265625,
+          79.6640625
+        ]
+      ]
+    },
+    {
+      "title": "E.1. Architectural and Implementation Details",
+      "heading_level": null,
+      "page_id": 17,
+      "polygon": [
+        [
+          52.59375,
+          177.0
+        ],
+        [
+          253.5,
+          177.0
+        ],
+        [
+          253.5,
+          187.171875
+        ],
+        [
+          52.59375,
+          187.171875
+        ]
+      ]
+    },
+    {
+      "title": "F. Ablations",
+      "heading_level": null,
+      "page_id": 20,
+      "polygon": [
+        [
+          52.59375,
+          67.67578125
+        ],
+        [
+          116.56694030761719,
+          67.67578125
+        ],
+        [
+          116.56694030761719,
+          79.80230712890625
+        ],
+        [
+          52.59375,
+          79.80230712890625
+        ]
+      ]
+    },
+    {
+      "title": "F.1. Mixing Schedule",
+      "heading_level": null,
+      "page_id": 20,
+      "polygon": [
+        [
+          51.3984375,
+          88.74945068359375
+        ],
+        [
+          144.01748657226562,
+          88.74945068359375
+        ],
+        [
+          144.01748657226562,
+          99.0
+        ],
+        [
+          51.3984375,
+          99.0
+        ]
+      ]
+    },
+    {
+      "title": "F.2. VLM Size",
+      "heading_level": null,
+      "page_id": 21,
+      "polygon": [
+        [
+          52.294921875,
+          69.37646484375
+        ],
+        [
+          115.77350616455078,
+          69.37646484375
+        ],
+        [
+          115.77350616455078,
+          79.33905029296875
+        ],
+        [
+          52.294921875,
+          79.33905029296875
+        ]
+      ]
+    },
+    {
+      "title": "F.3. VLM Prior",
+      "heading_level": null,
+      "page_id": 21,
+      "polygon": [
+        [
+          54.087890625,
+          533.344482421875
+        ],
+        [
+          121.3027572631836,
+          533.344482421875
+        ],
+        [
+          121.3027572631836,
+          543.3070831298828
+        ],
+        [
+          54.087890625,
+          543.3070831298828
+        ]
+      ]
+    },
+    {
+      "title": "F.4. Computational Overhead",
+      "heading_level": null,
+      "page_id": 22,
+      "polygon": [
+        [
+          54.087890625,
+          267.22265625
+        ],
+        [
+          181.5465850830078,
+          267.22265625
+        ],
+        [
+          181.5465850830078,
+          278.44012451171875
+        ],
+        [
+          54.087890625,
+          278.44012451171875
+        ]
+      ]
+    },
+    {
+      "title": "G. Experiments",
+      "heading_level": null,
+      "page_id": 23,
+      "polygon": [
+        [
+          53.7890625,
+          67.84710693359375
+        ],
+        [
+          135.8028564453125,
+          67.84710693359375
+        ],
+        [
+          135.8028564453125,
+          79.80230712890625
+        ],
+        [
+          53.7890625,
+          79.80230712890625
+        ]
+      ]
+    }
+  ],
+  "page_stats": [
+    {
+      "page_id": 0,
+      "text_extraction_method": "pdftext",
+      "block_counts": [
+        [
+          "Span",
+          341
+        ],
+        [
+          "Line",
+          141
+        ],
+        [
+          "Text",
+          11
+        ],
+        [
+          "SectionHeader",
+          4
+        ],
+        [
+          "Footnote",
+          1
+        ],
+        [
+          "PageFooter",
+          1
+        ]
+      ],
+      "block_metadata": {
+        "llm_request_count": 0,
+        "llm_error_count": 0,
+        "llm_tokens_used": 0,
+        "previous_text": "",
+        "previous_type": "",
+        "previous_order": 0
+      }
+    },
+    {
+      "page_id": 1,
+      "text_extraction_method": "surya",
+      "block_counts": [
+        [
+          "Line",
+          180
+        ],
+        [
+          "Span",
+          156
+        ],
+        [
+          "Text",
+          53
+        ],
+        [
+          "PageHeader",
+          2
+        ],
+        [
+          "Equation",
+          2
+        ],
+        [
+          "Caption",
+          1
+        ],
+        [
+          "SectionHeader",
+          1
+        ],
+        [
+          "PageFooter",
+          1
+        ],
+        [
+          "Reference",
+          1
+        ]
+      ],
+      "block_metadata": {
+        "llm_request_count": 0,
+        "llm_error_count": 0,
+        "llm_tokens_used": 0,
+        "previous_text": "",
+        "previous_type": "",
+        "previous_order": 0
+      }
+    },
+    {
+      "page_id": 2,
+      "text_extraction_method": "surya",
+      "block_counts": [
+        [
+          "Line",
+          211
+        ],
+        [
+          "Span",
+          142
+        ],
+        [
+          "Text",
+          61
+        ],
+        [
+          "Reference",
+          5
+        ],
+        [
+          "Footnote",
+          2
+        ],
+        [
+          "SectionHeader",
+          2
+        ],
+        [
+          "PageHeader",
+          1
+        ],
+        [
+          "Figure",
+          1
+        ],
+        [
+          "Caption",
+          1
+        ],
+        [
+          "Equation",
+          1
+        ],
+        [
+          "PageFooter",
+          1
+        ]
+      ],
+      "block_metadata": {
+        "llm_request_count": 0,
+        "llm_error_count": 0,
+        "llm_tokens_used": 0,
+        "previous_text": "",
+        "previous_type": "",
+        "previous_order": 0
+      }
+    },
+    {
+      "page_id": 3,
+      "text_extraction_method": "pdftext",
+      "block_counts": [
+        [
+          "Span",
+          555
+        ],
+        [
+          "Line",
+          169
+        ],
+        [
+          "Text",
+          19
+        ],
+        [
+          "Reference",
+          5
+        ],
+        [
+          "SectionHeader",
+          4
+        ],
+        [
+          "PageHeader",
+          1
+        ],
+        [
+          "Equation",
+          1
+        ],
+        [
+          "Footnote",
+          1
+        ],
+        [
+          "PageFooter",
+          1
+        ]
+      ],
+      "block_metadata": {
+        "llm_request_count": 0,
+        "llm_error_count": 0,
+        "llm_tokens_used": 0,
+        "previous_text": "",
+        "previous_type": "",
+        "previous_order": 0
+      }
+    },
+    {
+      "page_id": 4,
+      "text_extraction_method": "surya",
+      "block_counts": [
+        [
+          "Line",
+          148
+        ],
+        [
+          "Span",
+          91
+        ],
+        [
+          "Text",
+          23
+        ],
+        [
+          "PageHeader",
+          2
+        ],
+        [
+          "Reference",
+          2
+        ],
+        [
+          "Figure",
+          1
+        ],
+        [
+          "Caption",
+          1
+        ],
+        [
+          "SectionHeader",
+          1
+        ],
+        [
+          "PageFooter",
+          1
+        ],
+        [
+          "FigureGroup",
+          1
+        ]
+      ],
+      "block_metadata": {
+        "llm_request_count": 0,
+        "llm_error_count": 0,
+        "llm_tokens_used": 0,
+        "previous_text": "",
+        "previous_type": "",
+        "previous_order": 0
+      }
+    },
+    {
+      "page_id": 5,
+      "text_extraction_method": "pdftext",
+      "block_counts": [
+        [
+          "Span",
+          552
+        ],
+        [
+          "Line",
+          137
+        ],
+        [
+          "TableCell",
+          39
+        ],
+        [
+          "Text",
+          14
+        ],
+        [
+          "Reference",
+          3
+        ],
+        [
+          "PageHeader",
+          2
+        ],
+        [
+          "Caption",
+          2
+        ],
+        [
+          "Figure",
+          1
+        ],
+        [
+          "SectionHeader",
+          1
+        ],
+        [
+          "Table",
+          1
+        ],
+        [
+          "PageFooter",
+          1
+        ],
+        [
+          "FigureGroup",
+          1
+        ],
+        [
+          "TableGroup",
+          1
+        ]
+      ],
+      "block_metadata": {
+        "llm_request_count": 0,
+        "llm_error_count": 0,
+        "llm_tokens_used": 0,
+        "previous_text": "",
+        "previous_type": "",
+        "previous_order": 0
+      }
+    },
+    {
+      "page_id": 6,
+      "text_extraction_method": "pdftext",
+      "block_counts": [
+        [
+          "Span",
+          401
+        ],
+        [
+          "Line",
+          129
+        ],
+        [
+          "Text",
+          18
+        ],
+        [
+          "SectionHeader",
+          2
+        ],
+        [
+          "Reference",
+          2
+        ],
+        [
+          "PageHeader",
+          1
+        ],
+        [
+          "Figure",
+          1
+        ],
+        [
+          "Caption",
+          1
+        ],
+        [
+          "PageFooter",
+          1
+        ]
+      ],
+      "block_metadata": {
+        "llm_request_count": 0,
+        "llm_error_count": 0,
+        "llm_tokens_used": 0,
+        "previous_text": "",
+        "previous_type": "",
+        "previous_order": 0
+      }
+    },
+    {
+      "page_id": 7,
+      "text_extraction_method": "pdftext",
+      "block_counts": [
+        [
+          "Span",
+          367
+        ],
+        [
+          "Line",
+          152
+        ],
+        [
+          "Text",
+          14
+        ],
+        [
+          "Reference",
+          3
+        ],
+        [
+          "SectionHeader",
+          2
+        ],
+        [
+          "PageHeader",
+          1
+        ],
+        [
+          "Footnote",
+          1
+        ],
+        [
+          "PageFooter",
+          1
+        ]
+      ],
+      "block_metadata": {
+        "llm_request_count": 0,
+        "llm_error_count": 0,
+        "llm_tokens_used": 0,
+        "previous_text": "",
+        "previous_type": "",
+        "previous_order": 0
+      }
+    },
+    {
+      "page_id": 8,
+      "text_extraction_method": "pdftext",
+      "block_counts": [
+        [
+          "Span",
+          360
+        ],
+        [
+          "Line",
+          151
+        ],
+        [
+          "ListItem",
+          23
+        ],
+        [
+          "Reference",
+          22
+        ],
+        [
+          "Text",
+          2
+        ],
+        [
+          "ListGroup",
+          2
+        ],
+        [
+          "PageHeader",
+          1
+        ],
+        [
+          "PageFooter",
+          1
+        ]
+      ],
+      "block_metadata": {
+        "llm_request_count": 0,
+        "llm_error_count": 0,
+        "llm_tokens_used": 0,
+        "previous_text": "",
+        "previous_type": "",
+        "previous_order": 0
+      }
+    },
+    {
+      "page_id": 9,
+      "text_extraction_method": "pdftext",
+      "block_counts": [
+        [
+          "Span",
+          330
+        ],
+        [
+          "Line",
+          135
+        ],
+        [
+          "ListItem",
+          24
+        ],
+        [
+          "Reference",
+          24
+        ],
+        [
+          "Text",
+          12
+        ],
+        [
+          "ListGroup",
+          2
+        ],
+        [
+          "PageHeader",
+          1
+        ],
+        [
+          "PageFooter",
+          1
+        ]
+      ],
+      "block_metadata": {
+        "llm_request_count": 0,
+        "llm_error_count": 0,
+        "llm_tokens_used": 0,
+        "previous_text": "",
+        "previous_type": "",
+        "previous_order": 0
+      }
+    },
+    {
+      "page_id": 10,
+      "text_extraction_method": "pdftext",
+      "block_counts": [
+        [
+          "Span",
+          177
+        ],
+        [
+          "Line",
+          78
+        ],
+        [
+          "Text",
+          10
+        ],
+        [
+          "ListItem",
+          9
+        ],
+        [
+          "Reference",
+          9
+        ],
+        [
+          "PageHeader",
+          1
+        ],
+        [
+          "PageFooter",
+          1
+        ],
+        [
+          "ListGroup",
+          1
+        ]
+      ],
+      "block_metadata": {
+        "llm_request_count": 0,
+        "llm_error_count": 0,
+        "llm_tokens_used": 0,
+        "previous_text": "",
+        "previous_type": "",
+        "previous_order": 0
+      }
+    },
+    {
+      "page_id": 11,
+      "text_extraction_method": "surya",
+      "block_counts": [
+        [
+          "Line",
+          164
+        ],
+        [
+          "Span",
+          149
+        ],
+        [
+          "Text",
+          28
+        ],
+        [
+          "SectionHeader",
+          3
+        ],
+        [
+          "Reference",
+          2
+        ],
+        [
+          "PageHeader",
+          1
+        ],
+        [
+          "Code",
+          1
+        ],
+        [
+          "PageFooter",
+          1
+        ]
+      ],
+      "block_metadata": {
+        "llm_request_count": 0,
+        "llm_error_count": 0,
+        "llm_tokens_used": 0,
+        "previous_text": "",
+        "previous_type": "",
+        "previous_order": 0
+      }
+    },
+    {
+      "page_id": 12,
+      "text_extraction_method": "surya",
+      "block_counts": [
+        [
+          "Line",
+          135
+        ],
+        [
+          "Span",
+          73
+        ],
+        [
+          "Text",
+          27
+        ],
+        [
+          "ListItem",
+          3
+        ],
+        [
+          "PageHeader",
+          1
+        ],
+        [
+          "Equation",
+          1
+        ],
+        [
+          "PageFooter",
+          1
+        ],
+        [
+          "ListGroup",
+          1
+        ]
+      ],
+      "block_metadata": {
+        "llm_request_count": 0,
+        "llm_error_count": 0,
+        "llm_tokens_used": 0,
+        "previous_text": "",
+        "previous_type": "",
+        "previous_order": 0
+      }
+    },
+    {
+      "page_id": 13,
+      "text_extraction_method": "pdftext",
+      "block_counts": [
+        [
+          "Span",
+          209
+        ],
+        [
+          "Line",
+          73
+        ],
+        [
+          "Text",
+          20
+        ],
+        [
+          "ListItem",
+          6
+        ],
+        [
+          "SectionHeader",
+          4
+        ],
+        [
+          "ListGroup",
+          2
+        ],
+        [
+          "PageHeader",
+          1
+        ],
+        [
+          "PageFooter",
+          1
+        ],
+        [
+          "Reference",
+          1
+        ]
+      ],
+      "block_metadata": {
+        "llm_request_count": 0,
+        "llm_error_count": 0,
+        "llm_tokens_used": 0,
+        "previous_text": "",
+        "previous_type": "",
+        "previous_order": 0
+      }
+    },
+    {
+      "page_id": 14,
+      "text_extraction_method": "pdftext",
+      "block_counts": [
+        [
+          "Span",
+          289
+        ],
+        [
+          "Line",
+          88
+        ],
+        [
+          "Text",
+          14
+        ],
+        [
+          "SectionHeader",
+          7
+        ],
+        [
+          "PageHeader",
+          1
+        ],
+        [
+          "PageFooter",
+          1
+        ]
+      ],
+      "block_metadata": {
+        "llm_request_count": 0,
+        "llm_error_count": 0,
+        "llm_tokens_used": 0,
+        "previous_text": "",
+        "previous_type": "",
+        "previous_order": 0
+      }
+    },
+    {
+      "page_id": 15,
+      "text_extraction_method": "pdftext",
+      "block_counts": [
+        [
+          "Span",
+          299
+        ],
+        [
+          "Line",
+          90
+        ],
+        [
+          "Text",
+          5
+        ],
+        [
+          "PageHeader",
+          1
+        ],
+        [
+          "SectionHeader",
+          1
+        ],
+        [
+          "PageFooter",
+          1
+        ],
+        [
+          "Reference",
+          1
+        ]
+      ],
+      "block_metadata": {
+        "llm_request_count": 0,
+        "llm_error_count": 0,
+        "llm_tokens_used": 0,
+        "previous_text": "",
+        "previous_type": "",
+        "previous_order": 0
+      }
+    },
+    {
+      "page_id": 16,
+      "text_extraction_method": "surya",
+      "block_counts": [
+        [
+          "Line",
+          147
+        ],
+        [
+          "Span",
+          137
+        ],
+        [
+          "Text",
+          47
+        ],
+        [
+          "Equation",
+          5
+        ],
+        [
+          "PageHeader",
+          1
+        ],
+        [
+          "SectionHeader",
+          1
+        ],
+        [
+          "PageFooter",
+          1
+        ],
+        [
+          "Reference",
+          1
+        ]
+      ],
+      "block_metadata": {
+        "llm_request_count": 0,
+        "llm_error_count": 0,
+        "llm_tokens_used": 0,
+        "previous_text": "",
+        "previous_type": "",
+        "previous_order": 0
+      }
+    },
+    {
+      "page_id": 17,
+      "text_extraction_method": "surya",
+      "block_counts": [
+        [
+          "Line",
+          156
+        ],
+        [
+          "Span",
+          95
+        ],
+        [
+          "Text",
+          53
+        ],
+        [
+          "PageHeader",
+          1
+        ],
+        [
+          "SectionHeader",
+          1
+        ],
+        [
+          "PageFooter",
+          1
+        ]
+      ],
+      "block_metadata": {
+        "llm_request_count": 0,
+        "llm_error_count": 0,
+        "llm_tokens_used": 0,
+        "previous_text": "",
+        "previous_type": "",
+        "previous_order": 0
+      }
+    },
+    {
+      "page_id": 18,
+      "text_extraction_method": "pdftext",
+      "block_counts": [
+        [
+          "Span",
+          273
+        ],
+        [
+          "TableCell",
+          117
+        ],
+        [
+          "Line",
+          88
+        ],
+        [
+          "Text",
+          11
+        ],
+        [
+          "PageHeader",
+          1
+        ],
+        [
+          "Caption",
+          1
+        ],
+        [
+          "Table",
+          1
+        ],
+        [
+          "PageFooter",
+          1
+        ],
+        [
+          "Reference",
+          1
+        ]
+      ],
+      "block_metadata": {
+        "llm_request_count": 0,
+        "llm_error_count": 0,
+        "llm_tokens_used": 0,
+        "previous_text": "",
+        "previous_type": "",
+        "previous_order": 0
+      }
+    },
+    {
+      "page_id": 19,
+      "text_extraction_method": "pdftext",
+      "block_counts": [
+        [
+          "Span",
+          224
+        ],
+        [
+          "TableCell",
+          88
+        ],
+        [
+          "Line",
+          82
+        ],
+        [
+          "Text",
+          9
+        ],
+        [
+          "PageHeader",
+          1
+        ],
+        [
+          "Caption",
+          1
+        ],
+        [
+          "Table",
+          1
+        ],
+        [
+          "PageFooter",
+          1
+        ],
+        [
+          "Reference",
+          1
+        ]
+      ],
+      "block_metadata": {
+        "llm_request_count": 0,
+        "llm_error_count": 0,
+        "llm_tokens_used": 0,
+        "previous_text": "",
+        "previous_type": "",
+        "previous_order": 0
+      }
+    },
+    {
+      "page_id": 20,
+      "text_extraction_method": "pdftext",
+      "block_counts": [
+        [
+          "Span",
+          419
+        ],
+        [
+          "Line",
+          87
+        ],
+        [
+          "TableCell",
+          27
+        ],
+        [
+          "Text",
+          6
+        ],
+        [
+          "Caption",
+          4
+        ],
+        [
+          "Reference",
+          3
+        ],
+        [
+          "SectionHeader",
+          2
+        ],
+        [
+          "Figure",
+          2
+        ],
+        [
+          "PageHeader",
+          1
+        ],
+        [
+          "Table",
+          1
+        ],
+        [
+          "PageFooter",
+          1
+        ],
+        [
+          "TableGroup",
+          1
+        ],
+        [
+          "FigureGroup",
+          1
+        ]
+      ],
+      "block_metadata": {
+        "llm_request_count": 0,
+        "llm_error_count": 0,
+        "llm_tokens_used": 0,
+        "previous_text": "",
+        "previous_type": "",
+        "previous_order": 0
+      }
+    },
+    {
+      "page_id": 21,
+      "text_extraction_method": "pdftext",
+      "block_counts": [
+        [
+          "Span",
+          276
+        ],
+        [
+          "Line",
+          88
+        ],
+        [
+          "TableCell",
+          20
+        ],
+        [
+          "Text",
+          7
+        ],
+        [
+          "Reference",
+          4
+        ],
+        [
+          "SectionHeader",
+          2
+        ],
+        [
+          "Caption",
+          2
+        ],
+        [
+          "PageHeader",
+          1
+        ],
+        [
+          "Table",
+          1
+        ],
+        [
+          "Figure",
+          1
+        ],
+        [
+          "PageFooter",
+          1
+        ],
+        [
+          "TableGroup",
+          1
+        ],
+        [
+          "FigureGroup",
+          1
+        ]
+      ],
+      "block_metadata": {
+        "llm_request_count": 0,
+        "llm_error_count": 0,
+        "llm_tokens_used": 0,
+        "previous_text": "",
+        "previous_type": "",
+        "previous_order": 0
+      }
+    },
+    {
+      "page_id": 22,
+      "text_extraction_method": "pdftext",
+      "block_counts": [
+        [
+          "Span",
+          213
+        ],
+        [
+          "Line",
+          80
+        ],
+        [
+          "TableCell",
+          16
+        ],
+        [
+          "Text",
+          13
+        ],
+        [
+          "Figure",
+          4
+        ],
+        [
+          "Caption",
+          3
+        ],
+        [
+          "Reference",
+          3
+        ],
+        [
+          "PageHeader",
+          1
+        ],
+        [
+          "SectionHeader",
+          1
+        ],
+        [
+          "Table",
+          1
+        ],
+        [
+          "PageFooter",
+          1
+        ],
+        [
+          "FigureGroup",
+          1
+        ],
+        [
+          "TableGroup",
+          1
+        ]
+      ],
+      "block_metadata": {
+        "llm_request_count": 0,
+        "llm_error_count": 0,
+        "llm_tokens_used": 0,
+        "previous_text": "",
+        "previous_type": "",
+        "previous_order": 0
+      }
+    },
+    {
+      "page_id": 23,
+      "text_extraction_method": "pdftext",
+      "block_counts": [
+        [
+          "Span",
+          563
+        ],
+        [
+          "Line",
+          96
+        ],
+        [
+          "TableCell",
+          93
+        ],
+        [
+          "Text",
+          10
+        ],
+        [
+          "Caption",
+          2
+        ],
+        [
+          "Table",
+          2
+        ],
+        [
+          "Reference",
+          2
+        ],
+        [
+          "PageHeader",
+          1
+        ],
+        [
+          "SectionHeader",
+          1
+        ],
+        [
+          "PageFooter",
+          1
+        ],
+        [
+          "TableGroup",
+          1
+        ]
+      ],
+      "block_metadata": {
+        "llm_request_count": 0,
+        "llm_error_count": 0,
+        "llm_tokens_used": 0,
+        "previous_text": "",
+        "previous_type": "",
+        "previous_order": 0
+      }
+    },
+    {
+      "page_id": 24,
+      "text_extraction_method": "pdftext",
+      "block_counts": [
+        [
+          "Span",
+          139
+        ],
+        [
+          "Line",
+          66
+        ],
+        [
+          "PageHeader",
+          1
+        ],
+        [
+          "Figure",
+          1
+        ],
+        [
+          "Caption",
+          1
+        ],
+        [
+          "PageFooter",
+          1
+        ],
+        [
+          "Text",
+          1
+        ],
+        [
+          "FigureGroup",
+          1
+        ],
+        [
+          "Reference",
+          1
+        ]
+      ],
+      "block_metadata": {
+        "llm_request_count": 0,
+        "llm_error_count": 0,
+        "llm_tokens_used": 0,
+        "previous_text": "",
+        "previous_type": "",
+        "previous_order": 0
+      }
+    }
+  ],
+  "debug_data_path": "debug_data/c39d243f-0c59-4f70-8ee6-d9e742174491"
+}

icml26/c39d243f-0c59-4f70-8ee6-d9e742174491/model_text_v3.txt

@@ -0,0 +1,242 @@
+[p. 1 | section: Abstract | type: Text]
+Recent advances in Large Language Models (LLMs) and Vision-Language Models (VLMs) have enabled powerful semantic and multimodal reasoning capabilities, creating new opportunities to enhance sample efficiency, high-level planning, and interpretability in reinforcement learning (RL). While prior work has integrated LLMs and VLMs into various components of RL, the replay buffer, a core component for storing and reusing experiences, remains unexplored. We propose addressing this gap by leveraging VLMs to guide the prioritization of experiences in the replay buffer. Our key idea is to use a frozen, pre-trained VLM (requiring no fine-tuning) as an automated evaluator to identify and prioritize promising sub-trajectories from the agent's experiences. Across scenarios, including game-playing and robotics, spanning both discrete and continuous domains, agents trained with our proposed prioritization method achieve 11–52% higher average success rates and improve sample efficiency by 19–45% compared to previous approaches.
