Daily Papers

Recurrent fast-weight memories and selective state-space models compress an expanding context into a fixed-size recurrent state, making the state transition an online learning rule. We study this rule under read-after-write autoregressive semantics. For the prefix-prediction objective considered here, the local fast-memory example revealed at step t is the prefix-aligned pair (x_t,y_t)=(ϕ(k_{t-1}),v_t). The common same-step association (ϕ(k_t),v_t) remains causal, but optimizes a different internal objective. We derive normalized first-order updates for squared-error regression and negative inner-product objectives. The regression family comprises Falcon-1 (a scalar NLMS update), Falcon-2 (its per-column extension), and Falcon-3 (a sliding-window mini-batch update); Falcon-1A/Falcon-2A/Falcon-3A are the corresponding inner-product variants. We provide recurrent, masked-parallel, and chunk-parallel forms, together with numerically stable positive-decay renormalization. Representative variants remain competitive in language modeling and improve length extrapolation on variable-digit addition. This framework separates temporal alignment, plasticity, forgetting, and bounded rehearsal in recurrent sequence models.

Joint-Embedding Predictive Architectures (JEPAs), including recent LeWorldModel (LeWM), have become a promising foundation for reconstruction-free visual world models. For visual planning, however, LeWM evaluates candidate action sequences by repeatedly applying a local one-step latent transition model. This autoregressive rollout makes planning computationally expensive and exposes the predicted trajectory to accumulated latent errors as the horizon grows. We propose Fast LeWorldModel (Fast-LeWM), a fast latent world model that replaces repeated local rollout with action-prefix prediction. Given the current latent and a candidate action sequence, Fast-LeWM encodes its prefixes and predicts the future latents reached after executing those prefixes in parallel. By making action prefixes the basic prediction unit, Fast-LeWM directly models action effects accumulated to different extents over multiple horizons. This prefix-level supervision forces the model to learn how states continuously evolve under different action prefixes, rather than only fitting one-step state transitions. During planning, the predictor can use the last prefix token from the encoded action sequence to evaluate the corresponding future latent without explicitly rolling through each intermediate imagined state. Across multiple tasks, Fast-LeWM improves average success over LeWM while substantially reducing planning time, achieving lower open-loop latent loss whose growth becomes significantly slower as the rollout horizon increases.

Using a diffusion model for parallel drafting is a promising approach for speculative decoding. By predicting tokens at multiple future positions in a single forward pass, diffusion drafters substantially reduce drafting latency. However, this shifts the bottleneck to verification: verifying a single sequence limits acceptance length, while verifying large draft trees incurs excessive target-model latency. We identify a key mismatch in existing draft-tree methods: existing diffusion-tree methods rank nodes by the marginal probability, ignoring that verification is prefix-conditioned. As a result, they may verify unreachable descendants of rejected prefixes, increasing latency with limited acceptance gains. To address this, we propose TAPS, a target-aware prefix selection method that turns diffusion marginals into path-conditioned acceptance estimates. TAPS then selects a compact prefix-closed subtree under a fixed verification budget, improving the acceptance-cost tradeoff rather than simply expanding the draft tree. Experiments across diverse datasets and model families demonstrate that TAPS achieves up to 7.9x lossless end-to-end speedup over vanilla autoregressive decoding, outperforming state-of-the-art DFlash and DDTree by 1.36x and 1.74x respectively. Our work is available at https://anonymous.4open.science/r/TAPS-EMNLP2026-53DD

Large Language Models (LLMs) are increasingly evaluated on multiple-choice question answering (MCQA) tasks using *first-token probability* (FTP), which selects the answer option whose initial token has the highest likelihood. While efficient, FTP can be fragile: models may assign high probability to unrelated tokens (*misalignment*) or use a valid token merely as part of a generic preamble rather than as a clear answer choice (*misinterpretation*), undermining the reliability of symbolic evaluation. We propose a simple solution: the *prefilling attack*, a structured natural-language prefix (e.g., "*The correct option is:*") prepended to the model output. Originally explored in AI safety, we repurpose prefilling to steer the model to respond with a clean, valid option, without modifying its parameters. Empirically, the FTP with prefilling strategy substantially improves accuracy, calibration, and output consistency across a broad set of LLMs and MCQA benchmarks. It outperforms standard FTP and often matches the performance of open-ended generation approaches that require full decoding and external classifiers, while being significantly more efficient. Our findings suggest that prefilling is a simple, robust, and low-cost method to enhance the reliability of FTP-based evaluation in multiple-choice settings.