+
+[p. 1 | section: 1. Introduction | type: Text]
+Reinforcement learning (RL) has achieved impressive results across domains ranging from robotics to natural language systems (Levine et al., 2016; Ouyang et al., 2022) and logistics management (Tang et al., 2025) . Despite this success, RL remains notoriously sample-inefficient, sensitive to hyperparameters, and heavily dependent on carefully designed reward functions (Kalashnikov et al., 2021; Kroemer et al., 2021; Zhu et al., 2020) .
+
+[p. 1 | section: 1. Introduction | type: Text]
+A well-established line of research—off-policy RL (Lin, 1992; Mnih et al., 2013) —addresses the efficiency bottleneck through experience reuse. During training, transitions
+
+[p. 1 | section: 1. Introduction | type: Text]
+are collected into a dataset known as a replay buffer , and sampled multiple times to update the policy. Ideally, a sampling mechanism would prioritize the most meaningful transitions: those that provide the richest learning signal, irrespective of when they were collected. The core challenge is to identify which experiences are truly meaningful for learning process.
+
+[p. 1 | section: 1. Introduction | type: Text]
+Conventional methods such as Prioritized Experience Replay (PER, Schaul et al. 2015b) approximate this ideal by prioritizing transitions with high temporal-difference (TD) error. The TD error measures the discrepancy between a bootstrapped value estimate and its target (Sutton et al., 1998) . In effect, this approach promotes transitions which are misaligned with the agent's future outcome predictions. While TD error prioritization can correct value estimates, it is fundamentally limited by its lack of semantic awareness. It cannot distinguish transitions which reflect genuine progress toward task completion from those that do not.
+
+[p. 1 | section: 1. Introduction | type: Text]
+This limitation is particularly evident in long-horizon, sparse-reward tasks, such as robotic manipulation in OG-Bench (Park et al., 2024) . For instance, critical transitions—such as unlatching a door or grasping a tool—may yield low TD errors early in training. This is a consequence of delayed credit assignment preventing the reward signal from propagating to these earlier states. Conversely, visually distinct but task-irrelevant motions can yield large TD errors, despite contributing little to actual task completion.
+
+[p. 1 | section: 1. Introduction | type: Text]
+To bridge this gap, we propose replacing heuristic inductive biases with external sources of semantic understanding. Recent literature highlights the potential of integrating largescale Vision-Language Models (VLMs) with RL to leverage pre-trained world knowledge (Liang et al., 2024; Zhang et al., 2024) . Because VLMs combine visual perception with language reasoning, they can interpret complex environments and mitigate RL-specific challenges—such as sparse rewards or inefficient exploration—by providing enriched context (Cao et al., 2024; Schoepp et al., 2025) . This motivates our primary research question:
+
+[p. 1 | section: 1. Introduction | type: Text]
+Can the semantic knowledge of pre-trained VLMs be used to prioritize meaningful experiences in replay buffers?
+
+[p. 1 | section: 1. Introduction | type: Text]
+To answer this, we introduce VLM-RB (Fig .1) , which integrates a pre-trained VLM directly into the experi-
+
+[p. 2 | section: 1. Introduction | type: Caption]
+Figure 1. System Diagram: (1) Data is collected with the current policy \pi_k interacting with multiple instances of the environment. (2) Transitions (s, a, r, s') are kept in a prioritized replay buffer with a default priority \bar{\mathbf{p}} . (3) Asynchronously, after each insertion, a VLM worker scores the corresponding rendered clip \tau^O under a prompt P and writes the resulting priority \mathbf{p^{VLM}} back to the replay buffer (see Section 3.1). (4) The learner samples a minibatch using the mixture distribution q_t which interpolates between uniform and VLM-prioritized replay (see Section 3.1). (5) The agent policy is updated to \pi_{k+1} .
+
+[p. 2 | section: 1. Introduction | type: Text]
+ence prioritization pipeline. In contrast to heuristics relying on statistical proxies such as TD error, uncertainty (Carrasco-Davis et al., 2025), or density ratios (Zhao & Tresp, 2019), VLM-RB leverages the reasoning capabilities of VLMs to score experiences according to their semantic relevance.
+
+[p. 2 | section: 1. Introduction | type: Text]
+We design this framework as a modular plug-and-play prioritization layer that can be integrated with any off-policy algorithm that uses a replay buffer. By using frozen off-the-shelf VLMs, we leverage strong semantic priors while avoiding the computational cost of fine-tuning or additional inference latency at deployment. Within the training pipeline, VLM-RB promotes efficiency by querying the VLM asynchronously, thereby sidestepping the blocking inference bottleneck that often constrains the throughput of LLM/VLM-enhanced RL methods.
+
+[p. 2 | section: 1. Introduction | type: Text]
+We empirically show that pre-trained VLMs yield semantically grounded prioritization signals which are closely aligned with task objectives. Leveraging this semantic guidance, VLM-RB consistently outperforms existing baselines in both discrete game-playing and continuous robotic manipulation. Importantly, the moderate throughput overhead (about 12%) is more than offset by substantial gains in sample efficiency (19–45%) and average success rates (11–52%).
+
+[p. 2 | section: 2. Preliminaries | type: Text]
+Off-Policy Reinforcement Learning. We consider the standard RL framework, where an agent interacts with an environment modeled as an infinite-horizon Markov Decision Process (MDP) \mathcal{M} = (\mathcal{S}, \mathcal{A}, P, r, \gamma, \rho_0) , with state space \mathcal{S} , action space \mathcal{A} (either discrete or continuous),
+
+[p. 2 | section: 2. Preliminaries | type: Text]
+transition kernel P: \mathcal{S} \times \mathcal{A} \to \Delta(\mathcal{S}) , reward function r: \mathcal{S} \times \mathcal{A} \to \mathbb{R} , discount factor \gamma \in (0,1) and initial-state distribution \rho_0 \in \Delta(\mathcal{S}) . We further assume access to a rendering function \psi: \mathcal{S} \to \mathcal{O} \subseteq \mathbb{R}^{H \times W \times 3} which maps states to H \times W -sized visual observations o_t = \psi(s_t) . The agent learns a stochastic policy \pi_\theta: \mathcal{S} \to \Delta(\mathcal{A}) to maximize the expected discounted return
+
+[p. 2 | section: 2. Preliminaries | type: Equation]
+J(\pi) = \mathbb{E}_{\tau \sim p_{\pi}(\tau)} \left[ \sum_{t=0}^{\infty} \gamma^{t} r_{t} \right],
+
+[p. 2 | section: 2. Preliminaries | type: Text]
+where \tau = (s_0, a_0, r_0, s_1, a_1, r_1, ...) denotes a trajectory and p_{\pi}(\tau) is the trajectory distribution induced by policy \pi .
+
+[p. 2 | section: 2. Preliminaries | type: Text]
+To evaluate and improve policies, value functions are used to estimate expected returns. The action-value function (Q-function) Q^{\pi}: \mathcal{S} \times \mathcal{A} \to \mathbb{R} represents the expected discounted return from state s after taking action a and then following policy \pi :
+
+[p. 2 | section: 2. Preliminaries | type: Equation]
+Q^{\pi}(s, a) = \mathbb{E}_{\tau \sim p_{\pi}(\tau \mid s_0 = s, a_0 = a)} \left[ \sum_{t=0}^{\infty} \gamma^t r_t \right].
+
+[p. 2 | section: 2. Preliminaries | type: Text]
+In modern RL methods, Q^{\pi} is typically a learned function (represented by a neural network), denoted the critic .
+
+[p. 2 | section: 2. Preliminaries | type: Text]
+In off-policy RL, rather than exclusively using on-policy data, optimization is performed using previously collected experience stored in a replay buffer (Mnih et al., 2013). Let \mathcal{D}_t denote the replay buffer at time t. We sample training tuples (s, a, r, s') \in \mathcal{D}_t according to a time-dependent distribution q_t \in \Delta(\mathcal{D}_t) . In standard uniform replay (UER), q_t(i) = 1/|\mathcal{D}_t| for all i \in \mathcal{D}_t . This formulation highlights that the optimization dynamics depend on the evolving sampling distribution q_t .
+
+[p. 3 | section: 2. Preliminaries | type: Caption]
+Figure 2. Temporal context resolves visual ambiguity. From the initial state o_i (left), multiple futures are possible. The top sub-trajectory shows a successful grasp, while the bottom shows stagnation. By scoring sub-trajectories rather than single frames, the VLM has sufficient context to distinguish meaningful progress from failure modes.
+
+[p. 3 | section: 2. Preliminaries | type: Text]
+Experience Replay Prioritization. As opposed to uniform replay, prioritized replay methods bias retrieval toward transitions expected to yield a stronger learning signal. For clarity, we formalize this process as three distinct steps: scoring, prioritization, and sampling. Scoring assigns each transition i \in \mathcal{D}_t a scalar value \mathbf{p}_i \in \mathbb{R} reflecting its heuristic utility. Prioritization maps these raw scores into a probability distribution over the buffer, q_t^{\mathbf{P}} \in \Delta(\mathcal{D}_t) . Sampling then draws a minibatch of indices \{i_k\}_{k=1}^B from a target distribution q_t , or a mixture of several such distributions.
+
+[p. 3 | section: 2. Preliminaries | type: Text]
+Prioritized Experience Replay ( PER , Schaul et al. 2015b) is a specific instance of this framework. Its scoring function utilizes the temporal-difference (TD) error magnitude, \mathbf{p}_i = |\delta_i| + \epsilon^1 , where \delta_i is the discrepancy between the value estimate and its target<sup>2</sup>. The prioritized distribution is defined as q_t^{\mathbf{P}}(i) \propto \mathbf{p}_i^{\alpha} , where \alpha \in [0,1] determines the degree of prioritization. While PER effectively focuses updates on transitions with larger TD errors, it relies on value estimates which may be arbitrary and noisy early in training, particularly in sparse-reward settings.
+
+[p. 3 | section: 2. Preliminaries | type: Text]
+Vision–Language Models. Recent advances in large-scale representation learning have given rise to Vision–Language Models (VLMs), which learn aligned embeddings of visual and textual modalities through joint pre-training on large-scale image–text corpora (Radford et al., 2021; Jia et al., 2021; Zhai et al., 2023). Formally, a VLM defines encoders f_{\text{img}}: \mathcal{I} \to \mathbb{R}^d and f_{\text{txt}}: \mathcal{T} \to \mathbb{R}^d mapping images and text to a shared latent space, typically optimized via contrastive objectives which encourage semantic correspondence between paired inputs (Cherti et al., 2023; Zhai et al., 2022; Goel et al., 2022). These models have demonstrated strong generalization across a wide
+
+[p. 3 | section: 2. Preliminaries | type: Text]
+range of downstream tasks, including zero-shot recognition, goal specification, and reward inference (Ahn et al., 2022; Rocamonde et al., 2023). While standard VLMs process static images, Video Question Answering (Video-QA) models extend these capabilities to ingest a temporal sequence of frames (a video clip) along with a natural-language query. By aggregating visual information across time, these models can reason about events, motion, and causal relationships. Concretely, for the rest of this paper, we employ Perception-LM (Cho et al., 2025), a state-of-the-art opensource Video-QA family of models with multiple model sizes (1B/3B/8B).
+
+[p. 3 | section: 3. VLM-RB | type: Text]
+In this section, we present VLM-RB (Fig.1 and Algorithm 1), a plug-and-play framework designed for any off-policy algorithm that leverages a replay buffer. The central idea is to use a VLM to assign semantic scores to subtrajectories, subsequently biasing the sampling distribution toward higher-scoring experiences during policy optimization.
+
+[p. 3 | section: 3.1. Scoring, Prioritization, and Sampling | type: Text]
+We now define how semantic scores are extracted from collected data, how these scores induce a prioritization distribution, and how VLM-RB samples from this distribution.
+
+[p. 3 | section: 3.1. Scoring, Prioritization, and Sampling | type: Text]
+Scoring. VLM-RB uses a pre-trained frozen VLM to extract semantic scores from collected experiences. To enable this, we first construct temporal visual sequences from the agent's history. Formally, let \tau_i^O = (o_i, o_{i+1}, \dots, o_{i+L-1}) denote a visual clip comprising L rendered frames. The VLM maps this clip and a text prompt P to a scalar score:
+
+[p. 3 | section: 3.1. Scoring, Prioritization, and Sampling | type: Equation]
+\mathbf{p}^{\mathbf{VLM}} = f_{\mathbf{VLM}}(\tau^O, \mathsf{P}) \in \mathbb{R}. (1)
+
+[p. 3 | section: 3.1. Scoring, Prioritization, and Sampling | type: Text]
+The prompt P directs the VLM's scoring mechanism, leveraging the model's inherent world knowledge to identify meaningful behaviors even in the absence of dense reward signals. While this interface supports the injection of detailed and task-specific priors, we find that the VLM's intrinsic scene understanding is sufficient for our purposes. We therefore employ a general task-agnostic prompt (see Appendix C for details).
+
+[p. 3 | section: 3.1. Scoring, Prioritization, and Sampling | type: Text]
+For simplicity, we set f_{\rm VLM} as a binary indicator, assigning a score of 1 to clips exhibiting seemingly meaningful behavior, and 0 otherwise. This effectively labels useful sequences without the need for hand-crafted task-specific definitions. Crucially, because the VLM is frozen and its evaluation depends only on static visual content, each clip is scored exactly once. This offers a significant efficiency advantage over methods such as PER , which require ongoing priority updates as the value function evolves.
+
+[p. 3 | section: 3.1. Scoring, Prioritization, and Sampling | type: Footnote]
+& lt;sup>1</sup>Here \epsilon > 0 ensures non-zero probability.
+
+[p. 3 | section: 3.1. Scoring, Prioritization, and Sampling | type: Footnote]
+& lt;sup>2</sup>Namely, \delta_i = Q_{\theta}(s_t, a_t) - (r + \gamma Q_{\theta}(s_{t+1}, a')) where \theta are the critic parameters and a' is the action predicted by the current policy in s_{t+1} . In Q-learning methods, a' is the argmax action, and in actor-critic methods for continuous control, it is \pi(s_{t+1}) .
+
+[p. 4 | section: 3.1. Scoring, Prioritization, and Sampling | type: Text]
+A natural question is whether scoring individual frames, rather than clips, would suffice. We argue that single frames are fundamentally limited by semantic ambiguity . To illustrate this, consider Fig. 2: a single observation of a robotic gripper hovering above an object. This static frame is visually identical across two vastly different temporal contexts: the onset of a successful grasp or the aftermath of a failed attempt.
+
+[p. 4 | section: 3.1. Scoring, Prioritization, and Sampling | type: Text]
+Without temporal context a VLM cannot disambiguate these scenarios, potentially assigning high scores to frames which actually belong to failure modes. By scoring sub-trajectories instead, we increase the likelihood of the VLM having sufficient temporal information to distinguish meaningful behaviors from failures. While a sliding-window variant could potentially provide denser labels, we leave this (computationally intensive) alternative for future work.
+
+[p. 4 | section: 3.1. Scoring, Prioritization, and Sampling | type: Text]
+Prioritization. We construct a prioritized distribution q P by propagating the VLM score of each clip to all transitions within that clip. Under our binary scoring scheme (p i ∈ {0, 1}), q P becomes uniform over the subset of transitions labeled as semantically meaningful.
+
+[p. 4 | section: 3.1. Scoring, Prioritization, and Sampling | type: Text]
+Sampling. If we were to sample only from q P , we would discard all transitions labeled uninteresting by the VLM. To avoid wasting collected data and to ensure the agent explores the full state space, we instead use a mixture strategy q t interpolating between VLM-guided prioritization and uniform replay:
+
+[p. 4 | section: 3.1. Scoring, Prioritization, and Sampling | type: Equation]
+q_t(i) = \lambda_t q^{\mathbf{P}}(i) + (1 - \lambda_t) q^{\mathbf{U}}(i), \tag{2}
+
+[p. 4 | section: 3.1. Scoring, Prioritization, and Sampling | type: Text]
+where λ t ∈ [0, 1] controls the strength of the VLM guidance. In practice, each batch draws a λ t fraction from q P and the remainder uniformly. We use a linear warm-up schedule: starting with λ 0 = 0 (pure uniform sampling), we gradually anneal to λmax = 0.5 over the first half of training. We hypothesize that this schedule is essential: early in training, broad coverage stabilizes value learning, while later updates bias toward high-utility regions. Our ablation studies (Section 4.4) support this design: purely prioritized sampling is detrimental, but the proposed mixture schedule yields significant efficiency gains.
+
+[p. 4 | section: 3.2. Efficient Implementation | type: Text]
+Incorporating a VLM into the RL loop introduces significant computational overhead relative to standard components. We address this with two key design choices. First, we leverage the decoupling between data collection and policy optimization in off-policy learning to asynchronously score experiences. As illustrated in Fig. 1, the VLM interacts with the replay buffer in the background, ensuring that policy optimization is never blocked by inference latency.
+
+[p. 4 | section: 3.2. Efficient Implementation | type: Text]
+Beyond minimizing latency, this decoupled architecture renders VLM-RB agnostic to the policy input modality: while the VLM requires rendered frames, the policy itself can operate on arbitrary observation spaces such as lowdimensional states, enabling more sample-efficient learning. This architecture further allows a single VLM instance to efficiently serve multiple parallel environments. Second, we use a lightweight 1B parameter model (Cho et al., 2025) . Our ablations show that this model is sufficient to identify meaningful behaviors while maintaining high throughput 3 (see Appendix F.2) .
+
+[p. 4 | section: 3.3. Boosting with TD-error | type: Text]
+We can further refine prioritization by incorporating the TD-error, defining q P (i) ∝ p VLM i · |δ i |, where δ i emphasizes transitions with high prediction error (where the value function is inaccurate), and p VLM i emphasizes semantic relevance. Because the VLM score is a binary indicator, it effectively masks "irrelevant" transitions and promotes the remaining transitions based on their TD errors.
+
+[p. 4 | section: 3.3. Boosting with TD-error | type: Text]
+In this scheme, we maintain two scores per transition in the buffer: q P (i) ∝ p VLM i · |δ i | and δ i . The TD error δ i is updated each time the transition is sampled, while q P (i) ∝ p i · |δ i | is updated only once, as described in Section 3.2.
+
+[p. 4 | section: 4. Experiments | type: Text]
+In this section, we evaluate the efficacy of VLM-RB in leveraging semantic priors for efficient exploration. We first ask: does the VLM signal provide meaningful guidance for exploration? To answer this, we analyze its correlation with learned value estimates and its dependence on visual semantics (Section 4.2) . Next, we benchmark VLM-RB against UER, PER, and alternative prioritization schemes, evaluating both performance and sample efficiency across a range of discrete and continuous control tasks (Section 4.3) . Finally, we examine which design choices most affect the performance of VLM-RB, focusing on the sampling mixture and VLM model size (Section 4.4) .
+
+[p. 4 | section: 4.1. Experimental Setup | type: Text]
+Tasks. We evaluate VLM-RB on two domains with discrete and continuous action spaces: (i) DoorKey from MiniGrid (Chevalier-Boisvert et al., 2023) , using grid sizes 8x8, 12x12, and 16x16 to vary exploration difficulty; and (ii) scene from OGBench (Park et al., 2024) , using the predefined tasks 3, 4, and 5, which require increasingly long-horizon compositional manipulation (unlocking/locking and coordinated object placement). In all experiments, agents receive state-based observations (rather
+
+[p. 4 | section: 4.1. Experimental Setup | type: Footnote]
+3 Scaling to larger variants (3B or 8B) yields diminishing returns in downstream RL performance (see Appendix F.2) .
+
+[p. 5 | section: 4.1. Experimental Setup | type: FigureGroup]
+Figure 3. Frozen VLM scoring anticipates learned value. We visualize a single reference episode in MiniGrid/DoorKey-16x16 which contains both goal-directed actions and random noise. Left: Visualizations of two 32-frame sub-trajectories. clip A: captures a semantically relevant sequence. clip B: shows a mostly random walk which eventually reaches the goal. Right Top: A timeline of ground-truth events; sparse reward is only received at the final goal. Right Bottom: The rose curve shows the frozen VLM score for L=32-frame clips. The gray curves show the temporal value difference, \Delta Q(t;L)=Q(t+\frac{L}{2})-Q(t-\frac{L}{2}) , calculated from critic checkpoints at increasing training steps (light to dark). Early in training (light gray), the critic is uninformative, while later checkpoints (dark gray) increasingly assign positive value to the same semantic events the VLM identified. This demonstrates that the VLM provides a helpful signal long before the critic successfully converges.
+
+[p. 5 | section: 4.1. Experimental Setup | type: Text]
+than pixels) to minimize perception-related confounders and isolate the effect of replay prioritization on exploration and sample efficiency. More details about the various tasks are provided in Appendix B.
+
+[p. 5 | section: 4.1. Experimental Setup | type: Text]
+Baselines. We compare VLM-RB to two standard replay sampling methods: Uniform Experience Replay (UER), which samples transitions uniformly from the replay buffer, and PER, which prioritizes sampling by TD-error. We report UER and PER comparisons for four different algorithms: DQN (Van Hasselt et al., 2016), IQN (Dabney et al., 2018), TD3 (Fujimoto et al., 2018), and SAC (Haarnoja et al., 2018). In addition, to compare VLM-RB against alternative replay prioritization methods, we run a focused ablation for DQN comparing to Attentive Experience Replay (AER, Sun et al. 2020), Experience Replay Optimization (ERO, Zha et al. 2019), and Reducible Loss Prioritization (ReLO, Sujit et al. 2023). Implementation details for all baselines are provided in Appendix E.
+
+[p. 5 | section: 4.1. Experimental Setup | type: Text]
+Method Configuration. To adapt to the specific characteristics of the experiment domains, we employ two variants of our prioritization scheme. For the discrete MiniGrid tasks, where semantic progress is binary (e.g., carrying key vs. not), we rely strictly on the binary VLM semantic filter as defined in Section 3.1. Conversely, for the continuous control tasks in OGBench, where fine-grained motion control is required, we utilize the TD-error boosted variant described in Section 3.3. This allows the agent to combine the high-level semantic filtering of the VLM with information
+
+[p. 5 | section: 4.1. Experimental Setup | type: Text]
+from the TD error.
+
+[p. 5 | section: 4.1. Experimental Setup | type: Text]
+Metrics. We report the SUCCESS RATE (SR), defined as the fraction of episodes in which the agent completes the task within the maximum episode length, T_{\rm max} . The AVERAGE SUCCESS RATE (ASR) is the mean SR over N=32 evaluation episodes for a given seed. Final results are the ASR averaged over M=5 seeds, with the standard error of the mean (SEM) shown as shaded regions in plots.
+
+[p. 5 | section: 4.2. Do VLMs Contain Human-Like Priors? | type: Text]
+Before turning to downstream RL results, we first ask whether the frozen VLM offers a signal that is actually useful for prioritizing data. To understand the semantic grounding of our prioritization, we examine how the frozen VLM scores correlate with the critic value estimates ( \Delta Q ) as learning progresses. Concretely, we track how the value estimates of the critic evolve on a fixed reference episode, using checkpoints from a successful training run. As shown in Fig. 3, the critic's estimates are initially flat and uninformative (light gray lines). Over the course of training, the critic gradually learns to assign high value to the same semantic events (such as picking up a key or opening a door) that the VLM identified from the outset. This observation suggests that VLM-RB can immediately identify semantically relevant data, enabling faster learning than approaches which must wait for the critic to converge.