Recent think-answer approaches in VLMs, such as Qwen3-VL-Thinking, boost reasoning performance by leveraging intermediate thinking steps before the final answer, but their high computational cost limits real-world deployment. To distill such capabilities into compact think-answer VLMs, a primary objective is to improve the student's ability to utilize visual evidence throughout its reasoning trace. To this end, we introduce a novel think-answer distillation framework that encourages the student to anchor its thinking on visual information by masking the student's salient reasoning prefixes. To compensate for such masked textual cues, the student is encouraged to rely more on visual evidence as an alternative source of information during distillation. Our masking strategies include: 1) token-wise salient reasoning-prefix masking, which masks high-influence reasoning prefixes selectively for each next-token prediction, and 2) self-paced masking budget scheduling, which gradually increases the masking scale according to distillation difficulty, {measured by discrepancy between teacher--student distributions. In the distillation phase, the student is guided by our salient reasoning-prefix mask, which blocks both future tokens and salient reasoning cues, in place of the standard causal mask used for auto-regressive language modeling. Experimental results show that our approach outperforms recent open-source VLMs, VLM distillation, and self-distillation methods on multimodal reasoning benchmarks, while further analyses confirm enhanced visual utilization along the student thinking process.

Recently, large vision-language models (LVLMs) have rapidly gained popularity for their strong generation and reasoning capabilities given diverse multimodal inputs. However, these models incur significant computational and memory overhead during inference, which greatly hinders the efficient deployment in practical scenarios. The extensive key-value (KV) cache, necessitated by the lengthy input and output sequences, notably contributes to the high inference cost. Based on this, recent works have investigated ways to reduce the KV cache size for higher efficiency. Although effective, they generally overlook the distinct importance distributions of KV vectors across layers and maintain the same cache size for each layer during the next token prediction. This results in the significant contextual information loss for certain layers, leading to notable performance decline. To address this, we present PrefixKV. It reframes the challenge of determining KV cache sizes for all layers into the task of searching for the optimal global prefix configuration. With an adaptive layer-wise KV retention recipe based on binary search, the maximum contextual information can thus be preserved in each layer, facilitating the generation. Extensive experiments demonstrate that our method achieves the state-of-the-art performance compared with others. It exhibits superior inference efficiency and generation quality trade-offs, showing promising potential for practical applications. Code is available at https://github.com/THU-MIG/PrefixKV.

LLM-based multi-agent systems are increasingly deployed on long-horizon tasks, but a single decisive error is often accepted by downstream agents and cascades into trajectory-level failure. Existing work frames this as post-hoc failure attribution, diagnosing the responsible agent and step after the trajectory has ended. However, this paradigm forfeits any opportunity to intervene while trajectory is still unfolding. In this work, we introduce AgentForesight, a framework that reframes this problem as online auditing: at each step of an unfolding trajectory, an auditor observes only the current prefix and must either continue the run or alarm at the earliest decisive error, without access to future steps. To this end, we curate AFTraj-2K, a corpus of agentic trajectories across Coding, Math, and Agentic domains, in which safe trajectories are retained under a strict curation pipeline and unsafe trajectories are annotated at the step of their decisive error via consensus among multiple LLM judges. Built on that, we develop AgentForesight-7B, a compact online auditor trained with a coarse-to-fine reinforcement learning recipe that first equips it with a risk-anticipation prior at the failure boundary on adjacent safe/unsafe prefix pairs, then sharpens this prior into precise step-level localization under a three-axis reward jointly targeting the what, where, and who of an audit verdict. Across AFTraj-2K and an external Who\&When benchmark, AgentForesight-7B outperforms leading proprietary models, including GPT-4.1 and DeepSeek-V4-Pro, achieving up to +19.9% performance gain and 3times lower step localization error, opening the loop from post-hoc failures detection to enabling deployment-time intervention. Project page: https://zbox1005.github.io/agent-foresight/