+
+[p. 5 | section: 4.2. Do VLMs Contain Human-Like Priors? | type: Text]
+Having demonstrated that VLM scores correlate with well-trained Q-value predictors, we next ask: does this
+
+[p. 6 | section: 4.2. Do VLMs Contain Human-Like Priors? | type: FigureGroup]
+Figure 4. Success depends on alignment with semantic priors. We compare the performance of VLM-RB across frames rendered with Standard visuals (dark circles), Misleading swapped sprites (medium squares), and Abstract textures (light diamonds). The dashed line indicates the ASR of PER. Crucially, the agent's actual observations and the underlying MDP are identical across all settings; only the visual input to the VLM is altered (see Appendix F.3 for visual samples of these modifications). The significant performance drop in the Misleading and Abstract settings confirms that our method relies on the VLM correctly recognizing specific semantic objects (e.g., keys and doors) to accelerate learning.
+
+[p. 6 | section: 4.2. Do VLMs Contain Human-Like Priors? | type: Text]
+signal reflect interpretable, human-like priors? To test whether VLM-RB specifically leverages the VLM's pretrained visual semantics, we employ the "Modified Game" paradigm (Dubey et al., 2018) . We modify only the rendered frames used for VLM scoring, leaving the underlying MDP and agent observations untouched. We consider three settings: (i) the Standard game, with unaltered visuals; (ii) the Misleading game, where key objects are swapped (for example, traps appear as goals); and (iii) the Abstract game, where objects are replaced with random noise patterns (See Appendix F.3 for visual examples).
+
+[p. 6 | section: 4.2. Do VLMs Contain Human-Like Priors? | type: Text]
+As shown in Fig. 4, the Misleading and Abstract variants significantly slow learning and increase variance. Notably, in the Abstract setting, performance degrades to the level of PER (dashed line). Because the underlying MDP remains unchanged, this degradation demonstrates that the VLM's prioritization is only effective when the visual input aligns with natural semantic priors.
+
+[p. 6 | section: 4.2. Do VLMs Contain Human-Like Priors? | type: Text]
+In summary, these results confirm that VLM-RB scores are indicative of meaningful semantic behavior. We now turn to the question: does this semantic guidance actually yield improved sample efficiency and asymptotic performance in off-policy RL?
+
+[p. 6 | section: 4.3. Main Results: VLMs are Useful for RL Data Prioritization | type: Text]
+We evaluate VLM-RB and baselines across both discrete and continuous environments, as detailed in Section 4.1.
+
+[p. 6 | section: 4.3. Main Results: VLMs are Useful for RL Data Prioritization | type: TableGroup]
+Table 1. Improvement with respect to baselines. Performance denotes the best ASR (M = 5 seeds). Sample Efficiency tracks the steps required to reach the baseline's best performance. Values in parentheses indicate the relative improvement over the baseline. Both metrics are averaged across the aggregated algorithms. The corresponding training curves are provided in Fig. 9. Improvement Algorithm Lvl. Base. Perf. (↑) Sample Eff. (↓) 8×8 UER PER +0.0% (1.00/1.00) +0.0% (1.00/1.00) +19.1% (144K/178K) +23.0% (144K/187K) DQN+IQN (DoorKey) 12×12 UER +61.3% (1.00/0.62) +52.8% (302K/640K) PER +22.0% (1.00/0.82) +32.1% (317K/467K) 16×16 UER +241.7% (0.82/0.24) +37.6% (557K/893K) PER +70.8% (0.82/0.48) +24.1% (472K/622K) 3 UER PER +0.0% (1.00/1.00) +0.0% (1.00/1.00) +40.7% (210K/354K) +21.1% (210K/266K) SAC+TD3 (Scene) 4 PER +2.0% (1.00/0.98) UER +22.0% (1.00/0.82) +44.6% (430K/776K) +19.7% (509K/634K) 5 UER +119.4% (0.79/0.36) +46.3% (440K/819K) PER +49.1% (0.79/0.53) +17.9% (661K/805K)
+
+[p. 6 | section: 4.3. Main Results: VLMs are Useful for RL Data Prioritization | type: Text]
+Table 1 summarizes the aggregate improvements in both performance and sample efficiency. VLM-RB consistently outperforms all baselines, regardless of algorithm or task difficulty. The largest gains appear in the most challenging settings: on DoorKey-16x16, VLM-RB achieves a +241.7% improvement over UER and +70.8% over PER; on Scene-5, the improvement over UER is +119.4% and +49.1% over PER. Even in cases where baselines eventually solve the task, VLM-RB achieves substantially better sample efficiency, reducing the number of steps needed to reach baseline performance by as much as 52.8%.
+
+[p. 6 | section: 4.3. Main Results: VLMs are Useful for RL Data Prioritization | type: Text]
+Comparison to Baselines. Figure 5 expands this comparison, showing that VLM-RB is the only method to consistently solve the task when compared to alternative prioritization schemes (AER, ERO, ReLo). The baselines struggle primarily due to the combination of sparse rewards and long sequential dependencies. ERO and ReLo, for example, fail to prioritize task-relevant transitions because their feedback signals (reward gradients and TD-error differences) are either dominated by noise or remain zero throughout the long pre-success phase. On the other hand, similarity-based methods like AER suffer from a structural misalignment: their local similarity metrics lead to the over-sampling of current states while neglecting critical past dependencies (such as key acquisition), which impedes temporal credit assignment over long horizons. Further analysis is provided in Appendix D.
+
+[p. 6 | section: 4.3. Main Results: VLMs are Useful for RL Data Prioritization | type: Text]
+Training Time Finally, we address the computational trade-off. Although VLM inference introduces a moderate throughput overhead (approximately 12% wall-clock
+
+[p. 7 | section: 4.3. Main Results: VLMs are Useful for RL Data Prioritization | type: Caption]
+prioritization in sparse-reward tasks. Aggregated success rates on MiniGrid/DoorKey across grid sizes (8x8, 12x12, 16x16). While alternative methods (AER, ERO, ReLo) fail as they depend on dense rewards or local similarity metrics, VLM-RB successfully bridges the long-horizon dependencies. Curves show the mean across 5 seeds, with shaded standard errors.
+
+[p. 7 | section: 4.3. Main Results: VLMs are Useful for RL Data Prioritization | type: Text]
+time per step, see Appendix F.4) , this cost is outweighed by the reduction in training steps. In the vast majority of configurations (21/24), VLM-RB reduces the absolute wall-clock time to reach peak performance by up to 58%. In the rare cases where wall-clock time increased, the slowdown was marginal (<5%). By contrast, in more challenging exploration settings, VLM-RB consistently achieved speedups exceeding 20-40% (full results can be seen in Table 7 in Appendix G) . Overall, VLM-RB achieves a favorable tradeoff, substantially accelerating training in difficult sparsereward tasks, while incurring negligible overhead in simpler scenarios.
+
+[p. 7 | section: 4.4. Which Design Choices Matter? Mixing and Scale | type: Text]
+Sampling Mixing Schedule. A central component of our approach is to mix VLM-scored samples with uniformly sampled data (see Eq. 2) . To assess the importance of this feature, we vary only the maximal mixing coefficient, λmax, holding all other parameters fixed. The results shown in Table 4 indicate that retaining a fraction of uniform sampling is necessary to stabilize value learning. See Appendix F.1 for the full analysis of the results.
+
+[p. 7 | section: 4.4. Which Design Choices Matter? Mixing and Scale | type: Text]
+VLM Size. We also ablated the VLM size and found that scaling the model beyond 1B parameters dramatically increases computational cost without yielding consistent improvements in downstream RL performance. Detailed analysis of the trade-offs between throughput, memory, and performance is provided in Appendix F.2.
+
+[p. 7 | section: 5. Related Work | type: Text]
+Prioritized Experience Replay. Extensive research has focused on defining which transitions maximize learning utility. The seminal work, PER (Schaul et al., 2015b) , prioritizes transitions with high TD error, serving as a proxy for "surprising" events. Subsequent methods have proposed alternative utility metrics based on state similarity (Sun et al., 2020) , potential loss reduction (Sujit et al., 2023) , target quality/discrepancy criteria (Kumar et al., 2020) , or by explicitly learning a replay policy (Zha et al., 2019) . However, these approaches typically derive priorities from internal, training-dependent signals (e.g., value estimates or TD errors), which can be noisy or undefined early in training. In contrast, VLM-RB leverages a pre-trained VLM to assign semantic priorities, yielding a robust utility score from the very first update.
+
+[p. 7 | section: 5. Related Work | type: Text]
+Orthogonal to the choice of what to prioritize, several works modify how transitions are sampled. Hong et al. (2022) argue for enforcing the ordering of updates within a collected episode, to better align with the Bellman loss that is common in off-policy learning. Another example, Lahire et al. (2021) , focuses on the "staleness" of prioritization scores and suggests a double-sampling approach. These directions are largely complementary to VLM-RB and can be combined with VLM-based priorities. Furthermore, because VLM-RB computes priorities from a frozen VLM, it naturally avoids the score staleness issue addressed by the latter.
+
+[p. 7 | section: 5. Related Work | type: Text]
+Finally, rather than treating prioritization as a standalone sampling heuristic, a distinct line of work alters the underlying optimization objective itself. These methods often derive reweighting schemes from first principles, using fdivergence regularization (Li et al., 2024) , trust-region constraints (Novati & Koumoutsakos, 2019) , or density-ratio estimation (Liu et al., 2021; Sinha et al., 2022) to justify non-uniform data usage. Unlike these approaches, which fundamentally change the loss function, VLM-RB preserves the underlying training objective, ensuring it remains modular and compatible with a wide range of existing algorithms.
+
+[p. 7 | section: 5. Related Work | type: Text]
+LLM/VLM-Enhanced RL. Recent work has increasingly used foundation models to improve RL by injecting semantic knowledge into the learning loop. Broadly, these methods leverage LLMs/VLMs to (i) define richer learning signals (reward supervision), (ii) propose plans and exploration targets (high-level guidance), or (iii) assist action selection (policy components). Comprehensive surveys can be found in Cao et al. (2024) and Schoepp et al. (2025) .
+
+[p. 7 | section: 5. Related Work | type: Text]
+A major thread treats LLMs and VLMs as sources of reward. Off-the-shelf VLMs have been used as zero-shot reward scorers to judge state–goal alignment (Rocamonde et al.,
+
+[p. 8 | section: 5. Related Work | type: Text]
+2023) ; LLMs can synthesize dense shaping functions from natural-language task descriptions (Xie et al., 2023) ; and several methods replace or augment human feedback by using VLM/LLM judgments (absolute ratings or pairwise preferences) to train reward models (Fu et al., 2024; Luu et al., 2025; Singh et al., 2025; Ghosh et al., 2025) . Taken together, these approaches reframe reward design as semantic labeling, allowing pre-trained models to provide informative supervision in settings where environment rewards are sparse or absent.
+
+[p. 8 | section: 5. Related Work | type: Text]
+A second strand uses foundation models for planning and exploration. LLMs decompose instructions into temporallyextended plans grounded in affordances (Ahn et al., 2022; Huang et al., 2022) , and guide exploration by proposing semantically diverse subgoals or novelty-oriented objectives using learned embeddings (Ma et al., 2025; Gupta et al., 2022) . This high-level guidance is particularly useful in settings where unguided exploration rarely visits meaningful states.
+
+[p. 8 | section: 5. Related Work | type: Text]
+A third direction integrates LLMs/VLMs more directly into control, either as action advisors or as policy components. Examples range from generating executable policy code (Liang et al., 2022) to large-scale vision–language– action (VLA) models that map multimodal inputs to actions (Zitkovich et al., 2023) .
+
+[p. 8 | section: 5. Related Work | type: Text]
+Despite this breadth, most prior work uses these models to shape what the agent optimizes (rewards), where it searches (plans and exploration), or how it acts (policy), which commonly introduces additional training stages, fine-tuning, or computationally expensive inference. In contrast, VLM-RB targets the replay buffer—the core mechanism enabling sample reuse in off-policy RL—a component which remains underexplored for semantic foundation-model supervision. Moreover, as our experiments show, open-source pre-trained VLMs can already yield consistent gains when used to prioritize replay, without any additional model training or finetuning.
+
+[p. 8 | section: 6. Conclusions | type: Text]
+In this work, we presented VLM-RB, a framework that integrates the semantic reasoning capabilities of pre-trained VLMs directly into the experience replay mechanism. By shifting prioritization from statistical proxies such as TDerror to semantic evaluations of task progress, VLM-RB effectively identifies and promotes meaningful experiences even in sparse-reward regimes where traditional metrics fail. Our empirical results demonstrate that this approach significantly improves both sample efficiency and asymptotic performance across discrete and continuous domains, without requiring fine-tuning or gradient updates to the VLM.
+
+[p. 8 | section: 6. Conclusions | type: Text]
+Limitations. While our approach yields substantial improvements, it is important to clarify the specific assumptions and constraints under which these gains are realized. First, VLM-RB fundamentally relies on the assumption that task progress can be reliably inferred from visual observations. As a result, VLM-RB does not apply to domains where the underlying state lacks a visual representation, such as non-spatial biological systems or abstract network control tasks. In these settings, the VLM cannot ground its reasoning. Second, querying the VLM introduces additional computational overhead. Empirically, we found that this overhead reduced throughput by approximately 12% in our experiments. It is important to note that this computational cost is not intrinsic to the method; the precise overhead depends on hardware configuration and implementation choices, so the trade-off may vary across different setups 4 .
+
+[p. 8 | section: 6. Conclusions | type: Text]
+Future Work. Our findings suggest several concrete directions for future research. A natural extension is the application of VLM-RB to Goal-Conditioned RL (Schaul et al., 2015a) , where the textual description of the current goal can be injected into the VLM prompt to dynamically score alignment, or hindsight textual goals could promote meaningful sub-goals which may have some semantic relation but no direct relevance to the current task (Luu & Yoo, 2021; Jiang et al., 2019) . Another promising direction is to use the replay buffer as a curriculum mechanism, by evolving the prompt over the course of training. For example, the system could prioritize specific skills such as "open drawer" in the early stages, and later shift focus to more compositional behaviors. Finally, the prioritization scheme could be extended beyond single VLM scoring, for example, by incorporating more robust inference techniques such as majority voting or LLM-as-a-Judge.
+
+[p. 8 | section: 6. Conclusions | type: Text]
+Impact Statement. This paper presents work aimed at advancing the field of machine learning, specifically in the domain of sample-efficient reinforcement learning. There are many potential societal consequences of our work, none of which we feel must be specifically highlighted here.

icml26/c39d243f-0c59-4f70-8ee6-d9e742174491/parse_report.json

@@ -0,0 +1,78 @@
+{
+  "paper_id": "c39d243f-0c59-4f70-8ee6-d9e742174491",
+  "pipeline": "marker_non_llm_v3",
+  "parser": "marker_single",
+  "formats": [
+    "markdown",
+    "chunks"
+  ],
+  "llm_enabled": false,
+  "pdf_path": "/network/scratch/j/jianan.zhao/ReviewAgent/data/processed_papers/icml26_20260429_1952_duequeue/raw/c39d243f-0c59-4f70-8ee6-d9e742174491.pdf",
+  "pdf_sha256": "09e6f1c83579f310aa39f8fa0d9a2b04d87624df346bb40d0e7df8c3c194f682",
+  "bytes": 2607069,
+  "source": "https://koala.science/storage/pdfs/c39d243f-0c59-4f70-8ee6-d9e742174491.pdf",
+  "page_count": 25,
+  "ok": true,
+  "elapsed_seconds": 196.06,
+  "paper2markdown_v3": {
+    "paper_id": "c39d243f-0c59-4f70-8ee6-d9e742174491",
+    "pipeline": "Paper2Markdown-V3",
+    "ok": true,
+    "page_count": 25,
+    "chunk_count": 183,
+    "main_body_chunk_count": 81,
+    "appendix_chunk_count": 91,
+    "reference_chunk_count": 11,
+    "model_text_chars": 41640,
+    "raw_markdown_chars": 104035,
+    "sanitized_chars": 93277,
+    "page_provenance": {
+      "min_page": 1,
+      "max_page": 25,
+      "invalid_count": 0
+    },
+    "marker_block_type_counts": {
+      "Caption": 9,
+      "Code": 1,
+      "Equation": 10,
+      "Figure": 6,
+      "FigureGroup": 6,
+      "Footnote": 5,
+      "ListGroup": 8,
+      "PageFooter": 25,
+      "PageHeader": 27,
+      "SectionHeader": 40,
+      "Table": 3,
+      "TableGroup": 5,
+      "Text": 488
+    },
+    "asset_count_raw": 12,
+    "asset_count_model_kept": 10,
+    "asset_count_rejected": 2,
+    "asset_reject_reasons": {
+      "kept": 10,
+      "too_small": 2
+    },
+    "artifact_leak_audit": {
+      "ok": true,
+      "hits": {
+        "Anonymous Authors": [],
+        "ACKNOWLEDGMENT": [],
+        "OpenReview": [],
+        "\"accept_label\"": [],
+        "\"decision\"": [],
+        "\"decision_tier\"": [],
+        "\"source_status\"": [],
+        "Meta-review": [],
+        "Official Review": [],
+        "official_reviews": [],
+        "meta_reviews": [],
+        "suggested_verdict_score": []
+      },
+      "artifact_count": 2
+    },
+    "default_model_input": "model_text_v3.txt",
+    "appendix_input": "appendix_text_v3.txt",
+    "reference_input": "reference_text_v3.txt"
+  }
+}

icml26/c39d243f-0c59-4f70-8ee6-d9e742174491/reference_chunks.jsonl

@@ -0,0 +1,11 @@
+{"paper_id": "c39d243f-0c59-4f70-8ee6-d9e742174491", "chunk_id": "c39d243f-0c59-4f70-8ee6-d9e742174491:0081", "section": "References", "page_start": 8, "page_end": 8, "type": "Text", "text": "Ahn, M., Brohan, A., Brown, N., Chebotar, Y., Cortes, O., David, B., Finn, C., Fu, C., Gopalakrishnan, K., Hausman, K., et al. Do as i can, not as i say: Grounding language in robotic affordances. arXiv preprint arXiv:2204.01691 , 2022.", "source": "marker_v2", "marker_block_id": "/page/7/Text/11"}
+{"paper_id": "c39d243f-0c59-4f70-8ee6-d9e742174491", "chunk_id": "c39d243f-0c59-4f70-8ee6-d9e742174491:0082", "section": "References", "page_start": 8, "page_end": 8, "type": "Text", "text": "Bou, A., Bettini, M., Dittert, S., Kumar, V., Sodhani, S.,", "source": "marker_v2", "marker_block_id": "/page/7/Text/12"}
+{"paper_id": "c39d243f-0c59-4f70-8ee6-d9e742174491", "chunk_id": "c39d243f-0c59-4f70-8ee6-d9e742174491:0083", "section": "References", "page_start": 8, "page_end": 8, "type": "Footnote", "text": "4 Our implementation did not employ aggressive optimization techniques (e.g., quantization or dedicated serving stacks), suggesting that this overhead could be further minimized.", "source": "marker_v2", "marker_block_id": "/page/7/Footnote/13"}
+{"paper_id": "c39d243f-0c59-4f70-8ee6-d9e742174491", "chunk_id": "c39d243f-0c59-4f70-8ee6-d9e742174491:0084", "section": "References", "page_start": 9, "page_end": 9, "type": "Text", "text": "440 441 442 Yang, X., De Fabritiis, G., and Moens, V. Torchrl: A data-driven decision-making library for pytorch. arXiv preprint arXiv:2306.00577 , 2023.", "source": "marker_v2", "marker_block_id": "/page/8/Text/539"}
+{"paper_id": "c39d243f-0c59-4f70-8ee6-d9e742174491", "chunk_id": "c39d243f-0c59-4f70-8ee6-d9e742174491:0085", "section": "References", "page_start": 9, "page_end": 9, "type": "ListGroup", "text": "Cao, Y., Zhao, H., Cheng, Y., Shu, T., Chen, Y., Liu, G., Liang, G., Zhao, J., Yan, J., and Li, Y. Survey on large language model-enhanced reinforcement learning: Concept, taxonomy, and methods. IEEE Transactions on Neural Networks and Learning Systems , 2024. Carrasco-Davis, R., Lee, S., Clopath, C., and Dabney, W. Uncertainty prioritized experience replay. arXiv preprint arXiv:2506.09270 , 2025. Cherti, M., Beaumont, R., Wightman, R., Wortsman, M., Ilharco, G., Gordon, C., Schuhmann, C., Schmidt, L., and Jitsev, J. Reproducible scaling laws for contrastive language-image learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pp. 2818–2829, 2023. Chevalier-Boisvert, M., Dai, B., Towers, M., Perez-Vicente, R., Willems, L., Lahlou, S., Pal, S., Castro, P. S., and Terry, J. Minigrid & miniworld: Modular & customizable reinforcement learning environments for goal-oriented tasks. Advances in Neural Information Processing Sys tems , 36:73383–73394, 2023. Cho, J. H., Madotto, A., Mavroudi, E., Afouras, T., Nagarajan, T., Maaz, M., Song, Y., Ma, T., Hu, S., Rasheed, H., Sun, P., Huang, P.-Y., Bolya, D., Jain, S., Martin, M., Wang, H., Ravi, N., Jain, S., Stark, T., Moon, S., Damavandi, B., Lee, V., Westbury, A., Khan, S., Krahenb ¨ uhl, ¨ P., Dollar, P., Torresani, L., Grauman, K., and Feichten- ´ hofer, C. Perceptionlm: Open-access data and models for detailed visual understanding. arXiv:2504.13180 , 2025. Dabney, W., Ostrovski, G., Silver, D., and Munos, R. Implicit quantile networks for distributional reinforcement learning. In International conference on machine learn ing , pp. 1096–1105. PMLR, 2018. Dubey, R., Agrawal, P., Pathak, D., Griffiths, T. L., and Efros, A. A. Investigating human priors for playing video games. arXiv preprint arXiv:1802.10217 , 2018. Fu, Y., Zhang, H., Wu, D., Xu, W., and Boulet, B. Furl: Visual-language models as fuzzy rewards for reinforcement learning. arXiv preprint arXiv:2406.00645 , 2024. Fujimoto, S., Hoof, H., and Meger, D. Addressing function approximation error in actor-critic methods. In Interna tional conference on machine learning , pp. 1587–1596. PMLR, 2018. Ghosh, U., Raychaudhuri, D. S., Li, J., Karydis, K., and Roy-Chowdhury, A. Preference vlm: Leveraging vlms for scalable preference-based reinforcement learning. arXiv preprint arXiv:2502.01616 , 2025.", "source": "marker_v2", "marker_block_id": "/page/8/ListGroup/537"}
+{"paper_id": "c39d243f-0c59-4f70-8ee6-d9e742174491", "chunk_id": "c39d243f-0c59-4f70-8ee6-d9e742174491:0086", "section": "References", "page_start": 9, "page_end": 9, "type": "ListGroup", "text": "Goel, S., Bansal, H., Bhatia, S., Rossi, R., Vinay, V., and Grover, A. Cyclip: Cyclic contrastive language-image pretraining. Advances in Neural Information Processing Systems , 35:6704–6719, 2022. Gupta, T., Karkus, P., Che, T., Xu, D., and Pavone, M. Foundation models for semantic novelty in reinforcement learning. arXiv preprint arXiv:2211.04878 , 2022. Haarnoja, T., Zhou, A., Hartikainen, K., Tucker, G., Ha, S., Tan, J., Kumar, V., Zhu, H., Gupta, A., Abbeel, P., et al. Soft actor-critic algorithms and applications. arXiv preprint arXiv:1812.05905 , 2018. Hong, Z.-W., Chen, T., Lin, Y.-C., Pajarinen, J., and Agrawal, P. Topological experience replay. arXiv preprint arXiv:2203.15845 , 2022. Huang, W., Abbeel, P., Pathak, D., and Mordatch, I. Language models as zero-shot planners: Extracting actionable knowledge for embodied agents. In International conference on machine learning , pp. 9118–9147. PMLR, 2022. Jia, C., Yang, Y., Xia, Y., Chen, Y.-T., Parekh, Z., Pham, H., Le, Q., Sung, Y.-H., Li, Z., and Duerig, T. Scaling up visual and vision-language representation learning with noisy text supervision. In International conference on machine learning , pp. 4904–4916. PMLR, 2021. Jiang, Y., Gu, S. S., Murphy, K. P., and Finn, C. Language as an abstraction for hierarchical deep reinforcement learning. Advances in neural information processing systems , 32, 2019. Kalashnikov, D., Varley, J., Chebotar, Y., Swanson, B., Jonschkowski, R., Finn, C., Levine, S., and Hausman, K. Scaling up multi-task robotic reinforcement learning. In 5th Annual Conference on Robot Learning , 2021. Kroemer, O., Niekum, S., and Konidaris, G. A review of robot learning for manipulation: Challenges, representations, and algorithms. Journal of machine learning research , 22(30):1–82, 2021. Kumar, A., Gupta, A., and Levine, S. Discor: Corrective feedback in reinforcement learning via distribution correction. Advances in neural information processing systems , 33:18560–18572, 2020. Lahire, T., Geist, M., and Rachelson, E. Large batch experience replay. arXiv preprint arXiv:2110.01528 , 2021. Levine, S., Finn, C., Darrell, T., and Abbeel, P. End-to-end training of deep visuomotor policies. Journal of Machine Learning Research , 17(39):1–40, 2016.", "source": "marker_v2", "marker_block_id": "/page/8/ListGroup/538"}
+{"paper_id": "c39d243f-0c59-4f70-8ee6-d9e742174491", "chunk_id": "c39d243f-0c59-4f70-8ee6-d9e742174491:0087", "section": "References", "page_start": 10, "page_end": 10, "type": "Text", "text": "Li, C., Hong, Z.-W., Agrawal, P., Garg, D., and Pajarinen, J. Roer: Regularized optimal experience replay. arXiv preprint arXiv:2407.03995 , 2024.", "source": "marker_v2", "marker_block_id": "/page/9/Text/504"}
+{"paper_id": "c39d243f-0c59-4f70-8ee6-d9e742174491", "chunk_id": "c39d243f-0c59-4f70-8ee6-d9e742174491:0088", "section": "References", "page_start": 10, "page_end": 10, "type": "ListGroup", "text": "Li, J., Li, D., Savarese, S., and Hoi, S. Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models. In International conference on machine learning , pp. 19730–19742. PMLR, 2023. Liang, J., Huang, W., Xia, F., Xu, P., Hausman, K., Ichter, B., Florence, P., and Zeng, A. Code as policies: Language model programs for embodied control. arXiv preprint arXiv:2209.07753 , 2022. Liang, Z., Xu, Y., Hong, Y., Shang, P., Wang, Q., Fu, Q., and Liu, K. A survey of multimodel large language models. In Proceedings of the 3rd International Conference on Com puter, Artificial Intelligence and Control Engineering , pp. 405–409, 2024. Lin, L.-J. Self-improving reactive agents based on reinforcement learning, planning and teaching. Machine learning , 8(3):293–321, 1992. Liu, X.-H., Xue, Z., Pang, J., Jiang, S., Xu, F., and Yu, Y. Regret minimization experience replay in off-policy reinforcement learning. Advances in neural information processing systems , 34:17604–17615, 2021. Luu, T. M. and Yoo, C. D. Hindsight goal ranking on replay buffer for sparse reward environment. IEEE Access , 9: 51996–52007, 2021. Luu, T. M., Lee, Y., Lee, D., Kim, S., Kim, M. J., and Yoo, C. D. Enhancing rating-based reinforcement learning to effectively leverage feedback from large vision-language models. arXiv preprint arXiv:2506.12822 , 2025. Ma, R., Luijkx, J., Ajanovic, Z., and Kober, J. Explorllm: ´ Guiding exploration in reinforcement learning with large language models. In 2025 IEEE International Confer ence on Robotics and Automation (ICRA) , pp. 9011–9017. IEEE, 2025. Mnih, V., Kavukcuoglu, K., Silver, D., Graves, A., Antonoglou, I., Wierstra, D., and Riedmiller, M. Playing atari with deep reinforcement learning. arXiv preprint arXiv:1312.5602 , 2013. Novati, G. and Koumoutsakos, P. Remember and forget for experience replay. In International Conference on Machine Learning , pp. 4851–4860. PMLR, 2019. Ouyang, L., Wu, J., Jiang, X., Almeida, D., Wainwright, C., Mishkin, P., Zhang, C., Agarwal, S., Slama, K., Ray, A., et al. Training language models to follow instructions with human feedback. Advances in neural information processing systems , 35:27730–27744, 2022.", "source": "marker_v2", "marker_block_id": "/page/9/ListGroup/502"}
+{"paper_id": "c39d243f-0c59-4f70-8ee6-d9e742174491", "chunk_id": "c39d243f-0c59-4f70-8ee6-d9e742174491:0089", "section": "References", "page_start": 10, "page_end": 10, "type": "ListGroup", "text": "Park, S., Frans, K., Eysenbach, B., and Levine, S. Ogbench: Benchmarking offline goal-conditioned rl. arXiv preprint arXiv:2410.20092 , 2024. Radford, A., Kim, J. W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al. Learning transferable visual models from natural language supervision. In International conference on machine learning , pp. 8748–8763. PmLR, 2021. Rocamonde, J., Montesinos, V., Nava, E., Perez, E., and Lindner, D. Vision-language models are zero-shot reward models for reinforcement learning. arXiv preprint arXiv:2310.12921 , 2023. Schaul, T., Horgan, D., Gregor, K., and Silver, D. Universal value function approximators. In International conference on machine learning , pp. 1312–1320. PMLR, 2015a. Schaul, T., Quan, J., Antonoglou, I., and Silver, D. Prioritized experience replay. arXiv preprint arXiv:1511.05952 , 2015b. Schoepp, S., Jafaripour, M., Cao, Y., Yang, T., Abdollahi, F., Golestan, S., Sufiyan, Z., Zaiane, O. R., and Taylor, M. E. The evolving landscape of llm-and vlm-integrated reinforcement learning. arXiv preprint arXiv:2502.15214 , 2025. Singh, A., Bhaskar, A., Yu, P., Chakraborty, S., Dasyam, R., Bedi, A., and Tokekar, P. Varp: Reinforcement learning from vision-language model feedback with agent regularized preferences. arXiv preprint arXiv:2503.13817 , 2025. Sinha, S., Song, J., Garg, A., and Ermon, S. Experience replay with likelihood-free importance weights. In Learn ing for Dynamics and Control Conference , pp. 110–123. PMLR, 2022. Sujit, S., Nath, S., Braga, P., and Ebrahimi Kahou, S. Prioritizing samples in reinforcement learning with reducible loss. Advances in Neural Information Processing Systems , 36:23237–23258, 2023. Sun, P., Zhou, W., and Li, H. Attentive experience replay. In Proceedings of the AAAI conference on artificial intel ligence , volume 34, pp. 5900–5907, 2020. Sutton, R. S., Barto, A. G., et al. Reinforcement learning: An introduction , volume 1. MIT press Cambridge, 1998. Tang, C., Abbatematteo, B., Hu, J., Chandra, R., Mart´ın-Mart´ın, R., and Stone, P. Deep reinforcement learning for robotics: A survey of real-world successes. In Proceed ings of the AAAI Conference on Artificial Intelligence , volume 39, pp. 28694–28698, 2025.", "source": "marker_v2", "marker_block_id": "/page/9/ListGroup/503"}
+{"paper_id": "c39d243f-0c59-4f70-8ee6-d9e742174491", "chunk_id": "c39d243f-0c59-4f70-8ee6-d9e742174491:0090", "section": "References", "page_start": 11, "page_end": 11, "type": "Text", "text": "550 Van Hasselt, H., Guez, A., and Silver, D. Deep reinforcement learning with double q-learning. In Proceedings of the AAAI conference on artificial intelligence , volume 30, 2016.", "source": "marker_v2", "marker_block_id": "/page/10/Text/276"}
+{"paper_id": "c39d243f-0c59-4f70-8ee6-d9e742174491", "chunk_id": "c39d243f-0c59-4f70-8ee6-d9e742174491:0091", "section": "References", "page_start": 11, "page_end": 11, "type": "ListGroup", "text": "Xie, T., Zhao, S., Wu, C. H., Liu, Y., Luo, Q., Zhong, V., Yang, Y., and Yu, T. Text2reward: Reward shaping with language models for reinforcement learning. arXiv preprint arXiv:2309.11489 , 2023. Zha, D., Lai, K.-H., Zhou, K., and Hu, X. Experience replay optimization. arXiv preprint arXiv:1906.08387 , 2019. Zhai, X., Wang, X., Mustafa, B., Steiner, A., Keysers, D., Kolesnikov, A., and Beyer, L. Lit: Zero-shot transfer with locked-image text tuning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pp. 18123–18133, 2022. Zhai, X., Mustafa, B., Kolesnikov, A., and Beyer, L. Sigmoid loss for language image pre-training. In Proceed ings of the IEEE/CVF international conference on com puter vision , pp. 11975–11986, 2023. Zhang, J., Huang, J., Jin, S., and Lu, S. Vision-language models for vision tasks: A survey. IEEE transactions on pattern analysis and machine intelligence , 46(8):5625– 5644, 2024. Zhao, R. and Tresp, V. Curiosity-driven experience prioritization via density estimation. arXiv preprint arXiv:1902.08039 , 2019. Zhu, H., Yu, J., Gupta, A., Shah, D., Hartikainen, K., Singh, A., Kumar, V., and Levine, S. The ingredients of realworld robotic reinforcement learning. arXiv preprint arXiv:2004.12570 , 2020. Zitkovich, B., Yu, T., Xu, S., Xu, P., Xiao, T., Xia, F., Wu, J., Wohlhart, P., Welker, S., Wahid, A., et al. Rt-2: Vision-language-action models transfer web knowledge to robotic control. In Conference on Robot Learning , pp. 2165–2183. PMLR, 2023.", "source": "marker_v2", "marker_block_id": "/page/10/ListGroup/275"}

icml26/c39d243f-0c59-4f70-8ee6-d9e742174491/reference_text_v3.txt

@@ -0,0 +1,32 @@
+[p. 8 | section: References | type: Text]
+Ahn, M., Brohan, A., Brown, N., Chebotar, Y., Cortes, O., David, B., Finn, C., Fu, C., Gopalakrishnan, K., Hausman, K., et al. Do as i can, not as i say: Grounding language in robotic affordances. arXiv preprint arXiv:2204.01691 , 2022.
+
+[p. 8 | section: References | type: Text]
+Bou, A., Bettini, M., Dittert, S., Kumar, V., Sodhani, S.,
+
+[p. 8 | section: References | type: Footnote]
+4 Our implementation did not employ aggressive optimization techniques (e.g., quantization or dedicated serving stacks), suggesting that this overhead could be further minimized.
+
+[p. 9 | section: References | type: Text]
+440 441 442 Yang, X., De Fabritiis, G., and Moens, V. Torchrl: A data-driven decision-making library for pytorch. arXiv preprint arXiv:2306.00577 , 2023.
+
+[p. 9 | section: References | type: ListGroup]
+Cao, Y., Zhao, H., Cheng, Y., Shu, T., Chen, Y., Liu, G., Liang, G., Zhao, J., Yan, J., and Li, Y. Survey on large language model-enhanced reinforcement learning: Concept, taxonomy, and methods. IEEE Transactions on Neural Networks and Learning Systems , 2024. Carrasco-Davis, R., Lee, S., Clopath, C., and Dabney, W. Uncertainty prioritized experience replay. arXiv preprint arXiv:2506.09270 , 2025. Cherti, M., Beaumont, R., Wightman, R., Wortsman, M., Ilharco, G., Gordon, C., Schuhmann, C., Schmidt, L., and Jitsev, J. Reproducible scaling laws for contrastive language-image learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pp. 2818–2829, 2023. Chevalier-Boisvert, M., Dai, B., Towers, M., Perez-Vicente, R., Willems, L., Lahlou, S., Pal, S., Castro, P. S., and Terry, J. Minigrid & miniworld: Modular & customizable reinforcement learning environments for goal-oriented tasks. Advances in Neural Information Processing Sys tems , 36:73383–73394, 2023. Cho, J. H., Madotto, A., Mavroudi, E., Afouras, T., Nagarajan, T., Maaz, M., Song, Y., Ma, T., Hu, S., Rasheed, H., Sun, P., Huang, P.-Y., Bolya, D., Jain, S., Martin, M., Wang, H., Ravi, N., Jain, S., Stark, T., Moon, S., Damavandi, B., Lee, V., Westbury, A., Khan, S., Krahenb ¨ uhl, ¨ P., Dollar, P., Torresani, L., Grauman, K., and Feichten- ´ hofer, C. Perceptionlm: Open-access data and models for detailed visual understanding. arXiv:2504.13180 , 2025. Dabney, W., Ostrovski, G., Silver, D., and Munos, R. Implicit quantile networks for distributional reinforcement learning. In International conference on machine learn ing , pp. 1096–1105. PMLR, 2018. Dubey, R., Agrawal, P., Pathak, D., Griffiths, T. L., and Efros, A. A. Investigating human priors for playing video games. arXiv preprint arXiv:1802.10217 , 2018. Fu, Y., Zhang, H., Wu, D., Xu, W., and Boulet, B. Furl: Visual-language models as fuzzy rewards for reinforcement learning. arXiv preprint arXiv:2406.00645 , 2024. Fujimoto, S., Hoof, H., and Meger, D. Addressing function approximation error in actor-critic methods. In Interna tional conference on machine learning , pp. 1587–1596. PMLR, 2018. Ghosh, U., Raychaudhuri, D. S., Li, J., Karydis, K., and Roy-Chowdhury, A. Preference vlm: Leveraging vlms for scalable preference-based reinforcement learning. arXiv preprint arXiv:2502.01616 , 2025.
+
+[p. 9 | section: References | type: ListGroup]
+Goel, S., Bansal, H., Bhatia, S., Rossi, R., Vinay, V., and Grover, A. Cyclip: Cyclic contrastive language-image pretraining. Advances in Neural Information Processing Systems , 35:6704–6719, 2022. Gupta, T., Karkus, P., Che, T., Xu, D., and Pavone, M. Foundation models for semantic novelty in reinforcement learning. arXiv preprint arXiv:2211.04878 , 2022. Haarnoja, T., Zhou, A., Hartikainen, K., Tucker, G., Ha, S., Tan, J., Kumar, V., Zhu, H., Gupta, A., Abbeel, P., et al. Soft actor-critic algorithms and applications. arXiv preprint arXiv:1812.05905 , 2018. Hong, Z.-W., Chen, T., Lin, Y.-C., Pajarinen, J., and Agrawal, P. Topological experience replay. arXiv preprint arXiv:2203.15845 , 2022. Huang, W., Abbeel, P., Pathak, D., and Mordatch, I. Language models as zero-shot planners: Extracting actionable knowledge for embodied agents. In International conference on machine learning , pp. 9118–9147. PMLR, 2022. Jia, C., Yang, Y., Xia, Y., Chen, Y.-T., Parekh, Z., Pham, H., Le, Q., Sung, Y.-H., Li, Z., and Duerig, T. Scaling up visual and vision-language representation learning with noisy text supervision. In International conference on machine learning , pp. 4904–4916. PMLR, 2021. Jiang, Y., Gu, S. S., Murphy, K. P., and Finn, C. Language as an abstraction for hierarchical deep reinforcement learning. Advances in neural information processing systems , 32, 2019. Kalashnikov, D., Varley, J., Chebotar, Y., Swanson, B., Jonschkowski, R., Finn, C., Levine, S., and Hausman, K. Scaling up multi-task robotic reinforcement learning. In 5th Annual Conference on Robot Learning , 2021. Kroemer, O., Niekum, S., and Konidaris, G. A review of robot learning for manipulation: Challenges, representations, and algorithms. Journal of machine learning research , 22(30):1–82, 2021. Kumar, A., Gupta, A., and Levine, S. Discor: Corrective feedback in reinforcement learning via distribution correction. Advances in neural information processing systems , 33:18560–18572, 2020. Lahire, T., Geist, M., and Rachelson, E. Large batch experience replay. arXiv preprint arXiv:2110.01528 , 2021. Levine, S., Finn, C., Darrell, T., and Abbeel, P. End-to-end training of deep visuomotor policies. Journal of Machine Learning Research , 17(39):1–40, 2016.
+
+[p. 10 | section: References | type: Text]
+Li, C., Hong, Z.-W., Agrawal, P., Garg, D., and Pajarinen, J. Roer: Regularized optimal experience replay. arXiv preprint arXiv:2407.03995 , 2024.
+
+[p. 10 | section: References | type: ListGroup]
+Li, J., Li, D., Savarese, S., and Hoi, S. Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models. In International conference on machine learning , pp. 19730–19742. PMLR, 2023. Liang, J., Huang, W., Xia, F., Xu, P., Hausman, K., Ichter, B., Florence, P., and Zeng, A. Code as policies: Language model programs for embodied control. arXiv preprint arXiv:2209.07753 , 2022. Liang, Z., Xu, Y., Hong, Y., Shang, P., Wang, Q., Fu, Q., and Liu, K. A survey of multimodel large language models. In Proceedings of the 3rd International Conference on Com puter, Artificial Intelligence and Control Engineering , pp. 405–409, 2024. Lin, L.-J. Self-improving reactive agents based on reinforcement learning, planning and teaching. Machine learning , 8(3):293–321, 1992. Liu, X.-H., Xue, Z., Pang, J., Jiang, S., Xu, F., and Yu, Y. Regret minimization experience replay in off-policy reinforcement learning. Advances in neural information processing systems , 34:17604–17615, 2021. Luu, T. M. and Yoo, C. D. Hindsight goal ranking on replay buffer for sparse reward environment. IEEE Access , 9: 51996–52007, 2021. Luu, T. M., Lee, Y., Lee, D., Kim, S., Kim, M. J., and Yoo, C. D. Enhancing rating-based reinforcement learning to effectively leverage feedback from large vision-language models. arXiv preprint arXiv:2506.12822 , 2025. Ma, R., Luijkx, J., Ajanovic, Z., and Kober, J. Explorllm: ´ Guiding exploration in reinforcement learning with large language models. In 2025 IEEE International Confer ence on Robotics and Automation (ICRA) , pp. 9011–9017. IEEE, 2025. Mnih, V., Kavukcuoglu, K., Silver, D., Graves, A., Antonoglou, I., Wierstra, D., and Riedmiller, M. Playing atari with deep reinforcement learning. arXiv preprint arXiv:1312.5602 , 2013. Novati, G. and Koumoutsakos, P. Remember and forget for experience replay. In International Conference on Machine Learning , pp. 4851–4860. PMLR, 2019. Ouyang, L., Wu, J., Jiang, X., Almeida, D., Wainwright, C., Mishkin, P., Zhang, C., Agarwal, S., Slama, K., Ray, A., et al. Training language models to follow instructions with human feedback. Advances in neural information processing systems , 35:27730–27744, 2022.
+
+[p. 10 | section: References | type: ListGroup]
+Park, S., Frans, K., Eysenbach, B., and Levine, S. Ogbench: Benchmarking offline goal-conditioned rl. arXiv preprint arXiv:2410.20092 , 2024. Radford, A., Kim, J. W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al. Learning transferable visual models from natural language supervision. In International conference on machine learning , pp. 8748–8763. PmLR, 2021. Rocamonde, J., Montesinos, V., Nava, E., Perez, E., and Lindner, D. Vision-language models are zero-shot reward models for reinforcement learning. arXiv preprint arXiv:2310.12921 , 2023. Schaul, T., Horgan, D., Gregor, K., and Silver, D. Universal value function approximators. In International conference on machine learning , pp. 1312–1320. PMLR, 2015a. Schaul, T., Quan, J., Antonoglou, I., and Silver, D. Prioritized experience replay. arXiv preprint arXiv:1511.05952 , 2015b. Schoepp, S., Jafaripour, M., Cao, Y., Yang, T., Abdollahi, F., Golestan, S., Sufiyan, Z., Zaiane, O. R., and Taylor, M. E. The evolving landscape of llm-and vlm-integrated reinforcement learning. arXiv preprint arXiv:2502.15214 , 2025. Singh, A., Bhaskar, A., Yu, P., Chakraborty, S., Dasyam, R., Bedi, A., and Tokekar, P. Varp: Reinforcement learning from vision-language model feedback with agent regularized preferences. arXiv preprint arXiv:2503.13817 , 2025. Sinha, S., Song, J., Garg, A., and Ermon, S. Experience replay with likelihood-free importance weights. In Learn ing for Dynamics and Control Conference , pp. 110–123. PMLR, 2022. Sujit, S., Nath, S., Braga, P., and Ebrahimi Kahou, S. Prioritizing samples in reinforcement learning with reducible loss. Advances in Neural Information Processing Systems , 36:23237–23258, 2023. Sun, P., Zhou, W., and Li, H. Attentive experience replay. In Proceedings of the AAAI conference on artificial intel ligence , volume 34, pp. 5900–5907, 2020. Sutton, R. S., Barto, A. G., et al. Reinforcement learning: An introduction , volume 1. MIT press Cambridge, 1998. Tang, C., Abbatematteo, B., Hu, J., Chandra, R., Mart´ın-Mart´ın, R., and Stone, P. Deep reinforcement learning for robotics: A survey of real-world successes. In Proceed ings of the AAAI Conference on Artificial Intelligence , volume 39, pp. 28694–28698, 2025.
+
+[p. 11 | section: References | type: Text]
+550 Van Hasselt, H., Guez, A., and Silver, D. Deep reinforcement learning with double q-learning. In Proceedings of the AAAI conference on artificial intelligence , volume 30, 2016.
+
+[p. 11 | section: References | type: ListGroup]
+Xie, T., Zhao, S., Wu, C. H., Liu, Y., Luo, Q., Zhong, V., Yang, Y., and Yu, T. Text2reward: Reward shaping with language models for reinforcement learning. arXiv preprint arXiv:2309.11489 , 2023. Zha, D., Lai, K.-H., Zhou, K., and Hu, X. Experience replay optimization. arXiv preprint arXiv:1906.08387 , 2019. Zhai, X., Wang, X., Mustafa, B., Steiner, A., Keysers, D., Kolesnikov, A., and Beyer, L. Lit: Zero-shot transfer with locked-image text tuning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pp. 18123–18133, 2022. Zhai, X., Mustafa, B., Kolesnikov, A., and Beyer, L. Sigmoid loss for language image pre-training. In Proceed ings of the IEEE/CVF international conference on com puter vision , pp. 11975–11986, 2023. Zhang, J., Huang, J., Jin, S., and Lu, S. Vision-language models for vision tasks: A survey. IEEE transactions on pattern analysis and machine intelligence , 46(8):5625– 5644, 2024. Zhao, R. and Tresp, V. Curiosity-driven experience prioritization via density estimation. arXiv preprint arXiv:1902.08039 , 2019. Zhu, H., Yu, J., Gupta, A., Shah, D., Hartikainen, K., Singh, A., Kumar, V., and Levine, S. The ingredients of realworld robotic reinforcement learning. arXiv preprint arXiv:2004.12570 , 2020. Zitkovich, B., Yu, T., Xu, S., Xu, P., Xiao, T., Xia, F., Wu, J., Wohlhart, P., Welker, S., Wahid, A., et al. Rt-2: Vision-language-action models transfer web knowledge to robotic control. In Conference on Robot Learning , pp. 2165–2183. PMLR, 2023.