Autoregressive language models are constrained by their inherently sequential nature, generating one token at a time. This paradigm limits inference speed and parallelism, especially during later stages of generation when the direction and semantics of text are relatively certain. In this work, we propose a novel framework that leverages the inherent knowledge of vanilla autoregressive language models about future tokens, combining techniques to realize this potential and enable simultaneous prediction of multiple subsequent tokens. Our approach introduces several key innovations: (1) a masked-input formulation where multiple future tokens are jointly predicted from a common prefix; (2) a gated LoRA formulation that preserves the original LLM's functionality, while equipping it for multi-token prediction; (3) a lightweight, learnable sampler module that generates coherent sequences from the predicted future tokens; (4) a set of auxiliary training losses, including a consistency loss, to enhance the coherence and accuracy of jointly generated tokens; and (5) a speculative generation strategy that expands tokens quadratically in the future while maintaining high fidelity. Our method achieves significant speedups through supervised fine-tuning on pretrained models. For example, it generates code and math nearly 5x faster, and improves general chat and knowledge tasks by almost 2.5x. These gains come without any loss in quality.

Coding agents are a fast-growing LLM application, executing as long-running closed-loop sessions in which LLM generations alternate with external tool calls. Yet, unlike chat workloads, their serving behavior has not been studied extensively. We address this gap by collecting a dataset of real-world coding assistant traces. Our analysis shows that coding agent sessions repeatedly reuse large prefixes and create sustained KVCache pressure that conventional LLM serving policies handle poorly. Based on our analysis, we present CacheWise, a KVCache management layer that improves KVCache reuse for coding agent workloads. CacheWise combines prefix-aware scheduling with reuse-aware eviction guided by lightweight predictions from tool call metadata. Implemented in vLLM and evaluated on the collected traces, CacheWise reduces KVCache evictions by up to 2-2.6x and improves total agent session completion time by up to 3.5x.

Generating diverse responses is crucial for test-time scaling of large language models (LLMs), yet standard stochastic sampling mostly yields surface-level lexical variation, limiting semantic exploration. In this paper, we propose Exploratory Sampling (ESamp), a decoding approach that explicitly encourages semantic diversity during generation. ESamp is motivated by the well-known observation that neural networks tend to make lower-error predictions on inputs similar to those encountered before, and incur higher prediction error on novel ones. Building on this property, we train a lightweight Distiller at test time to predict deep-layer hidden representations of the LLM from its shallow-layer representations to model the LLM's depth-wise representation transitions. During decoding, the Distiller continuously adapts to the mappings induced by the current generation context. ESamp uses the prediction error as a novelty signal to reweight candidate token extensions conditioned on the current prefix, thereby biasing decoding toward less-explored semantic patterns. ESamp is implemented with an asynchronous training--inference pipeline, with less than 5% worst case overhead (1.2% in the optimized release). Empirical results show that ESamp significantly boosts the Pass@k efficiency of reasoning models, showing superior or comparable performance to strong stochastic and heuristic baselines. Notably, ESamp achieves robust generalization across mathematics, science, and code generation benchmarks and breaks the trade-off between diversity and coherence in creative writing. Our code has released at: https://github.com/LinesHogan/tLLM.

Membership inference attacks serves as useful tool for fair use of language models, such as detecting potential copyright infringement and auditing data leakage. However, many current state-of-the-art attacks require access to models' hidden states or probability distribution, which prevents investigation into more widely-used, API-access only models like GPT-4. In this work, we introduce N-Gram Coverage Attack, a membership inference attack that relies solely on text outputs from the target model, enabling attacks on completely black-box models. We leverage the observation that models are more likely to memorize and subsequently generate text patterns that were commonly observed in their training data. Specifically, to make a prediction on a candidate member, N-Gram Coverage Attack first obtains multiple model generations conditioned on a prefix of the candidate. It then uses n-gram overlap metrics to compute and aggregate the similarities of these outputs with the ground truth suffix; high similarities indicate likely membership. We first demonstrate on a diverse set of existing benchmarks that N-Gram Coverage Attack outperforms other black-box methods while also impressively achieving comparable or even better performance to state-of-the-art white-box attacks - despite having access to only text outputs. Interestingly, we find that the success rate of our method scales with the attack compute budget - as we increase the number of sequences generated from the target model conditioned on the prefix, attack performance tends to improve. Having verified the accuracy of our method, we use it to investigate previously unstudied closed OpenAI models on multiple domains. We find that more recent models, such as GPT-4o, exhibit increased robustness to membership inference, suggesting an evolving trend toward improved privacy protections.