icml26/c39d243f-0c59-4f70-8ee6-d9e742174491/sanitization_report.json

@@ -0,0 +1,61 @@
+{
+  "paper_id": "c39d243f-0c59-4f70-8ee6-d9e742174491",
+  "pipeline": "Paper2Markdown-V3",
+  "ok": true,
+  "page_count": 25,
+  "chunk_count": 183,
+  "main_body_chunk_count": 81,
+  "appendix_chunk_count": 91,
+  "reference_chunk_count": 11,
+  "model_text_chars": 41640,
+  "raw_markdown_chars": 104035,
+  "sanitized_chars": 93277,
+  "page_provenance": {
+    "min_page": 1,
+    "max_page": 25,
+    "invalid_count": 0
+  },
+  "marker_block_type_counts": {
+    "Caption": 9,
+    "Code": 1,
+    "Equation": 10,
+    "Figure": 6,
+    "FigureGroup": 6,
+    "Footnote": 5,
+    "ListGroup": 8,
+    "PageFooter": 25,
+    "PageHeader": 27,
+    "SectionHeader": 40,
+    "Table": 3,
+    "TableGroup": 5,
+    "Text": 488
+  },
+  "asset_count_raw": 12,
+  "asset_count_model_kept": 10,
+  "asset_count_rejected": 2,
+  "asset_reject_reasons": {
+    "kept": 10,
+    "too_small": 2
+  },
+  "artifact_leak_audit": {
+    "ok": true,
+    "hits": {
+      "Anonymous Authors": [],
+      "ACKNOWLEDGMENT": [],
+      "OpenReview": [],
+      "\"accept_label\"": [],
+      "\"decision\"": [],
+      "\"decision_tier\"": [],
+      "\"source_status\"": [],
+      "Meta-review": [],
+      "Official Review": [],
+      "official_reviews": [],
+      "meta_reviews": [],
+      "suggested_verdict_score": []
+    },
+    "artifact_count": 2
+  },
+  "default_model_input": "model_text_v3.txt",
+  "appendix_input": "appendix_text_v3.txt",
+  "reference_input": "reference_text_v3.txt"
+}

icml26/c39d243f-0c59-4f70-8ee6-d9e742174491/sanitized_v3.txt

@@ -0,0 +1,494 @@
+{0}
+# Abstract
+Recent advances in Large Language Models (LLMs) and Vision-Language Models (VLMs) have enabled powerful semantic and multimodal reasoning capabilities, creating new opportunities to enhance sample efficiency, high-level planning, and interpretability in reinforcement learning (RL). While prior work has integrated LLMs and VLMs into various components of RL, the replay buffer, a core component for storing and reusing experiences, remains unexplored. We propose addressing this gap by leveraging VLMs to guide the prioritization of experiences in the replay buffer. Our key idea is to use a frozen, pre-trained VLM (requiring no fine-tuning) as an automated evaluator to identify and prioritize promising sub-trajectories from the agent's experiences. Across scenarios, including game-playing and robotics, spanning both discrete and continuous domains, agents trained with our proposed prioritization method achieve 11–52% higher average success rates and improve sample efficiency by 19–45% compared to previous approaches.
+# 1. Introduction
+Reinforcement learning (RL) has achieved impressive results across domains ranging from robotics to natural language systems [\(Levine et al.,](#page-8-0) [2016;](#page-8-0) [Ouyang et al.,](#page-9-0) [2022\)](#page-9-0) and logistics management [\(Tang et al.,](#page-9-1) [2025\)](#page-9-1). Despite this success, RL remains notoriously sample-inefficient, sensitive to hyperparameters, and heavily dependent on carefully designed reward functions [\(Kalashnikov et al.,](#page-8-1) [2021;](#page-8-1) [Kroemer](#page-8-2) [et al.,](#page-8-2) [2021;](#page-8-2) [Zhu et al.,](#page-10-0) [2020\)](#page-10-0).
+A well-established line of research—off-policy RL [\(Lin,](#page-9-2) [1992;](#page-9-2) [Mnih et al.,](#page-9-3) [2013\)](#page-9-3)—addresses the efficiency bottleneck through experience reuse. During training, transitions
+are collected into a dataset known as a *replay buffer*, and sampled multiple times to update the policy. Ideally, a sampling mechanism would prioritize the most meaningful transitions: those that provide the richest learning signal, irrespective of when they were collected. The core challenge is to identify which experiences are truly meaningful for learning process.
+Conventional methods such as Prioritized Experience Replay (PER, [Schaul et al.](#page-9-4) [2015b\)](#page-9-4) approximate this ideal by prioritizing transitions with high temporal-difference (TD) error. The TD error measures the discrepancy between a bootstrapped value estimate and its target [\(Sutton et al.,](#page-9-5) [1998\)](#page-9-5). In effect, this approach promotes transitions which are misaligned with the agent's future outcome predictions. While TD error prioritization can correct value estimates, it is fundamentally limited by its lack of semantic awareness. It cannot distinguish transitions which reflect genuine progress toward task completion from those that do not.
+This limitation is particularly evident in long-horizon, sparse-reward tasks, such as robotic manipulation in OG-Bench [\(Park et al.,](#page-9-6) [2024\)](#page-9-6). For instance, critical transitions—such as unlatching a door or grasping a tool—may yield low TD errors early in training. This is a consequence of delayed credit assignment preventing the reward signal from propagating to these earlier states. Conversely, visually distinct but task-irrelevant motions can yield large TD errors, despite contributing little to actual task completion.
+To bridge this gap, we propose replacing heuristic inductive biases with external sources of semantic understanding. Recent literature highlights the potential of integrating largescale Vision-Language Models (VLMs) with RL to leverage pre-trained world knowledge [\(Liang et al.,](#page-9-7) [2024;](#page-9-7) [Zhang](#page-10-1) [et al.,](#page-10-1) [2024\)](#page-10-1). Because VLMs combine visual perception with language reasoning, they can interpret complex environments and mitigate RL-specific challenges—such as sparse rewards or inefficient exploration—by providing enriched context [\(Cao et al.,](#page-8-3) [2024;](#page-8-3) [Schoepp et al.,](#page-9-8) [2025\)](#page-9-8). This motivates our primary research question:
+*Can the semantic knowledge of pre-trained VLMs be used to prioritize meaningful experiences in replay buffers?*
+To answer this, we introduce VLM-RB (Fig[.1\)](#page-1-0), which integrates a pre-trained VLM directly into the experi-
+{1}------------------------------------------------
+Figure 1. System Diagram: (1) Data is collected with the current policy $\pi_k$ interacting with multiple instances of the environment. (2) Transitions (s, a, r, s') are kept in a prioritized replay buffer with a default priority $\bar{\mathbf{p}}$ . (3) Asynchronously, after each insertion, a VLM worker scores the corresponding rendered clip $\tau^O$ under a prompt P and writes the resulting priority $\mathbf{p^{VLM}}$ back to the replay buffer (see Section 3.1). (4) The learner samples a minibatch using the mixture distribution $q_t$ which interpolates between uniform and VLM-prioritized replay (see Section 3.1). (5) The agent policy is updated to $\pi_{k+1}$ .
+ence prioritization pipeline. In contrast to heuristics relying on statistical proxies such as TD error, uncertainty (Carrasco-Davis et al., 2025), or density ratios (Zhao & Tresp, 2019), **VLM-RB** leverages the reasoning capabilities of VLMs to *score* experiences according to their semantic relevance.
+We design this framework as a modular *plug-and-play* prioritization layer that can be integrated with any off-policy algorithm that uses a replay buffer. By using frozen off-the-shelf VLMs, we leverage strong semantic priors while avoiding the computational cost of fine-tuning or additional inference latency at deployment. Within the training pipeline, VLM-RB promotes efficiency by querying the VLM asynchronously, thereby sidestepping the blocking inference bottleneck that often constrains the throughput of LLM/VLM-enhanced RL methods.
+We empirically show that pre-trained VLMs yield semantically grounded prioritization signals which are closely aligned with task objectives. Leveraging this semantic guidance, VLM-RB consistently outperforms existing baselines in both discrete game-playing and continuous robotic manipulation. Importantly, the moderate throughput overhead (about 12%) is more than offset by substantial gains in sample efficiency (19–45%) and average success rates (11–52%).
+#### 2. Preliminaries
+<span id="page-1-0"></span>
+**Off-Policy Reinforcement Learning.** We consider the standard RL framework, where an agent interacts with an environment modeled as an infinite-horizon Markov Decision Process (MDP) $\mathcal{M} = (\mathcal{S}, \mathcal{A}, P, r, \gamma, \rho_0)$ , with state space $\mathcal{S}$ , action space $\mathcal{A}$ (either discrete or continuous),
+transition kernel $P: \mathcal{S} \times \mathcal{A} \to \Delta(\mathcal{S})$ , reward function $r: \mathcal{S} \times \mathcal{A} \to \mathbb{R}$ , discount factor $\gamma \in (0,1)$ and initial-state distribution $\rho_0 \in \Delta(\mathcal{S})$ . We further assume access to a rendering function $\psi: \mathcal{S} \to \mathcal{O} \subseteq \mathbb{R}^{H \times W \times 3}$ which maps states to $H \times W$ -sized visual observations $o_t = \psi(s_t)$ . The agent learns a stochastic policy $\pi_\theta: \mathcal{S} \to \Delta(\mathcal{A})$ to maximize the expected discounted return
+$$J(\pi) = \mathbb{E}_{\tau \sim p_{\pi}(\tau)} \left[ \sum_{t=0}^{\infty} \gamma^{t} r_{t} \right],$$
+where $\tau = (s_0, a_0, r_0, s_1, a_1, r_1, ...)$ denotes a trajectory and $p_{\pi}(\tau)$ is the trajectory distribution induced by policy $\pi$ .
+To evaluate and improve policies, value functions are used to estimate expected returns. The *action-value function* (Q-function) $Q^{\pi}: \mathcal{S} \times \mathcal{A} \to \mathbb{R}$ represents the expected discounted return from state s after taking action a and then following policy $\pi$ :
+$$Q^{\pi}(s, a) = \mathbb{E}_{\tau \sim p_{\pi}(\tau \mid s_0 = s, a_0 = a)} \left[ \sum_{t=0}^{\infty} \gamma^t r_t \right].$$
+In modern RL methods, $Q^{\pi}$ is typically a learned function (represented by a neural network), denoted the *critic*.
+In off-policy RL, rather than exclusively using on-policy data, optimization is performed using previously collected experience stored in a *replay buffer* (Mnih et al., 2013). Let $\mathcal{D}_t$ denote the replay buffer at time t. We sample training tuples $(s, a, r, s') \in \mathcal{D}_t$ according to a time-dependent distribution $q_t \in \Delta(\mathcal{D}_t)$ . In standard *uniform* replay (UER), $q_t(i) = 1/|\mathcal{D}_t|$ for all $i \in \mathcal{D}_t$ . This formulation highlights that the optimization dynamics depend on the evolving sampling distribution $q_t$ .
+{2}------------------------------------------------
+<span id="page-2-3"></span>![](_page_2_Figure_1.jpeg)
+Figure 2. Temporal context resolves visual ambiguity. From the initial state $o_i$ (left), multiple futures are possible. The top sub-trajectory shows a successful grasp, while the bottom shows stagnation. By scoring sub-trajectories rather than single frames, the VLM has sufficient context to distinguish meaningful progress from failure modes.
+Experience Replay Prioritization. As opposed to uniform replay, prioritized replay methods bias retrieval toward transitions expected to yield a stronger learning signal. For clarity, we formalize this process as three distinct steps: scoring, prioritization, and sampling. Scoring assigns each transition $i \in \mathcal{D}_t$ a scalar value $\mathbf{p}_i \in \mathbb{R}$ reflecting its heuristic utility. Prioritization maps these raw scores into a probability distribution over the buffer, $q_t^{\mathbf{P}} \in \Delta(\mathcal{D}_t)$ . Sampling then draws a minibatch of indices $\{i_k\}_{k=1}^B$ from a target distribution $q_t$ , or a mixture of several such distributions.
+Prioritized Experience Replay (**PER**, Schaul et al. 2015b) is a specific instance of this framework. Its scoring function utilizes the temporal-difference (TD) error magnitude, $\mathbf{p}_i = |\delta_i| + \epsilon^1$ , where $\delta_i$ is the discrepancy between the value estimate and its target<sup>2</sup>. The prioritized distribution is defined as $q_t^{\mathbf{P}}(i) \propto \mathbf{p}_i^{\alpha}$ , where $\alpha \in [0,1]$ determines the degree of prioritization. While **PER** effectively focuses updates on transitions with larger TD errors, it relies on value estimates which may be arbitrary and noisy early in training, particularly in sparse-reward settings.
+**Vision–Language Models.** Recent advances in large-scale representation learning have given rise to *Vision–Language Models* (VLMs), which learn aligned embeddings of visual and textual modalities through joint pre-training on large-scale image–text corpora (Radford et al., 2021; Jia et al., 2021; Zhai et al., 2023). Formally, a VLM defines encoders $f_{\text{img}}: \mathcal{I} \to \mathbb{R}^d$ and $f_{\text{txt}}: \mathcal{T} \to \mathbb{R}^d$ mapping images and text to a shared latent space, typically optimized via contrastive objectives which encourage semantic correspondence between paired inputs (Cherti et al., 2023; Zhai et al., 2022; Goel et al., 2022). These models have demonstrated strong generalization across a wide
+range of downstream tasks, including zero-shot recognition, goal specification, and reward inference (Ahn et al., 2022; Rocamonde et al., 2023). While standard VLMs process static images, *Video Question Answering* (Video-QA) models extend these capabilities to ingest a temporal sequence of frames (a video clip) along with a natural-language query. By aggregating visual information across time, these models can reason about events, motion, and causal relationships. Concretely, for the rest of this paper, we employ Perception-LM (Cho et al., 2025), a state-of-the-art opensource Video-QA family of models with multiple model sizes (1B/3B/8B).
+#### 3. VLM-RB
+In this section, we present **VLM-RB** (Fig.1 and Algorithm 1), a plug-and-play framework designed for any off-policy algorithm that leverages a replay buffer. The central idea is to use a VLM to assign semantic scores to subtrajectories, subsequently biasing the sampling distribution toward higher-scoring experiences during policy optimization.
+#### <span id="page-2-0"></span>3.1. Scoring, Prioritization, and Sampling
+We now define how semantic scores are extracted from collected data, how these scores induce a prioritization distribution, and how **VLM-RB** samples from this distribution.
+**Scoring. VLM-RB** uses a pre-trained frozen VLM to extract semantic scores from collected experiences. To enable this, we first construct temporal visual sequences from the agent's history. Formally, let $\tau_i^O = (o_i, o_{i+1}, \dots, o_{i+L-1})$ denote a visual clip comprising L rendered frames. The VLM maps this clip and a text prompt P to a scalar score:
+<span id="page-2-4"></span>
+$$\mathbf{p}^{\mathbf{VLM}} = f_{\mathbf{VLM}}(\tau^O, \mathsf{P}) \in \mathbb{R}.$$
+(1)
+The prompt P directs the VLM's scoring mechanism, leveraging the model's inherent world knowledge to identify meaningful behaviors even in the absence of dense reward signals. While this interface supports the injection of detailed and task-specific priors, we find that the VLM's intrinsic scene understanding is sufficient for our purposes. We therefore employ a general task-agnostic prompt (see Appendix C for details).
+For simplicity, we set $f_{\rm VLM}$ as a binary indicator, assigning a score of 1 to clips exhibiting seemingly meaningful behavior, and 0 otherwise. This effectively labels useful sequences without the need for hand-crafted task-specific definitions. Crucially, because the VLM is frozen and its evaluation depends only on static visual content, each clip is scored exactly once. This offers a significant efficiency advantage over methods such as **PER**, which require ongoing priority updates as the value function evolves.
+<span id="page-2-2"></span><span id="page-2-1"></span><sup>&</sup>lt;sup>1</sup>Here $\epsilon > 0$ ensures non-zero probability.
+<sup>&</sup>lt;sup>2</sup>Namely, $\delta_i = Q_{\theta}(s_t, a_t) - (r + \gamma Q_{\theta}(s_{t+1}, a'))$ where $\theta$ are the critic parameters and a' is the action predicted by the current policy in $s_{t+1}$ . In Q-learning methods, a' is the argmax action, and in actor-critic methods for continuous control, it is $\pi(s_{t+1})$ .
+{3}------------------------------------------------
+A natural question is whether scoring individual frames, rather than clips, would suffice. We argue that single frames are fundamentally limited by *semantic ambiguity*. To illustrate this, consider Fig. [2:](#page-2-3) a single observation of a robotic gripper hovering above an object. This static frame is visually identical across two vastly different temporal contexts: the onset of a successful grasp or the aftermath of a failed attempt.
+218 219 Without temporal context a VLM cannot disambiguate these scenarios, potentially assigning high scores to frames which actually belong to failure modes. By scoring sub-trajectories instead, we increase the likelihood of the VLM having sufficient temporal information to distinguish meaningful behaviors from failures. While a sliding-window variant could potentially provide denser labels, we leave this (computationally intensive) alternative for future work.
+Prioritization. We construct a prioritized distribution q P by propagating the VLM score of each clip to all transitions within that clip. Under our binary scoring scheme (p<sup>i</sup> ∈ {0, 1}), q <sup>P</sup> becomes uniform over the subset of transitions labeled as semantically meaningful.
+Sampling. If we were to sample only from q <sup>P</sup>, we would discard all transitions labeled uninteresting by the VLM. To avoid wasting collected data and to ensure the agent explores the full state space, we instead use a mixture strategy q<sup>t</sup> interpolating between VLM-guided prioritization and uniform replay:
+<span id="page-3-4"></span>
+$$q_t(i) = \lambda_t q^{\mathbf{P}}(i) + (1 - \lambda_t) q^{\mathbf{U}}(i), \tag{2}$$
+where λ<sup>t</sup> ∈ [0, 1] controls the strength of the VLM guidance. In practice, each batch draws a λ<sup>t</sup> fraction from q <sup>P</sup> and the remainder uniformly. We use a linear warm-up schedule: starting with λ<sup>0</sup> = 0 (pure uniform sampling), we gradually anneal to λmax = 0.5 over the first half of training. We hypothesize that this schedule is essential: early in training, broad coverage stabilizes value learning, while later updates bias toward high-utility regions. Our ablation studies (Section [4.4\)](#page-6-0) support this design: purely prioritized sampling is detrimental, but the proposed mixture schedule yields significant efficiency gains.
+### <span id="page-3-1"></span>3.2. Efficient Implementation
+Incorporating a VLM into the RL loop introduces significant computational overhead relative to standard components. We address this with two key design choices. First, we leverage the decoupling between data collection and policy optimization in off-policy learning to asynchronously score experiences. As illustrated in Fig. [1,](#page-1-0) the VLM interacts with the replay buffer in the background, ensuring that policy optimization is never blocked by inference latency. Beyond minimizing latency, this decoupled architecture renders VLM-RB agnostic to the policy input modality: while the VLM requires rendered frames, the policy itself can operate on arbitrary observation spaces such as lowdimensional states, enabling more sample-efficient learning. This architecture further allows a single VLM instance to efficiently serve multiple parallel environments. Second, we use a lightweight 1B parameter model [\(Cho et al.,](#page-8-8) [2025\)](#page-8-8). Our ablations show that this model is sufficient to identify meaningful behaviors while maintaining high throughput[3](#page-3-0) (see Appendix [F.2\)](#page-21-0).
+### <span id="page-3-2"></span>3.3. Boosting with TD-error
+We can further refine prioritization by incorporating the TD-error, defining q <sup>P</sup>(i) ∝ p VLM i · |δ<sup>i</sup> |, where δ<sup>i</sup> emphasizes transitions with high prediction error (where the value function is inaccurate), and p VLM i emphasizes semantic relevance. Because the VLM score is a binary indicator, it effectively masks "irrelevant" transitions and promotes the remaining transitions based on their TD errors.
+In this scheme, we maintain two scores per transition in the buffer: q <sup>P</sup>(i) ∝ p VLM i · |δ<sup>i</sup> | and δ<sup>i</sup> . The TD error δ<sup>i</sup> is updated each time the transition is sampled, while q <sup>P</sup>(i) ∝ p i · |δ<sup>i</sup> | is updated only once, as described in Section [3.2.](#page-3-1)
+# 4. Experiments
+In this section, we evaluate the efficacy of VLM-RB in leveraging semantic priors for efficient exploration. We first ask: does the VLM signal provide meaningful guidance for exploration? To answer this, we analyze its correlation with learned value estimates and its dependence on visual semantics (Section [4.2\)](#page-4-0). Next, we benchmark VLM-RB against UER, PER, and alternative prioritization schemes, evaluating both performance and sample efficiency across a range of discrete and continuous control tasks (Section [4.3\)](#page-5-0). Finally, we examine which design choices most affect the performance of VLM-RB, focusing on the sampling mixture and VLM model size (Section [4.4\)](#page-6-0).
+### <span id="page-3-3"></span>4.1. Experimental Setup
+Tasks. We evaluate VLM-RB on two domains with discrete and continuous action spaces: (i) DoorKey from MiniGrid [\(Chevalier-Boisvert et al.,](#page-8-9) [2023\)](#page-8-9), using grid sizes 8x8, 12x12, and 16x16 to vary exploration difficulty; and (ii) scene from OGBench [\(Park et al.,](#page-9-6) [2024\)](#page-9-6), using the predefined tasks 3, 4, and 5, which require increasingly long-horizon compositional manipulation (unlocking/locking and coordinated object placement). In all experiments, agents receive *state-based* observations (rather
+<span id="page-3-0"></span><sup>3</sup> Scaling to larger variants (3B or 8B) yields diminishing returns in downstream RL performance (see Appendix [F.2\)](#page-21-0).
+{4}------------------------------------------------
+<span id="page-4-1"></span>![](_page_4_Figure_2.jpeg)
+Figure 3. Frozen VLM scoring anticipates learned value. We visualize a single reference episode in MiniGrid/DoorKey-16x16 which contains both goal-directed actions and random noise. Left: Visualizations of two 32-frame sub-trajectories. clip A: captures a semantically relevant sequence. clip B: shows a mostly random walk which eventually reaches the goal. Right Top: A timeline of ground-truth events; sparse reward is only received at the final goal. Right Bottom: The rose curve shows the frozen VLM score for L=32-frame clips. The gray curves show the temporal value difference, $\Delta Q(t;L)=Q(t+\frac{L}{2})-Q(t-\frac{L}{2})$ , calculated from critic checkpoints at increasing training steps (light to dark). Early in training (light gray), the critic is uninformative, while later checkpoints (dark gray) increasingly assign positive value to the same semantic events the VLM identified. This demonstrates that the VLM provides a helpful signal long before the critic successfully converges.
+than pixels) to minimize perception-related confounders and isolate the effect of replay prioritization on exploration and sample efficiency. More details about the various tasks are provided in Appendix B.
+Baselines. We compare VLM-RB to two standard replay sampling methods: Uniform Experience Replay (UER), which samples transitions uniformly from the replay buffer, and PER, which prioritizes sampling by TD-error. We report UER and PER comparisons for four different algorithms: DQN (Van Hasselt et al., 2016), IQN (Dabney et al., 2018), TD3 (Fujimoto et al., 2018), and SAC (Haarnoja et al., 2018). In addition, to compare VLM-RB against alternative replay prioritization methods, we run a focused ablation for DQN comparing to Attentive Experience Replay (AER, Sun et al. 2020), Experience Replay Optimization (ERO, Zha et al. 2019), and Reducible Loss Prioritization (ReLO, Sujit et al. 2023). Implementation details for all baselines are provided in Appendix E.
+Method Configuration. To adapt to the specific characteristics of the experiment domains, we employ two variants of our prioritization scheme. For the discrete MiniGrid tasks, where semantic progress is binary (e.g., carrying key vs. not), we rely strictly on the binary VLM semantic filter as defined in Section 3.1. Conversely, for the continuous control tasks in OGBench, where fine-grained motion control is required, we utilize the TD-error boosted variant described in Section 3.3. This allows the agent to combine the high-level semantic filtering of the VLM with information
+from the TD error.
+**Metrics.** We report the SUCCESS RATE (SR), defined as the fraction of episodes in which the agent completes the task within the maximum episode length, $T_{\rm max}$ . The AVERAGE SUCCESS RATE (ASR) is the mean SR over N=32 evaluation episodes for a given seed. Final results are the ASR averaged over M=5 seeds, with the standard error of the mean (SEM) shown as shaded regions in plots.
+### <span id="page-4-0"></span>4.2. Do VLMs Contain Human-Like Priors?
+Before turning to downstream RL results, we first ask whether the frozen VLM offers a signal that is actually useful for prioritizing data. To understand the semantic grounding of our prioritization, we examine how the frozen VLM scores correlate with the critic value estimates ( $\Delta Q$ ) as learning progresses. Concretely, we track how the value estimates of the critic evolve on a fixed reference episode, using checkpoints from a successful training run. As shown in Fig. 3, the critic's estimates are initially flat and uninformative (light gray lines). Over the course of training, the critic gradually learns to assign high value to the same semantic events (such as picking up a key or opening a door) that the VLM identified from the outset. This observation suggests that VLM-RB can immediately identify semantically relevant data, enabling faster learning than approaches which must wait for the critic to converge.
+Having demonstrated that VLM scores correlate with well-trained Q-value predictors, we next ask: does this
+{5}------------------------------------------------
+<span id="page-5-1"></span>![](_page_5_Figure_2.jpeg)
+*Figure 4.* Success depends on alignment with semantic priors. We compare the performance of VLM-RB across frames rendered with *Standard* visuals (dark circles), *Misleading* swapped sprites (medium squares), and *Abstract* textures (light diamonds). The dashed line indicates the ASR of PER. Crucially, the agent's actual observations and the underlying MDP are identical across all settings; only the visual input to the VLM is altered (see Appendix [F.3](#page-21-1) for visual samples of these modifications). The significant performance drop in the *Misleading* and *Abstract* settings confirms that our method relies on the VLM correctly recognizing specific semantic objects (e.g., keys and doors) to accelerate learning.
+signal reflect interpretable, human-like priors? To test whether VLM-RB specifically leverages the VLM's pretrained visual semantics, we employ the "Modified Game" paradigm [\(Dubey et al.,](#page-8-13) [2018\)](#page-8-13). We modify only the rendered frames used for VLM scoring, leaving the underlying MDP and agent observations untouched. We consider three settings: (i) the *Standard* game, with unaltered visuals; (ii) the *Misleading* game, where key objects are swapped (for example, traps appear as goals); and (iii) the *Abstract* game, where objects are replaced with random noise patterns (See Appendix [F.3](#page-21-1) for visual examples).
+As shown in Fig. [4,](#page-5-1) the *Misleading* and *Abstract* variants significantly slow learning and increase variance. Notably, in the *Abstract* setting, performance degrades to the level of PER (dashed line). Because the underlying MDP remains unchanged, this degradation demonstrates that the VLM's prioritization is only effective when the visual input aligns with natural semantic priors.
+In summary, these results confirm that VLM-RB scores are indicative of meaningful semantic behavior. We now turn to the question: does this semantic guidance actually yield improved sample efficiency and asymptotic performance in off-policy RL?
+### <span id="page-5-0"></span>4.3. Main Results: VLMs are Useful for RL Data Prioritization
+We evaluate VLM-RB and baselines across both discrete and continuous environments, as detailed in Section [4.1.](#page-3-3)
+<span id="page-5-2"></span>*Table 1.* Improvement with respect to baselines. Performance denotes the best ASR (M = 5 seeds). Sample Efficiency tracks the steps required to reach the *baseline's* best performance. Values in parentheses indicate the relative improvement over the baseline. Both metrics are averaged across the aggregated algorithms. The corresponding training curves are provided in Fig. [9.](#page-24-0)
+| | | | Improvement | | |
+|----------------------|------|------------|----------------------------------------|-----------------------------------------------------------------------------------------------|--|
+| Algorithm | Lvl. | Base. | Perf. (↑) | Sample Eff. (↓) | |
+| | 8×8 | UER<br>PER | +0.0% (1.00/1.00)<br>+0.0% (1.00/1.00) | +19.1% (144K/178K)<br>+23.0% (144K/187K) | |
+| DQN+IQN<br>(DoorKey) | | | | 12×12 UER +61.3% (1.00/0.62) +52.8% (302K/640K)<br>PER +22.0% (1.00/0.82) +32.1% (317K/467K) | |
+| | | | | 16×16 UER +241.7% (0.82/0.24) +37.6% (557K/893K)<br>PER +70.8% (0.82/0.48) +24.1% (472K/622K) | |
+| | 3 | UER<br>PER | +0.0% (1.00/1.00)<br>+0.0% (1.00/1.00) | +40.7% (210K/354K)<br>+21.1% (210K/266K) | |
+| SAC+TD3<br>(Scene) | 4 | PER | +2.0% (1.00/0.98) | UER +22.0% (1.00/0.82) +44.6% (430K/776K)<br>+19.7% (509K/634K) | |
+| | 5 | | | UER +119.4% (0.79/0.36) +46.3% (440K/819K)<br>PER +49.1% (0.79/0.53) +17.9% (661K/805K) | |
+Table [1](#page-5-2) summarizes the aggregate improvements in both performance and sample efficiency. VLM-RB consistently outperforms all baselines, regardless of algorithm or task difficulty. The largest gains appear in the most challenging settings: on DoorKey-16x16, VLM-RB achieves a +241.7% improvement over UER and +70.8% over PER; on Scene-5, the improvement over UER is +119.4% and +49.1% over PER. Even in cases where baselines eventually solve the task, VLM-RB achieves substantially better sample efficiency, reducing the number of steps needed to reach baseline performance by as much as 52.8%.
+Comparison to Baselines. Figure [5](#page-6-1) expands this comparison, showing that VLM-RB is the only method to consistently solve the task when compared to alternative prioritization schemes (AER, ERO, ReLo). The baselines struggle primarily due to the combination of sparse rewards and long sequential dependencies. ERO and ReLo, for example, fail to prioritize task-relevant transitions because their feedback signals (reward gradients and TD-error differences) are either dominated by noise or remain zero throughout the long pre-success phase. On the other hand, similarity-based methods like AER suffer from a structural misalignment: their local similarity metrics lead to the over-sampling of current states while neglecting critical past dependencies (such as key acquisition), which impedes temporal credit assignment over long horizons. Further analysis is provided in Appendix [D.](#page-15-0)
+Training Time Finally, we address the computational trade-off. Although VLM inference introduces a moderate throughput overhead (approximately 12% wall-clock
+{6}------------------------------------------------
+<span id="page-6-1"></span>![](_page_6_Figure_1.jpeg)
+prioritization in sparse-reward tasks. Aggregated success rates on MiniGrid/DoorKey across grid sizes (8x8, 12x12, 16x16). While alternative methods (AER, ERO, ReLo) fail as they depend on dense rewards or local similarity metrics, VLM-RB successfully bridges the long-horizon dependencies. Curves show the mean across 5 seeds, with shaded standard errors.
+time per step, see Appendix [F.4\)](#page-22-0), this cost is outweighed by the reduction in training steps. In the vast majority of configurations (21/24), VLM-RB reduces the absolute wall-clock time to reach peak performance by up to 58%. In the rare cases where wall-clock time increased, the slowdown was marginal (<5%). By contrast, in more challenging exploration settings, VLM-RB consistently achieved speedups exceeding 20-40% (full results can be seen in Table [7](#page-23-0) in Appendix [G\)](#page-23-1). Overall, VLM-RB achieves a favorable tradeoff, substantially accelerating training in difficult sparsereward tasks, while incurring negligible overhead in simpler scenarios.
+### <span id="page-6-0"></span>4.4. Which Design Choices Matter? Mixing and Scale
+Sampling Mixing Schedule. A central component of our approach is to mix VLM-scored samples with uniformly sampled data (see Eq. [2\)](#page-3-4). To assess the importance of this feature, we vary only the maximal mixing coefficient, λmax, holding all other parameters fixed. The results shown in Table [4](#page-20-0) indicate that retaining a fraction of uniform sampling is necessary to stabilize value learning. See Appendix [F.1](#page-20-1) for the full analysis of the results.
+VLM Size. We also ablated the VLM size and found that scaling the model beyond 1B parameters dramatically increases computational cost without yielding consistent improvements in downstream RL performance. Detailed analysis of the trade-offs between throughput, memory, and performance is provided in Appendix [F.2.](#page-21-0)
+## 5. Related Work
+Prioritized Experience Replay. Extensive research has focused on defining *which* transitions maximize learning utility. The seminal work, PER [\(Schaul et al.,](#page-9-4) [2015b\)](#page-9-4), prioritizes transitions with high TD error, serving as a proxy for "surprising" events. Subsequent methods have proposed alternative utility metrics based on state similarity [\(Sun](#page-9-12) [et al.,](#page-9-12) [2020\)](#page-9-12), potential loss reduction [\(Sujit et al.,](#page-9-13) [2023\)](#page-9-13), target quality/discrepancy criteria [\(Kumar et al.,](#page-8-14) [2020\)](#page-8-14), or by explicitly learning a replay policy [\(Zha et al.,](#page-10-6) [2019\)](#page-10-6). However, these approaches typically derive priorities from internal, training-dependent signals (e.g., value estimates or TD errors), which can be noisy or undefined early in training. In contrast, VLM-RB leverages a pre-trained VLM to assign semantic priorities, yielding a robust utility score from the very first update.
+Orthogonal to the choice of *what* to prioritize, several works modify *how* transitions are sampled. [Hong et al.](#page-8-15) [\(2022\)](#page-8-15) argue for enforcing the ordering of updates within a collected episode, to better align with the Bellman loss that is common in off-policy learning. Another example, [Lahire](#page-8-16) [et al.](#page-8-16) [\(2021\)](#page-8-16), focuses on the "staleness" of prioritization scores and suggests a double-sampling approach. These directions are largely complementary to VLM-RB and can be combined with VLM-based priorities. Furthermore, because VLM-RB computes priorities from a frozen VLM, it naturally avoids the score staleness issue addressed by the latter.
+Finally, rather than treating prioritization as a standalone sampling heuristic, a distinct line of work alters the underlying *optimization objective* itself. These methods often derive reweighting schemes from first principles, using fdivergence regularization [\(Li et al.,](#page-9-14) [2024\)](#page-9-14), trust-region constraints [\(Novati & Koumoutsakos,](#page-9-15) [2019\)](#page-9-15), or density-ratio estimation [\(Liu et al.,](#page-9-16) [2021;](#page-9-16) [Sinha et al.,](#page-9-17) [2022\)](#page-9-17) to justify non-uniform data usage. Unlike these approaches, which fundamentally change the loss function, VLM-RB preserves the underlying training objective, ensuring it remains modular and compatible with a wide range of existing algorithms.
+LLM/VLM-Enhanced RL. Recent work has increasingly used foundation models to improve RL by injecting semantic knowledge into the learning loop. Broadly, these methods leverage LLMs/VLMs to (i) define richer learning signals (reward supervision), (ii) propose plans and exploration targets (high-level guidance), or (iii) assist action selection (policy components). Comprehensive surveys can be found in [Cao et al.](#page-8-3) [\(2024\)](#page-8-3) and [Schoepp et al.](#page-9-8) [\(2025\)](#page-9-8).
+A major thread treats LLMs and VLMs as sources of reward. Off-the-shelf VLMs have been used as zero-shot reward scorers to judge state–goal alignment [\(Rocamonde et al.,](#page-9-11)
+{7}------------------------------------------------
+[2023\)](#page-9-11); LLMs can synthesize dense shaping functions from natural-language task descriptions [\(Xie et al.,](#page-10-7) [2023\)](#page-10-7); and several methods replace or augment human feedback by using VLM/LLM judgments (absolute ratings or pairwise preferences) to train reward models [\(Fu et al.,](#page-8-17) [2024;](#page-8-17) [Luu](#page-9-18) [et al.,](#page-9-18) [2025;](#page-9-18) [Singh et al.,](#page-9-19) [2025;](#page-9-19) [Ghosh et al.,](#page-8-18) [2025\)](#page-8-18). Taken together, these approaches reframe reward design as semantic labeling, allowing pre-trained models to provide informative supervision in settings where environment rewards are sparse or absent.
+A second strand uses foundation models for planning and exploration. LLMs decompose instructions into temporallyextended plans grounded in affordances [\(Ahn et al.,](#page-7-0) [2022;](#page-7-0) [Huang et al.,](#page-8-19) [2022\)](#page-8-19), and guide exploration by proposing semantically diverse subgoals or novelty-oriented objectives using learned embeddings [\(Ma et al.,](#page-9-20) [2025;](#page-9-20) [Gupta et al.,](#page-8-20) [2022\)](#page-8-20). This high-level guidance is particularly useful in settings where unguided exploration rarely visits meaningful states.
+A third direction integrates LLMs/VLMs more directly into control, either as action advisors or as policy components. Examples range from generating executable policy code [\(Liang et al.,](#page-9-21) [2022\)](#page-9-21) to large-scale vision–language– action (VLA) models that map multimodal inputs to actions [\(Zitkovich et al.,](#page-10-8) [2023\)](#page-10-8).
+Despite this breadth, most prior work uses these models to shape what the agent optimizes (rewards), where it searches (plans and exploration), or how it acts (policy), which commonly introduces additional training stages, fine-tuning, or computationally expensive inference. In contrast, VLM-RB targets the replay buffer—the core mechanism enabling sample reuse in off-policy RL—a component which remains underexplored for semantic foundation-model supervision. Moreover, as our experiments show, open-source pre-trained VLMs can already yield consistent gains when used to prioritize replay, without any additional model training or finetuning.
+# 6. Conclusions
+In this work, we presented VLM-RB, a framework that integrates the semantic reasoning capabilities of pre-trained VLMs directly into the experience replay mechanism. By shifting prioritization from statistical proxies such as TDerror to semantic evaluations of task progress, VLM-RB effectively identifies and promotes meaningful experiences even in sparse-reward regimes where traditional metrics fail. Our empirical results demonstrate that this approach significantly improves both sample efficiency and asymptotic performance across discrete and continuous domains, without requiring fine-tuning or gradient updates to the VLM.
+Limitations. While our approach yields substantial improvements, it is important to clarify the specific assumptions and constraints under which these gains are realized. First, VLM-RB fundamentally relies on the assumption that task progress can be reliably inferred from visual observations. As a result, VLM-RB does not apply to domains where the underlying state lacks a visual representation, such as non-spatial biological systems or abstract network control tasks. In these settings, the VLM cannot ground its reasoning. Second, querying the VLM introduces additional computational overhead. Empirically, we found that this overhead reduced throughput by approximately 12% in our experiments. It is important to note that this computational cost is not intrinsic to the method; the precise overhead depends on hardware configuration and implementation choices, so the trade-off may vary across different setups[4](#page-7-1) .
+Future Work. Our findings suggest several concrete directions for future research. A natural extension is the application of VLM-RB to Goal-Conditioned RL [\(Schaul](#page-9-22) [et al.,](#page-9-22) [2015a\)](#page-9-22), where the textual description of the current goal can be injected into the VLM prompt to dynamically score alignment, or hindsight textual goals could promote meaningful sub-goals which may have some semantic relation but no direct relevance to the current task [\(Luu &](#page-9-23) [Yoo,](#page-9-23) [2021;](#page-9-23) [Jiang et al.,](#page-8-21) [2019\)](#page-8-21). Another promising direction is to use the replay buffer as a curriculum mechanism, by evolving the prompt over the course of training. For example, the system could prioritize specific skills such as "open drawer" in the early stages, and later shift focus to more compositional behaviors. Finally, the prioritization scheme could be extended beyond single VLM scoring, for example, by incorporating more robust inference techniques such as majority voting or LLM-as-a-Judge.
+Impact Statement. This paper presents work aimed at advancing the field of machine learning, specifically in the domain of sample-efficient reinforcement learning. There are many potential societal consequences of our work, none of which we feel must be specifically highlighted here.
+## References
+<span id="page-7-0"></span>Ahn, M., Brohan, A., Brown, N., Chebotar, Y., Cortes, O., David, B., Finn, C., Fu, C., Gopalakrishnan, K., Hausman, K., et al. Do as i can, not as i say: Grounding language in robotic affordances. *arXiv preprint arXiv:2204.01691*, 2022.
+<span id="page-7-2"></span>Bou, A., Bettini, M., Dittert, S., Kumar, V., Sodhani, S.,
+<span id="page-7-1"></span><sup>4</sup>Our implementation did not employ aggressive optimization techniques (e.g., quantization or dedicated serving stacks), suggesting that this overhead could be further minimized.
+{8}------------------------------------------------
+440 441 442 Yang, X., De Fabritiis, G., and Moens, V. Torchrl: A data-driven decision-making library for pytorch. *arXiv preprint arXiv:2306.00577*, 2023.
+- <span id="page-8-3"></span>Cao, Y., Zhao, H., Cheng, Y., Shu, T., Chen, Y., Liu, G., Liang, G., Zhao, J., Yan, J., and Li, Y. Survey on large language model-enhanced reinforcement learning: Concept, taxonomy, and methods. *IEEE Transactions on Neural Networks and Learning Systems*, 2024.
+- <span id="page-8-4"></span>Carrasco-Davis, R., Lee, S., Clopath, C., and Dabney, W. Uncertainty prioritized experience replay. *arXiv preprint arXiv:2506.09270*, 2025.
+- <span id="page-8-6"></span>Cherti, M., Beaumont, R., Wightman, R., Wortsman, M., Ilharco, G., Gordon, C., Schuhmann, C., Schmidt, L., and Jitsev, J. Reproducible scaling laws for contrastive language-image learning. In *Proceedings of the IEEE/CVF conference on computer vision and pattern recognition*, pp. 2818–2829, 2023.
+- <span id="page-8-9"></span>Chevalier-Boisvert, M., Dai, B., Towers, M., Perez-Vicente, R., Willems, L., Lahlou, S., Pal, S., Castro, P. S., and Terry, J. Minigrid & miniworld: Modular & customizable reinforcement learning environments for goal-oriented tasks. *Advances in Neural Information Processing Systems*, 36:73383–73394, 2023.
+- <span id="page-8-8"></span>Cho, J. H., Madotto, A., Mavroudi, E., Afouras, T., Nagarajan, T., Maaz, M., Song, Y., Ma, T., Hu, S., Rasheed, H., Sun, P., Huang, P.-Y., Bolya, D., Jain, S., Martin, M., Wang, H., Ravi, N., Jain, S., Stark, T., Moon, S., Damavandi, B., Lee, V., Westbury, A., Khan, S., Krahenb ¨ uhl, ¨ P., Dollar, P., Torresani, L., Grauman, K., and Feichten- ´ hofer, C. Perceptionlm: Open-access data and models for detailed visual understanding. *arXiv:2504.13180*, 2025.
+- <span id="page-8-10"></span>Dabney, W., Ostrovski, G., Silver, D., and Munos, R. Implicit quantile networks for distributional reinforcement learning. In *International conference on machine learning*, pp. 1096–1105. PMLR, 2018.
+- <span id="page-8-13"></span>Dubey, R., Agrawal, P., Pathak, D., Griffiths, T. L., and Efros, A. A. Investigating human priors for playing video games. *arXiv preprint arXiv:1802.10217*, 2018.
+- <span id="page-8-17"></span>Fu, Y., Zhang, H., Wu, D., Xu, W., and Boulet, B. Furl: Visual-language models as fuzzy rewards for reinforcement learning. *arXiv preprint arXiv:2406.00645*, 2024.
+- <span id="page-8-11"></span>Fujimoto, S., Hoof, H., and Meger, D. Addressing function approximation error in actor-critic methods. In *International conference on machine learning*, pp. 1587–1596. PMLR, 2018.
+- <span id="page-8-18"></span>Ghosh, U., Raychaudhuri, D. S., Li, J., Karydis, K., and Roy-Chowdhury, A. Preference vlm: Leveraging vlms for scalable preference-based reinforcement learning. *arXiv preprint arXiv:2502.01616*, 2025.
+- <span id="page-8-7"></span>Goel, S., Bansal, H., Bhatia, S., Rossi, R., Vinay, V., and Grover, A. Cyclip: Cyclic contrastive language-image pretraining. *Advances in Neural Information Processing Systems*, 35:6704–6719, 2022.
+- <span id="page-8-20"></span>Gupta, T., Karkus, P., Che, T., Xu, D., and Pavone, M. Foundation models for semantic novelty in reinforcement learning. *arXiv preprint arXiv:2211.04878*, 2022.
+- <span id="page-8-12"></span>Haarnoja, T., Zhou, A., Hartikainen, K., Tucker, G., Ha, S., Tan, J., Kumar, V., Zhu, H., Gupta, A., Abbeel, P., et al. Soft actor-critic algorithms and applications. *arXiv preprint arXiv:1812.05905*, 2018.
+- <span id="page-8-15"></span>Hong, Z.-W., Chen, T., Lin, Y.-C., Pajarinen, J., and Agrawal, P. Topological experience replay. *arXiv preprint arXiv:2203.15845*, 2022.
+- <span id="page-8-19"></span>Huang, W., Abbeel, P., Pathak, D., and Mordatch, I. Language models as zero-shot planners: Extracting actionable knowledge for embodied agents. In *International conference on machine learning*, pp. 9118–9147. PMLR, 2022.
+- <span id="page-8-5"></span>Jia, C., Yang, Y., Xia, Y., Chen, Y.-T., Parekh, Z., Pham, H., Le, Q., Sung, Y.-H., Li, Z., and Duerig, T. Scaling up visual and vision-language representation learning with noisy text supervision. In *International conference on machine learning*, pp. 4904–4916. PMLR, 2021.
+- <span id="page-8-21"></span>Jiang, Y., Gu, S. S., Murphy, K. P., and Finn, C. Language as an abstraction for hierarchical deep reinforcement learning. *Advances in neural information processing systems*, 32, 2019.
+- <span id="page-8-1"></span>Kalashnikov, D., Varley, J., Chebotar, Y., Swanson, B., Jonschkowski, R., Finn, C., Levine, S., and Hausman, K. Scaling up multi-task robotic reinforcement learning. In *5th Annual Conference on Robot Learning*, 2021.
+- <span id="page-8-2"></span>Kroemer, O., Niekum, S., and Konidaris, G. A review of robot learning for manipulation: Challenges, representations, and algorithms. *Journal of machine learning research*, 22(30):1–82, 2021.
+- <span id="page-8-14"></span>Kumar, A., Gupta, A., and Levine, S. Discor: Corrective feedback in reinforcement learning via distribution correction. *Advances in neural information processing systems*, 33:18560–18572, 2020.
+- <span id="page-8-16"></span>Lahire, T., Geist, M., and Rachelson, E. Large batch experience replay. *arXiv preprint arXiv:2110.01528*, 2021.
+- <span id="page-8-0"></span>Levine, S., Finn, C., Darrell, T., and Abbeel, P. End-to-end training of deep visuomotor policies. *Journal of Machine Learning Research*, 17(39):1–40, 2016.
+{9}------------------------------------------------
+<span id="page-9-14"></span>Li, C., Hong, Z.-W., Agrawal, P., Garg, D., and Pajarinen, J. Roer: Regularized optimal experience replay. *arXiv preprint arXiv:2407.03995*, 2024.