On-policy distillation (OPD) leverages dense teacher rewards to enhance reasoning models. However, scaling OPD to long-horizon tasks exposes a critical flaw: as the student's generated prefix inevitably diverges from the teacher's thought process, the teacher's dense reward loses local exploitability. Continuing to generate and evaluate tokens on these ``drifted'' trajectories not only degrades reward quality but also incurs massive computational waste. To address this, we introduce Prune-OPD, a framework that dynamically aligns training budgets with supervision quality. By continuously monitoring the local compatibility between student and teacher predictions (e.g., via top-k overlap), Prune-OPD detects prefix-drift events in real time. Upon detecting severe drift, it monotonically down-weights subsequent unreliable rewards and triggers dynamic rollout truncation. This allows the training process to halt futile generation and reallocate compute strictly to reliable teacher supervision. Across diverse teacher-student combinations, Prune-OPD consistently aligns computation with supervision reliability. When prefix drift makes dense teacher rewards unreliable, it reduces training time by 37.6\%--68.0\% while preserving, and often improving, performance on challenging benchmarks (AMC, AIME, HMMT). When student-teacher compatibility remains high, it automatically preserves long-context supervision by expanding the training window. These results suggest that Prune-OPD improves OPD not by blindly shortening rollouts, but by reallocating computation toward locally exploitable teacher rewards.

Deep autoregressive sequence-to-sequence models have demonstrated impressive performance across a wide variety of tasks in recent years. While common architecture classes such as recurrent, convolutional, and self-attention networks make different trade-offs between the amount of computation needed per layer and the length of the critical path at training time, generation still remains an inherently sequential process. To overcome this limitation, we propose a novel blockwise parallel decoding scheme in which we make predictions for multiple time steps in parallel then back off to the longest prefix validated by a scoring model. This allows for substantial theoretical improvements in generation speed when applied to architectures that can process output sequences in parallel. We verify our approach empirically through a series of experiments using state-of-the-art self-attention models for machine translation and image super-resolution, achieving iteration reductions of up to 2x over a baseline greedy decoder with no loss in quality, or up to 7x in exchange for a slight decrease in performance. In terms of wall-clock time, our fastest models exhibit real-time speedups of up to 4x over standard greedy decoding.

End-to-end autonomous driving via Vision-Language-Action (VLA) models demands a precarious balance between high-fidelity trajectory planning and efficient inference. Existing paradigms typically fall short: autoregressive (AR) VLAs are memory-bandwidth-bound on edge hardware and prone to exposure-bias drift, while full-sequence diffusion models preclude KV-cache reuse and suffer from "logical leakage" that violates the fundamental perceive-then-plan causality. We present Fast-dDrive, a block-diffusion VLA that performs bidirectional refinement within semantic units while enforcing strict causal ordering across them. Leveraging the observation that driving VLAs often emit structured JSON-like outputs, Fast-dDrive freezes structural tokens into a section scaffold and employs a section-aware training recipe that prioritizes safety-critical planning. We further introduce Scaffold Speculative Decoding to achieve AR-equivalent quality at significantly higher throughput. Finally, we propose a low-overhead test-time scaling scheme: by forking N stochastic trajectory rollouts from a single shared-prefix KV cache and averaging them, we effectively suppress prediction variance at a fractional computational cost. Empirical results demonstrate that Fast-dDrive redefines the speed-accuracy frontier for driving agents. On the WOD-E2E test set, Fast-dDrive achieves SOTA ADE@3s and ADE@5s, alongside the highest RFS among diffusion-based VLAs; on nuScenes, it reduces average L2 error to 0.32m (a 22% improvement). When integrated with SGLang, our framework delivers 12times throughput speedup over the AR baseline, narrowing the gap between high-capacity VLAs and the efficiency demands of real-time on-vehicle deployment.

Speculative decoding has emerged as a powerful approach to accelerate large language model (LLM) inference by employing lightweight draft models to propose candidate tokens that are subsequently verified by the target model. The effectiveness of this paradigm critically depends on the quality of the draft model. While recent advances such as the EAGLE series achieve state-of-the-art speedup, existing draft models remain limited by error accumulation: they condition only on the current prefix, causing their predictions to drift from the target model over steps. In this work, we propose ConFu (Contemplate the Future), a novel speculative decoding framework that enables draft models to anticipate the future direction of generation. ConFu introduces (i) contemplate tokens and soft prompts that allow the draft model to leverage future-oriented signals from the target model at negligible cost, (ii) a dynamic contemplate token mechanism with MoE to enable context-aware future prediction, and (iii) a training framework with anchor token sampling and future prediction replication that learns robust future prediction. Experiments demonstrate that ConFu improves token acceptance rates and generation speed over EAGLE-3 by 8--11% across various downstream tasks with Llama-3 3B and 8B models. We believe our work is the first to bridge speculative decoding with continuous reasoning tokens, offering a new direction for accelerating LLM inference.