+<span id="page-9-21"></span>504
+<span id="page-9-2"></span>513 514
+<span id="page-9-16"></span>516
+<span id="page-9-18"></span>524 525 526
+<span id="page-9-3"></span>536
+- <span id="page-9-10"></span>Li, J., Li, D., Savarese, S., and Hoi, S. Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models. In *International conference on machine learning*, pp. 19730–19742. PMLR, 2023.
+- Liang, J., Huang, W., Xia, F., Xu, P., Hausman, K., Ichter, B., Florence, P., and Zeng, A. Code as policies: Language model programs for embodied control. *arXiv preprint arXiv:2209.07753*, 2022.
+- <span id="page-9-7"></span>Liang, Z., Xu, Y., Hong, Y., Shang, P., Wang, Q., Fu, Q., and Liu, K. A survey of multimodel large language models. In *Proceedings of the 3rd International Conference on Computer, Artificial Intelligence and Control Engineering*, pp. 405–409, 2024.
+- Lin, L.-J. Self-improving reactive agents based on reinforcement learning, planning and teaching. *Machine learning*, 8(3):293–321, 1992.
+- Liu, X.-H., Xue, Z., Pang, J., Jiang, S., Xu, F., and Yu, Y. Regret minimization experience replay in off-policy reinforcement learning. *Advances in neural information processing systems*, 34:17604–17615, 2021.
+- <span id="page-9-23"></span>Luu, T. M. and Yoo, C. D. Hindsight goal ranking on replay buffer for sparse reward environment. *IEEE Access*, 9: 51996–52007, 2021.
+- Luu, T. M., Lee, Y., Lee, D., Kim, S., Kim, M. J., and Yoo, C. D. Enhancing rating-based reinforcement learning to effectively leverage feedback from large vision-language models. *arXiv preprint arXiv:2506.12822*, 2025.
+- <span id="page-9-20"></span>Ma, R., Luijkx, J., Ajanovic, Z., and Kober, J. Explorllm: ´ Guiding exploration in reinforcement learning with large language models. In *2025 IEEE International Conference on Robotics and Automation (ICRA)*, pp. 9011–9017. IEEE, 2025.
+- Mnih, V., Kavukcuoglu, K., Silver, D., Graves, A., Antonoglou, I., Wierstra, D., and Riedmiller, M. Playing atari with deep reinforcement learning. *arXiv preprint arXiv:1312.5602*, 2013.
+- <span id="page-9-15"></span>Novati, G. and Koumoutsakos, P. Remember and forget for experience replay. In *International Conference on Machine Learning*, pp. 4851–4860. PMLR, 2019.
+- <span id="page-9-0"></span>Ouyang, L., Wu, J., Jiang, X., Almeida, D., Wainwright, C., Mishkin, P., Zhang, C., Agarwal, S., Slama, K., Ray, A., et al. Training language models to follow instructions with human feedback. *Advances in neural information processing systems*, 35:27730–27744, 2022.
+- <span id="page-9-6"></span>Park, S., Frans, K., Eysenbach, B., and Levine, S. Ogbench: Benchmarking offline goal-conditioned rl. *arXiv preprint arXiv:2410.20092*, 2024.
+- <span id="page-9-9"></span>Radford, A., Kim, J. W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al. Learning transferable visual models from natural language supervision. In *International conference on machine learning*, pp. 8748–8763. PmLR, 2021.
+- <span id="page-9-11"></span>Rocamonde, J., Montesinos, V., Nava, E., Perez, E., and Lindner, D. Vision-language models are zero-shot reward models for reinforcement learning. *arXiv preprint arXiv:2310.12921*, 2023.
+- <span id="page-9-22"></span>Schaul, T., Horgan, D., Gregor, K., and Silver, D. Universal value function approximators. In *International conference on machine learning*, pp. 1312–1320. PMLR, 2015a.
+- <span id="page-9-4"></span>Schaul, T., Quan, J., Antonoglou, I., and Silver, D. Prioritized experience replay. *arXiv preprint arXiv:1511.05952*, 2015b.
+- <span id="page-9-8"></span>Schoepp, S., Jafaripour, M., Cao, Y., Yang, T., Abdollahi, F., Golestan, S., Sufiyan, Z., Zaiane, O. R., and Taylor, M. E. The evolving landscape of llm-and vlm-integrated reinforcement learning. *arXiv preprint arXiv:2502.15214*, 2025.
+- <span id="page-9-19"></span>Singh, A., Bhaskar, A., Yu, P., Chakraborty, S., Dasyam, R., Bedi, A., and Tokekar, P. Varp: Reinforcement learning from vision-language model feedback with agent regularized preferences. *arXiv preprint arXiv:2503.13817*, 2025.
+- <span id="page-9-17"></span>Sinha, S., Song, J., Garg, A., and Ermon, S. Experience replay with likelihood-free importance weights. In *Learning for Dynamics and Control Conference*, pp. 110–123. PMLR, 2022.
+- <span id="page-9-13"></span>Sujit, S., Nath, S., Braga, P., and Ebrahimi Kahou, S. Prioritizing samples in reinforcement learning with reducible loss. *Advances in Neural Information Processing Systems*, 36:23237–23258, 2023.
+- <span id="page-9-12"></span>Sun, P., Zhou, W., and Li, H. Attentive experience replay. In *Proceedings of the AAAI conference on artificial intelligence*, volume 34, pp. 5900–5907, 2020.
+- <span id="page-9-5"></span>Sutton, R. S., Barto, A. G., et al. *Reinforcement learning: An introduction*, volume 1. MIT press Cambridge, 1998.
+- <span id="page-9-1"></span>Tang, C., Abbatematteo, B., Hu, J., Chandra, R., Mart´ın-Mart´ın, R., and Stone, P. Deep reinforcement learning for robotics: A survey of real-world successes. In *Proceedings of the AAAI Conference on Artificial Intelligence*, volume 39, pp. 28694–28698, 2025.
+{10}------------------------------------------------
+<span id="page-10-5"></span>550 Van Hasselt, H., Guez, A., and Silver, D. Deep reinforcement learning with double q-learning. In *Proceedings of the AAAI conference on artificial intelligence*, volume 30, 2016.
+<span id="page-10-7"></span>554
+<span id="page-10-2"></span>576
+- Xie, T., Zhao, S., Wu, C. H., Liu, Y., Luo, Q., Zhong, V., Yang, Y., and Yu, T. Text2reward: Reward shaping with language models for reinforcement learning. *arXiv preprint arXiv:2309.11489*, 2023.
+- <span id="page-10-6"></span>Zha, D., Lai, K.-H., Zhou, K., and Hu, X. Experience replay optimization. *arXiv preprint arXiv:1906.08387*, 2019.
+- <span id="page-10-4"></span>Zhai, X., Wang, X., Mustafa, B., Steiner, A., Keysers, D., Kolesnikov, A., and Beyer, L. Lit: Zero-shot transfer with locked-image text tuning. In *Proceedings of the IEEE/CVF conference on computer vision and pattern recognition*, pp. 18123–18133, 2022.
+- <span id="page-10-3"></span>Zhai, X., Mustafa, B., Kolesnikov, A., and Beyer, L. Sigmoid loss for language image pre-training. In *Proceedings of the IEEE/CVF international conference on computer vision*, pp. 11975–11986, 2023.
+- <span id="page-10-1"></span>Zhang, J., Huang, J., Jin, S., and Lu, S. Vision-language models for vision tasks: A survey. *IEEE transactions on pattern analysis and machine intelligence*, 46(8):5625– 5644, 2024.
+- Zhao, R. and Tresp, V. Curiosity-driven experience prioritization via density estimation. *arXiv preprint arXiv:1902.08039*, 2019.
+- <span id="page-10-0"></span>Zhu, H., Yu, J., Gupta, A., Shah, D., Hartikainen, K., Singh, A., Kumar, V., and Levine, S. The ingredients of realworld robotic reinforcement learning. *arXiv preprint arXiv:2004.12570*, 2020.
+- <span id="page-10-8"></span>Zitkovich, B., Yu, T., Xu, S., Xu, P., Xiao, T., Xia, F., Wu, J., Wohlhart, P., Welker, S., Wahid, A., et al. Rt-2: Vision-language-action models transfer web knowledge to robotic control. In *Conference on Robot Learning*, pp. 2165–2183. PMLR, 2023.
+{11}------------------------------------------------
+#### A. VLM-RB
+632633634
+638639
+648649
+653654
+### <span id="page-11-0"></span>**Algorithm 1 VLM-Prioritized Replay Buffer**
+```
+Require: prompt P; clip length L.
+replay buffer \mathcal{D}; clip buffer \mathcal{C}; queues \mathcal{Q}_{\mathrm{in}} and \mathcal{Q}_{\mathrm{out}}.
+1: function VLMWORKER
+2:
+while true do
+(idxs, \tau^O) \leftarrow pop(Q_{in})
+3:
+\mathbf{p^{VLM}} = f_{\mathrm{VLM}}(\tau^O, \mathsf{P}) \in \mathbb{R}
+4:
+(Equation 1)
+\operatorname{push}(\mathcal{Q}_{\mathrm{out}},(\operatorname{idxs},\operatorname{\mathbf{p}^{\acute{\mathbf{VLM}}}}))
+5:
+6: Init: launch VLMWORKER asynchronously.
+7:
+while Training do
+8:
+⊳(1) Env step
+9:
+Step env to obtain (s, a, r, s') and visual observation o = \psi(s)
+10:
+idx \leftarrow INSERT(\mathcal{D}, (s, a, r, s'; \bar{p}))
+\mathcal{C} \leftarrow \mathcal{C} \cup (\text{idx}, o)
+11:
+12:
+⊳ (2) Enqueue clips (streaming)
+if |\mathcal{C}| = L or terminated or truncated then
+13:
+idxs, \tau^O \leftarrow \{(idx_i, o_i)\}_{i=0}^{|C|-1}
+14:
+\operatorname{push}(\mathcal{Q}_{\operatorname{in}},(\operatorname{idxs},\tau^O))
+15:
+\mathcal{C} \leftarrow \emptyset
+16:
+▷ (3) Apply VLM scores (drain output queue)
+17:
+while \mathcal{Q}_{\mathrm{out}} not empty do
+18:
+(\texttt{idxs}, \mathbf{p^{VLM}}) \leftarrow \mathbf{pop}(\mathcal{Q}_{\text{out}})
+19:
+20:
+for all idx \in idxs do
+\texttt{SETPRIORITY}(\mathcal{D}, \texttt{idx}, \mathbf{p^{VLM}})
+21:
+\bar{p} \leftarrow \text{CMA}(\bar{p}, \mathbf{p^{VLM}})
+22:
+\triangleright (4) Sample and learn
+23:
+\mathcal{B} \sim \lambda_t q_t^{\hat{P}} + (1 - \lambda_t) q_t^{U}
+(Equation 2)
+25:
+UPDATELEARNER(\mathcal{B}); update \lambda_t
+```
+In Algorithm 1, C denotes a temporary clip buffer, $Q_{in}$ and $Q_{out}$ are asynchronous communication queues, and CMA refers to the Cumulative Moving Average update rule for the global priority statistics.
+#### <span id="page-11-1"></span>**B.** Environment Details
+In this section, we describe the environments used in our experiments.
+MiniGrid / DoorKey. We use the <code>DoorKey</code> tasks from the MiniGrid suite (Chevalier-Boisvert et al., 2023) with sizes 8x8, 12x12, and 16x16. Each episode requires the agent to complete a sequence of subtasks: (i) navigate to and pick up the key, (ii) reach the locked door and open it using the <code>toggle</code> action (which is only possible while holding the key), and (iii) proceed to the goal tile. This structure enforces temporal dependencies and compositional reasoning. The reward function is sparse: the agent receives zero reward at all intermediate steps and a positive reward only upon reaching the goal, with the reward magnitude linearly decaying with the number of steps taken, following the standard MiniGrid protocol. Episodes terminate either upon successful completion or when the maximum step limit is reached.
+State observations are provided in a symbolic format, where each $s_t \in \{0, 1, \ldots\}^{N \times N \times 3}$ encodes the full grid. The three channels correspond to (i) object indices, (ii) color indices, and (iii) a state channel that includes both the door state and the agent's orientation. The action space is discrete with $|\mathcal{A}| = 5$ , consisting of turn left, turn right, move forward, pick up, and toggle actions.
+**OGBench / Scene.** We use the scene-play manipulation environment from OGBench (Park et al., 2024), which is a MuJoCo-based tabletop task involving a UR5e arm with a parallel gripper. The agent interacts with a set of objects: a cube, a sliding drawer, a sliding window, and two toggle buttons. Each button controls the lock state of either the drawer or the window, introducing dependencies that require the agent to execute explicit unlock, manipulate, and (re-)lock sequences to achieve many goals. Episodes are capped at a horizon of 750 steps. To facilitate learning, we employ potential-based reward
+{12}------------------------------------------------
+shaping with a per-step living cost, defined as:
+$$r_t = -1 \ + \ \gamma \, \phi(s_{t+1}) - \phi(s_t), \qquad \phi(s) = \frac{1}{K} \sum_{i=1}^K \mathbb{I} \left[ \mathrm{subgoal}_i(s) \right].$$
+Here, $\phi(s)$ denotes the fraction of satisfied subgoals, where each subgoal is a task-specific predicate over cube placement, button states, and the positions of the drawer and window. An episode terminates either when all subgoals are satisfied $(\phi(s_t) = 1)$ or when the maximum horizon is reached.
+Observations are provided as full state vectors, $s_t \in \mathbb{R}^{40}$ , and actions are specified as continuous end-effector controls, $a_t \in \mathbb{R}^5$ .
+To address the exploration challenges inherent in these long-horizon tasks, we initialize the replay buffer with a lightweight warm-start of 10 demonstration episodes, which are held fixed across all methods. This small set of demonstrations provides minimal task-relevant coverage, while leaving the remainder of the training protocol unchanged.
+We evaluate on a set of predefined Scene goals (tasks 3–5), which are designed to progressively increase the degree of temporal composition required for successful completion.
+- **Task 3 (rearrange-medium).** The agent must move the cube to a specified tabletop location, open the drawer, close the window, and terminate with both the drawer and window unlocked. Notably, the window begins in a locked and open state, so the policy must first unlock it before closing, while simultaneously coordinating the manipulation of the drawer and the relocation of the cube.
+- **Task 4 (put-in-drawer).** The agent is required to place the cube inside the drawer and terminate with the drawer closed and unlocked, while ensuring the window remains locked. This sequence involves unlocking the drawer, opening it, inserting the cube, and then closing the drawer.
+- **Task 5** (**rearrange-hard**). The agent must place the cube inside the (closed) drawer and leave the window open, while ensuring that both the drawer and window are locked at the end of the episode. Achieving this goal requires the agent to execute both unlock and relock sequences, coordinate drawer opening and closing, and correctly position the cube.
+{13}------------------------------------------------
+## <span id="page-13-0"></span>C. Prompts
+Task prompts used in our experiments. We use a binary "success visible" query with a strict output format.
+- 1. MiniGrid/DoorKey: *"Does this clip contain a clear instance of goal satisfaction anywhere in it? If no visible success occurs, answer No. Do not guess. Output exactly Answer: Yes or Answer: No."*
+- 2. OGBench/Scene: *"Is there at least one clear instance of goal satisfaction in these frames? Look for contact + displacement consistent with the goal (lift off surface, place into receptacle, open/close articulation, move to target zone). Do not guess. If not visible, answer No. Output exactly Answer: Yes or Answer: No."*
+Meta-prompt for generating task prompts. To facilitate applying our procedure to new environments, we use the following meta-prompt to generate a task prompt from a task identifier and optional human context. The meta-prompt encourages generic, visually grounded success criteria and enforces a strict output schema.
+### Meta-prompt template for generating task prompts
+User: You are an expert in Reinforcement Learning and Visual Language Models. I will provide a short clip from an agent rollout. Your job is to write a single text prompt that asks a VLM to output a binary judgment about whether goal satisfaction / competent task progress is clearly visible in the clip. Requirements:
+- Keep the prompt environment-agnostic: rely on visible physics and outcomes rather than simulator-specific rules.
+- Specify what to look for in broad categories when helpful (e.g., contact + displacement for manipulation; reaching a target region for navigation).
+- Explicitly instruct: do not guess; if success is not clearly visible, answer No.
+- End the prompt with the string: Output exactly Answer: Yes or Answer: No.
+#### Examples (few-shot):
+• Task: Minigrid-doorkey
+Prompt: "Does this clip contain a clear instance of goal satisfaction anywhere in it? If no visible success occurs, answer No. Do not guess. Output exactly Answer: Yes or Answer: No."
+• Task: OGBench-scene-task-4
+Prompt: "Is there at least one clear instance of goal satisfaction in these frames? Look for contact + displacement consistent with the goal (lift off surface, place into receptacle, open/close articulation, move to target zone). Do not guess. If not visible, answer No. Output exactly Answer: Yes or Answer: No."
+#### Current request:
+Task Name: {{TASK ID}}
+Optional Human Context: {{HUMAN CONTEXT}}
+{14}------------------------------------------------
+> 822 823 824
+Example: prompt families for a new benchmark. As an illustration, we can apply the meta-prompt to a suite such as DeepMind Control by grouping tasks by their visible success patterns (locomotion / balance / interaction) and instantiating one generic prompt per group.
+### Example: prompt families for DeepMind Control Suite
+User: Group DeepMind Control tasks by shared *visible* success patterns and propose one generic success-detection prompt per group (not task-specific). Each prompt must end with: "Output exactly Answer: Yes or Answer: No."
+Model: We used Gemini 3 Pro to generate these prompts. Here is the partition into three semantic groups based on the nature of the movement, with a generalized "vibe-check" prompt for each.
+#### Group 1: Going Somewhere (Locomotion)
+The Semantic Goal: The agent must make progress across the environment. The VLM needs to distinguish between an agent that is struggling/falling (chaos) and one that is moving purposefully (control).
+The Group: cheetah-run, dog-walk, dog-trot, dog-run, fish-swim, hopper-hop, humanoid-walk, humanoid-run, quadruped-walk, quadruped-run, walker-walk, walker-run.
+#### The Prompt:
+"Does the agent demonstrate controlled, purposeful movement through the environment? Look for the agent making progress in a specific direction (locomotion) rather than just flailing in place, falling over, or crashing. The movement should look stable and competent. If the agent falls, flips, or fails to move effectively, answer No. Output exactly Answer: Yes or Answer: No."
+#### Group 2: Defying Gravity (Stabilization)
+The Semantic Goal: The agent is fighting gravity. Success is defined by stillness or verticality. Whether it is a pole swinging up or a dog standing still, the "vibe" of success is reaching a stable, upright state and holding it.
+The Group: acrobot-swingup, cartpole-balance, cartpole-balance sparse, cartpole-swingup, cartpole-swingup sparse, dog-stand, hopper-stand, humanoid-stand, walker-stand, pendulum-swingup.
+#### The Prompt:
+"Does the agent successfully achieve and maintain a stable, upright posture against gravity? Look for the agent balancing itself or an object vertically. Success means the agent (or the object it controls) is standing up or swinging up to a balanced position and staying there without collapsing, spinning out of control, or falling over. If it looks unstable or falls, answer No. Output exactly Answer: Yes or Answer: No."
+#### Group 3: Touching the Thing (Interaction)
+The Semantic Goal: This is about precise manipulation. The agent is not trying to go somewhere or just stand up; it is trying to force an interaction between its end-effector and a target object.
+The Group: ball in cup-catch, finger-spin, finger-turn easy, finger-turn hard, reacher-easy, reacher-hard.
+#### The Prompt:
+"Does the agent successfully interact with the target object in this scene? Look for the agent reaching a target, catching an object, or spinning a mechanism as intended. There should be clear, controlled contact or manipulation consistent with solving a task. If the agent misses the target, drops the object, or flails without making meaningful contact, answer No. Output exactly Answer: Yes or Answer: No."
+{15}------------------------------------------------
+## <span id="page-15-0"></span>D. Baselines
+DoorKey presents a particularly challenging regime, characterized by extremely sparse rewards (identically zero until termination) and a long-horizon, sequential dependency structure (key to door to goal). In such settings, replay schemes relying on dense feedback, informative TD statistics, or a well-defined similarity metric are fundamentally misaligned with the task: they fail to prioritize transitions relevant to solving the task, particularly throughout the extended pre-success phase.
+Experience Replay Optimization (ERO) [\(Zha et al.,](#page-10-6) [2019\)](#page-10-6) ERO optimizes a learned rejection sampler using a REINFORCE-style update, where the scalar reward is defined as the change in evaluation return. In DoorKey, evaluation returns remain identically zero for a substantial portion of training (reflecting the absence of successes), which in turn implies rreplay ≈ 0 and results in near-zero gradients for the replay policy over extended periods. Furthermore, the replay-policy features used here, (r<sup>i</sup> , |δ<sup>i</sup> |, ti/Tmax), are weakly informative under sparse rewards: r<sup>i</sup> = 0 for almost all transitions and |δ<sup>i</sup> | is largely determined by bootstrap noise in the early stages of learning. As a result, the rejection policy effectively acts as an untrained stochastic filter for much of the training. When successes eventually occur, rreplay becomes highly variable, leading to unstable and non-stationary updates which can bias acceptance toward incidental correlates, such as late timesteps. The top-up rule, which fills with the highest-P<sup>i</sup> candidates when acceptance is low, further concentrates replay on a narrow subset of transitions. This reduces diversity and fails to provide a consistent signal for solving the task.
+ReLo (Reducible Loss Prioritization) [\(Sujit et al.,](#page-9-13) [2023\)](#page-9-13) ReLo prioritizes transitions based on the difference between the magnitudes of online and target TD residuals, pReLo = max(0, |δθ| − |δθ<sup>−</sup> |) +ϵ. In the sparse-reward DoorKey setting, both residuals are typically dominated by bootstrapping error rather than meaningful reward propagation throughout the prolonged pre-success phase. Moreover, the use of hard target updates periodically synchronizes the online and target networks, further diminishing any systematic separation between |δθ| and |δθ<sup>−</sup> |. Consequently, |δθ| − |δθ<sup>−</sup> | is often small or negative and is clipped to approximately ϵ, so the sampling distribution effectively reverts toward uniform, yet still incurs the variance and bias tradeoffs associated with prioritized replay.
+Attentive Experience Replay (AER) [\(Sun et al.,](#page-9-12) [2020\)](#page-9-12) AER selects samples closest (in a frozen embedding space) to the agent's current state, thereby inducing a strongly local replay distribution. In DoorKey this notion of locality is fundamentally misaligned with the task structure. Once trajectories reach later phases (near the door or goal), nearestneighbor replay disproportionately samples those regions and neglects earlier structural dependencies, such as key acquisition, thereby impeding temporal credit assignment across the long horizon. This effect is further amplified by the use of a randomly-MLinitialized frozen encoder: the resulting metric primarily captures superficial spatial proximity in the grid encoding, rather than functional or task-relevant similarity. As a result, the selection rule acts as a myopic location-based filter, rather than an attentive semantic sampler.
+{16}------------------------------------------------
+## <span id="page-16-0"></span>E. Implementation Details
+To ensure a fair comparison, all baselines are implemented with the same backbone architecture, optimizer, and training schedule as **VLM-RB**. Unless otherwise specified, each method employs a **PER**-style replay buffer parameterized by exponents $\alpha$ and $\beta$ .
+**Uniform Experience Replay** In this setting, each transition i in the buffer is sampled with equal probability, $p_i = 1/N_{curr}$ , where $N_{curr}$ denotes the current buffer size.
+**PER** (**Prioritized Experience Replay**) Here, the priority assigned to each transition i is given by $p_i = |\delta_i| + \epsilon$ , where $\delta_i$ denotes the most recent TD-error for that transition, and $\epsilon = 10^{-6}$ guarantees that every transition can be sampled. The probability of sampling transition i is then defined as
+$$\mathbf{p}_i = \frac{p_i^{\alpha}}{\sum_k p_k^{\alpha}},$$
+with $\alpha$ determining the extent to which prioritization influences sampling. To account for the bias from prioritized sampling, importance-sampling (IS) weights are computed as
+$$w_i = \left(\frac{1}{N} \cdot \frac{1}{p_i}\right)^{\beta}.$$
+These weights are normalized by $1/\max_i(w_i)$ prior to being used in the loss.
+**Experience Replay Optimization (ERO)** ERO (Zha et al., 2019) replaces standard replay sampling with a learned rejection policy applied to uniformly drawn candidates. The replay policy is parameterized as an MLP $\phi_{\psi}$ with two hidden layers of 64 units each (ReLU activations, sigmoid output), mapping transition features to a retention probability.
+$$p_i = \phi_{\psi}(r_i, |\delta_i|, t_i/T_{\text{max}}) \in (0, 1),$$
+$$\mathcal{L}_{ERO} = -r_{replay} \sum_{i \in \mathcal{B}} \log p_i,$$
+where gradients are applied only to $\psi$ .
+**ReLo** (Reducible Loss Prioritization) ReLo (Sujit et al., 2023) assigns priority to each transition based on a reducible-loss score, which is computed from the online and target Q-networks. Given a transition (s,a,r,s'), the target value is defined as $y=r+\gamma\max_{a'}Q_{\theta^-}(s',a')$ . The online and target TD-errors are then $\delta_{\text{online}}(s,a)=Q_{\theta}(s,a)-y$ and $\delta_{\text{target}}(s,a)=Q_{\theta^-}(s,a)-y$ , respectively. The sampling priority is
+$$p_{\text{ReLo}}(s, a) = \max(0, |\delta_{\text{online}}(s, a)| - |\delta_{\text{target}}(s, a)|) + \epsilon, \qquad \epsilon = 10^{-6}.$$
+Each new transition is initialized with the maximum priority value among all transitions currently in the buffer. The method sets $\alpha = 0.6$ and initializes $\beta = 0.4$ , annealing $\beta$ linearly to 1.0. Standard importance-sampling weights are applied.
+{17}------------------------------------------------
+Attentive Experience Replay (AER) AER (Sun et al., 2020) selects transitions using an attentive sampling mechanism. A dedicated encoder $\phi$ is constructed with the same backbone architecture as the Q-network. This encoder is randomly initialized and remains fixed throughout training. At each training step t, a candidate pool of size $N_{\text{cand}} = \lfloor \lambda_t B \rfloor$ is sampled uniformly from the buffer, where $\lambda_t$ decays linearly from $\lambda_0 = 4$ to 1 over the course of training. If the candidate pool size $N_{\text{cand}}$ is less than or equal to the batch size B, the method reverts to the default sampling strategy. Given the current state $s_{\text{curr}}$ and a set of candidate states $\{s_i\}$ , distances are computed in the frozen embedding space as $d(s_{\text{curr}}, s_i) = \|\phi(s_{\text{curr}}) - \phi(s_i)\|_2^2$ . The B candidates with the smallest distances are then selected deterministically to form the training batch. Following Sun et al. (2020), importance sampling is disabled.
+### E.1. Architectural and Implementation Details
+MiniGrid agents (DQN / IQN). Both DQN and IQN are trained on symbolic Minigrid observations $s_t \in \{0, \dots\}^{N \times N \times 3}$ , which encode object identities, colors, and state information such as door status and agent orientation. To process these inputs, we employ a shared encoder which first embeds each channel, then applies a residual dilated CNN with three blocks. The resulting features are aggregated using both average and max global pooling, yielding a 256-dimensional representation. This design aims to capture both local and global spatial structure in the environment. For DQN, the 256-dimensional feature vector is passed through a small MLP, which outputs scalar Q-values for each discrete action. In IQN, we replace the scalar output head with an implicit quantile head. Specifically, we embed sampled quantile fractions using cosine functions, project these embeddings to match the feature dimension, and combine them element-wise with the state features. The resulting representations are mapped to per-action quantile values. We optimize using the quantile Huber loss and select actions using either Double-DQN-style (Van Hasselt et al., 2016) or target-network-based strategies, as described in Appendix E. This approach allows IQN to model the full distribution over returns, rather than just the mean.
+OGBench Scene agents (SAC / TD3). For OGBench/Scene, both SAC and TD3 operate on flattened state-based observations $s_t \in \mathbb{R}^{40}$ . We adopt a unified architecture across tasks: the actor is a fully-connected network with three hidden layers of width 512 and ReLU activations. This network outputs either a tanh-squashed Gaussian policy (for SAC) or a deterministic tanh policy (for TD3), enabling flexible action selection in continuous spaces. The critic is implemented as an ensemble of N=10 Q-functions, each parameterized by a 3-layer MLP that receives the concatenated state and action as input and outputs a scalar value. For target computation, we randomly select M=2 ensemble members and aggregate their predictions using either a mean (in all main SAC/TD3 runs) or a min operator. This ensemble approach is intended to improve stability and reduce overestimation bias. For actor updates, we use the same Q-function ensemble, aggregating Q-values according to the actor reduction rule. In all reported experiments, we use the mean over ensemble members. This ensures consistency between actor and critic updates.
+Replay buffer and prioritization. All methods are implemented with a two-branch replay ensemble: one prioritized branch and one uniform branch which share an underlying storage, implemented using ReplayBufferEnsemble (Bou et al., 2023) with mixture weights controlled by the sampling ratio schedule in Tables 2–3. For the PER baseline, we assign priorities using the standard PER formula $p_i = |\delta_i| + \epsilon$ , where $\delta_i$ is the TD error. When importance-sampling corrections are enabled, we apply them as usual. New transitions are initially assigned a small default priority, and we update their TD-errors after each gradient step to ensure accurate prioritization. In VLM-RB, we augment each transition with both a TD-based metric and a VLM-based score. For MiniGrid/DoorKey, replay priorities are determined solely by the VLM score. In contrast, for OGBench/Scene, we set priorities as the product of the VLM score and the TD-based metric (see Section 3.3). This design allows us to tailor prioritization to the characteristics of each environment.
+**VLM worker.** The VLM worker uses Perception-LM-1B (Cho et al., 2025) with a fixed prompt to detect binary success. For each transition, it processes a clip of L=32 frames, applying left padding if the clip is shorter. To compute a scalar priority, we sum the probabilities assigned to all "Yes" and "No" token variants in the first generated token's logits, and calculate their ratio to obtain a probability-style score. We hard-threshold these scores at 0.5, assigning 1 or 0 accordingly. The resulting priorities are asynchronously streamed back to update the replay buffer, ensuring the main learner is not blocked during this process.
+{18}------------------------------------------------
+<span id="page-18-0"></span>*Table 2.* Hyperparameters and implementation details for OGBench experiments. Shared parameters are listed once; divergent parameters are compared side-by-side.
+| Hyperparameter | TD3 | SAC | |
+|---------------------------------------------|---------------------------|-----------------|--|
+| Network Architecture | | | |
+| Hidden Dimensions | | [512, 512, 512] | |
+| Q-Network Layer Norm | False | True | |
+| Optimization | | | |
+| Critic Learning Rate | | 3 × 10−4 | |
+| Actor Learning Rate | 3 × 10−4 | 1 × 10−3 | |
+| Alpha Learning Rate | – | 3 × 10−4 | |
+| Batch Size | | 256 | |
+| Discount Factor (γ) | | 0.99 | |
+| Target Smoothing (τ ) | | 0.005 | |
+| Max Grad Norm | | 10.0 | |
+| Learning Starts | | 10,000 steps | |
+| Ensemble & Update Ratios | | | |
+| Ensemble Size (N) | | 10 | |
+| Subsampled Q-Networks (M) | | 2 | |
+| Critic UTD Ratio | | 4 | |
+| Actor UTD Ratio | | 2 | |
+| Target Q Reduction | min | mean | |
+| Actor Q Reduction | min | mean | |
+| Data & Replay Buffer | | | |
+| Replay Buffer Size | | 1 × 106 | |
+| Expert Demos | | 10 | |
+| Algorithm Specifics | | | |
+| Policy Update Frequency | 2 | 1 | |
+| Exploration Noise (std) | 0.1 | – | |
+| Target Policy Noise (std) | 0.2 | – | |
+| Noise Clip | 0.5 | – | |
+| Entropy Coeff. (α) | – | 1.0 (Initial) | |
+| Target Entropy Scale | – | 0.5 | |
+| Auto Entropy Tuning | – | True | |
+| PER Baseline Settings | | | |
+| PER Alpha (α) | 0.7 | | |
+| PER Beta (β) | 1.0 | | |
+| Importance Sampling | False | | |
+| VLM-RB Settings | | | |
+| VLM Model | Facebook Perception-LM-1B | | |
+| Priority Mode | Hard Threshold (> 0.5) | | |
+| Trajectory Length (L) | 32 Frames | | |
+| Importance Sampling | False | | |
+| Prioritized Sampling Ratio (λ0<br>and λmax) | Annealed 0.0 → 0.5 | | |
+| Annealing Schedule (Tschedule) | Linear over 500k steps | | |
+{19}------------------------------------------------
+<span id="page-19-0"></span>*Table 3.* Hyperparameters and implementation details for MiniGrid experiments. Shared parameters are listed once; divergent parameters are compared side-by-side.
+| Hyperparameter | DQN | IQN |
+|---------------------------------------------|---------------------------|-----------------------------|
+| Optimization | | |
+| Learning Rate | 4 × 10−5 | |
+| Batch Size | | 128 |
+| Discount Factor (γ) | 0.95 | |
+| Target Update Frequency | | 1,000 steps |
+| Target Update Rate (τ ) | | 1.0 (Hard Update) |
+| Learning Starts | | 500 steps |
+| Train Frequency | | 4 steps |
+| Max Grad Norm | | 1.0 |
+| Exploration (Epsilon-Greedy) | | |
+| epsstart | | 1.0 |
+| epsend | | 0.05 |
+| Exploration Fraction | | 0.5 (First 50% of training) |
+| Algorithm Specifics | | |
+| Double DQN | | Enabled |
+| Noisy Nets | – | Disabled |
+| Num Quantiles (Policy) | – | 32 |
+| Num Quantiles (Train/Target) | – | 64 |
+| Num Cosine Basis Functions | – | 64 |
+| Huber Kappa (κ) | – | 1.0 |
+| PER Baseline Settings | | |
+| PER Alpha (α) | | 0.7 |
+| PER Beta (β) | 1.0 | |
+| Importance Sampling | True | |
+| Prioritized Sampling Ratio | 1.0 (Always Prioritized) | |
+| VLM-RB Settings | | |
+| VLM Model | Facebook Perception-LM-1B | |
+| Priority Mode | Hard Threshold (> 0.5) | |
+| Trajectory Length (L) | 32 Frames | |
+| Importance Sampling | False | |
+| Prioritized Sampling Ratio (λ0<br>and λmax) | Annealed 0.0 → 0.5 | |
+| Annealing Schedule (Tschedule) | Linear over 500k steps | |
+{20}------------------------------------------------
+## <span id="page-20-1"></span>F. Ablations
+### F.1. Mixing Schedule
+<span id="page-20-0"></span>*Table 4.* Ablation of mixing schedule (λmax) on DoorKey-16x16. Performance denotes the highest Average Success Rate (M = 5 seeds). Sample Efficiency tracks the steps required to reach the *baseline's* best performance.
+| | | Improvement | | | |
+|------|-----------------|------------------------------------|----------------------------------------------------------------------------------------------|--|--|
+| λmax | Base. | Perf. (↑) | Sample Eff. (↓) | | |
+| | None UER<br>PER | N/A (0.00/0.24)<br>N/A (0.00/0.80) | N/A (1000K/916K)<br>N/A (1000K/592K) | | |
+| | PER | +0.0% (0.80/0.80) | 0.25 UER +233.3% (0.80/0.24) +40.6% (544K/916K)<br>-0.3% (594K/592K) | | |
+| | | | 0.50 UER +316.7% (1.00/0.24) +62.9% (340K/916K)<br>PER +25.0% (1.00/0.80) +35.5% (382K/592K) | | |
+| | | | 0.75 UER +266.7% (0.88/0.24) +57.0% (394K/916K)<br>PER +10.0% (0.88/0.80) +17.2% (490K/592K) | | |
+| | | | 1.00 UER +316.7% (1.00/0.24) +69.0% (284K/916K)<br>PER +25.0% (1.00/0.80) +40.2% (354K/592K) | | |
+To understand the effect of the final mixing coefficient λmax, we conduct an ablation on the MiniGrid/DoorKey-16x16 task (Fig[.6\)](#page-20-2). We fix Tschedule = 5 · 10<sup>5</sup> , ensuring that λ<sup>t</sup> is annealed linearly from 0 (corresponding to purely uniform sampling) to λmax over the first half of training (right panel). We sweep λmax ∈ {0.25, 0.5, 0.75, 1.0} and include a **None** baseline, which disables the schedule and relies entirely on VLM-prioritized sampling. We observe that larger λmax values (0.75, 1.0) reach 50% success marginally earlier than smaller values, but their final success rates remain lower within the fixed training budget. In contrast, λmax = 0.5 emerges as the most reliable choice in this environment: it consistently achieves 100% success within the budget and attains the highest performance at the 90% success threshold among all options considered. The smallest value (λmax = 0.25) results in slower learning, and the fully prioritized variant (**None**) fails to solve the task under our setup, with success rates remaining near zero. This suggests that maintaining a non-trivial fraction of uniform sampling is essential for effective learning in this setting. Based on these findings, we fix λmax = 0.5
+![](_page_20_Figure_7.jpeg)
+for all main experiments without further per-task tuning.
+<span id="page-20-2"></span>![](_page_20_Figure_8.jpeg)
+*Figure 6.* MiniGrid/DoorKey-16x16: λmax Ablation (5 seeds), "None" corresponds to no scheduling, i.e., only VLM-prioritized sampling. Left: performance for different λmax values. Right: λmax value evolution.
+{21}------------------------------------------------
+### <span id="page-21-0"></span>F.2. VLM Size
+<span id="page-21-2"></span>*Table 5.* Performance comparison of Perception-LM [\(Cho et al.,](#page-8-8) [2025\)](#page-8-8) models. Values in parentheses denote the relative percentage change compared to 1B. We use an Nvidia RTX 4090 GPU, and run 100 batches of 32 frames.
+| Model | Load (GiB) | Peak (GiB) | Time (s) | FPS |
+|-------|---------------|---------------|--------------|--------------|
+| 1B | 2.86 | 3.77 | 0.46 | 69.27 |
+| 3B | 6.56 (+130%) | 8.16 (+116%) | 0.77 (+66%) | 41.75 (-40%) |
+| 8B | 18.25 (+539%) | 20.34 (+439%) | 2.15 (+366%) | 14.88 (-79%) |
+We investigate how scaling the VLM affects both inference overhead and downstream RL performance. Table [5](#page-21-2) quantifies the resource requirements of Perception-LM [\(Cho et al.,](#page-8-8) [2025\)](#page-8-8) variants for clip scoring (Nvidia RTX 4090 GPU; 100 batches of 32-frame clips). We observe that increasing the VLM size leads to a substantial increase in memory footprint and a corresponding reduction in throughput. Relative to the 1B model, the 3B variant increases peak memory by +116% and reduces FPS by 40%; the 8B variant increases peak memory by +439% and reduces FPS by 79%.
+<span id="page-21-3"></span>To assess the impact of VLM size on downstream performance, we compare the 1B, 3B, and 8B models on MiniGrid/DoorKey-16x16 (Fig. [7\)](#page-21-3). Notably, despite the increased inference cost of larger models, we do not observe consistent improvements in RL performance relative to the 1B configuration. These results suggest that in this setting, the clip-scoring signal saturates once the VLM is sufficiently reliable at separating task-relevant from irrelevant segments. Given this trade-off, we select the 1B Perception-LM as the default backbone for all main experiments.
+![](_page_21_Figure_6.jpeg)
+*Figure 7.* MiniGrid/DoorKey-16x16: VLM Size Ablation (5 seeds)
+### <span id="page-21-1"></span>F.3. VLM Prior
+Motivated by the "Modified Game" paradigm of [Dubey et al.](#page-8-13) [\(2018\)](#page-8-13), we probe whether VLM-RB leverages the VLM's pre-trained visual semantics to improve performance. To isolate the effect of visual semantics, we modify only the rendered frames provided to the VLM for scoring, while leaving both the underlying MDP and the agent's observations unchanged. If VLM-RB relies on semantic cues such as identifying keys, doors, or goal-relevant interactions, we expect its performance gains to diminish when these cues are distorted or removed.
+To test this, we introduce two renderer perturbations (Fig. [8\)](#page-22-1). First, Sprite Swap replaces object sprites with semantically conflicting alternatives, such as rendering keys as lava or doors as boxes, thereby introducing misleading visual priors. Second, Texture replaces all object appearances with abstract high-contrast patterns, removing naturalistic semantics entirely.
+Empirically, we observe that the unmodified setting achieves near-perfect success and converges the fastest (Fig. [4\)](#page-5-1). In contrast, both Sprite Swap and Texture slow learning and reduce final success rates, with Texture also leading to the largest variance across seeds. Because the control problem and agent inputs remain fixed, this degradation suggests that the VLM produces less informative priorities when visual evidence for goal-relevant events is either misleading or lacks
+{22}------------------------------------------------
+<span id="page-22-1"></span>semantic content.
+![](_page_22_Figure_2.jpeg)
+![](_page_22_Figure_3.jpeg)
+![](_page_22_Figure_4.jpeg)
+![](_page_22_Figure_5.jpeg)
+(c) Abstract patterns that remove naturalistic cues
+*Figure 8.* Samples of the modified visuals. We modify only the frames passed to the VLM for scoring (agent observations and environment dynamics unchanged).
+misleading sprites
+### <span id="page-22-2"></span><span id="page-22-0"></span>F.4. Computational Overhead
+*Table 6.* Training Throughput (steps/second) comparison on MiniGrid/DoorKey-16x16. Higher is better.
+| Hardware | PER | VLM-RB | Rel. Speed |
+|--------------|-----|--------|------------|
+| NVIDIA A100 | 111 | 97 | 87% |
+| NVIDIA A40 | 92 | 81 | 88% |
+| NVIDIA A4000 | 76 | 67 | 88% |
+How much does VLM-RB actually slow down training in practice? To answer this, we measure throughput (steps per second) on the MiniGrid/DoorKey-16x16 task using DQN, comparing PER and VLM-RB across three dual-GPU setups (NVIDIA A100, A40, and A4000). In each case, the RL learner and VLM are placed on separate devices. Notably, in this distributed configuration, the main bottleneck is inter-process communication (IPC) and data transfer, rather than competition for computational resources.
+The results, summarized in Table [6,](#page-22-2) reveal a consistent and modest throughput reduction of about 12% across all hardware types. This suggests that, even when VLM inference is separated from the RL learner, the method maintains efficient scaling. In other words, the additional overhead introduced by VLM-RB remains limited in practical distributed settings.
+It is important to note that these measurements use a standard inference setup, without any aggressive optimizations such as TensorRT or quantization (e.g., INT8 or FP4). We anticipate that employing a dedicated serving stack, such as vLLM or TGI, would further close the speed gap between VLM-RB and the baseline.
+{23}------------------------------------------------
+## <span id="page-23-1"></span>G. Experiments
+In this section, we present the full results of the baselines on all environments.
+<span id="page-23-0"></span>*Table 7.* Performance denotes the highest Average Success Rate (M = 5 seeds). Sample Efficiency tracks the steps required to reach the *baseline's* best performance. Wall-Clock Saving estimates the reduction in real training time, accounting for a 12% inference overhead per step.
+| Alg. | Task | Baseline | Performance (↑)<br>Best ASR | Sample Efficiency (↓)<br>Steps to Base. Best | Wall-Clock Saving (↑)<br>Time vs. Baseline |
+|------|---------------|------------|--------------------------------------------|----------------------------------------------|--------------------------------------------|
+| | Scene-3 | UER<br>PER | +0.0% (1.00/1.00)<br>+0.0% (1.00/1.00) | +54.6% (206K/454K)<br>+25.4% (206K/276K) | +49.2%<br>+16.4% |
+| SAC | Scene-4 | UER<br>PER | +4.2% (1.00/0.96)<br>+4.2% (1.00/0.96) | +28.3% (592K/826K)<br>+26.0% (592K/800K) | +19.7%<br>+17.1% |
+| | Scene-5 | UER<br>PER | +100.0% (0.88/0.44)<br>+41.9% (0.88/0.62) | +43.8% (514K/914K)<br>+6.3% (708K/756K) | +37.0%<br>-4.8% |
+| | Scene-3 | UER<br>PER | +0.0% (1.00/1.00)<br>+0.0% (1.00/1.00) | +15.7% (214K/254K)<br>+16.4% (214K/256K) | +5.6%<br>+6.4% |
+| TD3 | Scene-4 | UER<br>PER | +47.1% (1.00/0.68)<br>+0.0% (1.00/1.00) | +63.1% (268K/726K)<br>+9.0% (426K/468K) | +58.7%<br>-1.9% |
+| | Scene-5 | UER<br>PER | +150.0% (0.70/0.28)<br>+59.1% (0.70/0.44) | +49.4% (366K/724K)<br>+28.1% (614K/854K) | +43.4%<br>+19.5% |
+| | DoorKey-8x8 | UER<br>PER | +0.0% (1.00/1.00)<br>+0.0% (1.00/1.00) | +10.7% (150K/168K)<br>+27.9% (150K/208K) | +0.0%<br>+19.2% |
+| DQN | DoorKey-12x12 | UER<br>PER | +66.7% (1.00/0.60)<br>+0.0% (1.00/1.00) | +62.6% (216K/578K)<br>+6.1% (246K/262K) | +58.1%<br>-5.2% |
+| | DoorKey-16x16 | UER<br>PER | +316.7% (1.00/0.24)<br>+25.0% (1.00/0.80) | +62.9% (340K/916K)<br>+35.5% (382K/592K) | +58.4%<br>+27.7% |
+| | DoorKey-8x8 | UER<br>PER | +0.0% (1.00/1.00)<br>+0.0% (1.00/1.00) | +26.6% (138K/188K)<br>+16.9% (138K/166K) | +17.8%<br>+6.9% |
+| IQN | DoorKey-12x12 | UER<br>PER | +56.2% (1.00/0.64)<br>+56.2% (1.00/0.64) | +44.7% (388K/702K)<br>+42.3% (388K/672K) | +38.1%<br>+35.3% |
+| | DoorKey-16x16 | UER<br>PER | +166.7% (0.64/0.24)<br>+300.0% (0.64/0.16) | +11.0% (774K/870K)<br>+13.8% (562K/652K) | +0.4%<br>+3.5% |
+*Table 8.* Performance denotes the highest Average Success Rate (M = 5 seeds). Sample Efficiency tracks the steps required to reach the *baseline's* best performance. Both metrics are averaged across the aggregated algorithms and tasks.
+| Env Type | Agg. Algorithms | Baseline | Performance (↑)<br>Mean Best ASR | Sample Efficiency (↓)<br>Mean Steps to Base Peak |
+|----------|-----------------|------------|------------------------------------------|--------------------------------------------------|
+| Scene | (SAC + TD3) | UER<br>PER | +28.0% (0.93/0.73)<br>+11.2% (0.93/0.84) | +44.6% (360K/650K)<br>+19.1% (460K/568K) |
+| DoorKey | (DQN + IQN) | UER<br>PER | +51.6% (0.94/0.62)<br>+22.6% (0.94/0.77) | +41.4% (334K/570K)<br>+26.9% (311K/425K) |
+{24}------------------------------------------------
+<span id="page-24-0"></span>![](_page_24_Figure_1.jpeg)
+*Figure 9.* VLM-RB consistently outperforms baselines across continuous and discrete tasks. The plots show aggregated success rates for four algorithms (DQN, IQN, SAC, TD3) on MiniGrid and OGBench domains. Annotations highlight the relative improvement in sample efficiency (horizontal arrows, reaching peak performance faster) and the best success rate (vertical arrows). Shaded regions indicate standard deviation across seeds.