Large language models (LLMs) have demonstrated remarkable capabilities in tasks requiring reasoning and multi-step problem-solving through the use of chain-of-thought (CoT) prompting. However, generating the full CoT process results in significantly longer output sequences, leading to increased computational costs and latency during inference. To address this challenge, we propose a novel approach to compress the CoT process through semantic alignment, enabling more efficient decoding while preserving the benefits of CoT reasoning. Our method introduces an auxiliary CoT model that learns to generate and compress the full thought process into a compact special token representation semantically aligned with the original CoT output. This compressed representation is then integrated into the input of the Hidden Chain-of-Thought (HCoT) model. The training process follows a two-stage procedure: First, the CoT model is optimized to generate the compressed token representations aligned with the ground-truth CoT outputs using a contrastive loss. Subsequently, with the CoT model parameters frozen, the HCoT model is fine-tuned to generate accurate subsequent predictions conditioned on the prefix instruction and the compressed CoT representations from the CoT model. Extensive experiments across three challenging domains - mathematical reasoning, agent invocation, and question answering - demonstrate that our semantic compression approach achieves competitive or improved performance compared to the full CoT baseline, while providing significant speedups of at least 1.5x in decoding time. Moreover, incorporating contrastive learning objectives further enhances the quality of the compressed representations, leading to better CoT prompting and improved task accuracy. Our work paves the way for more efficient exploitation of multi-step reasoning capabilities in LLMs across a wide range of applications.

Fast weight architectures offer a promising alternative to attention-based transformers for long-context modeling by maintaining constant memory overhead regardless of context length. However, their potential is limited by the next-token prediction (NTP) training paradigm. NTP optimizes single-token predictions and ignores semantic coherence across multiple tokens following a prefix. Consequently, fast weight models, which dynamically update their parameters to store contextual information, learn suboptimal representations that fail to capture long-range dependencies. We introduce REFINE (Reinforced Fast weIghts with Next sEquence prediction), a reinforcement learning framework that trains fast weight models under the next-sequence prediction (NSP) objective. REFINE selects informative token positions based on prediction entropy, generates multi-token rollouts, assigns self-supervised sequence-level rewards, and optimizes the model with group relative policy optimization (GRPO). REFINE is applicable throughout the training lifecycle of pre-trained language models: mid-training, post-training, and test-time training. Our experiments on LaCT-760M and DeltaNet-1.3B demonstrate that REFINE consistently outperforms supervised fine-tuning with NTP across needle-in-a-haystack retrieval, long-context question answering, and diverse tasks in LongBench. REFINE provides an effective and versatile framework for improving long-context modeling in fast weight architectures.

Generative recommendation has recently emerged as a promising paradigm in sequential recommendation. It formulates the task as an autoregressive generation process, predicting discrete tokens of the next item conditioned on user interaction histories. Existing generative recommendation models are typically trained with token-level likelihood objectives, such as cross-entropy loss, while employing multi-step beam search during inference to generate ranked item candidates. However, this leads to a fundamental training-inference inconsistency: standard training assumes ground-truth history is always available, ignoring the fact that beam search prunes low-probability branches during inference. Consequently, the correct item may be prematurely discarded simply because its initial tokens (prefixes) have low scores. To address this issue, we propose the Adaptive Prefix-Aware Optimization (APAO) framework, which introduces prefix-level optimization losses to better align the training objective with the inference setting. Furthermore, we design an adaptive worst-prefix optimization strategy that dynamically focuses on the most vulnerable prefixes during training, thereby enhancing the model's ability to retain correct candidates under beam search constraints. We provide theoretical analyses to demonstrate the effectiveness and efficiency of our framework. Extensive experiments on multiple datasets further show that APAO consistently alleviates the training-inference inconsistency and improves performance across various generative recommendation backbones. Our codes are publicly available at https://github.com/yuyq18/APAO.

We present an approach to pose object recognition as next token prediction. The idea is to apply a language decoder that auto-regressively predicts the text tokens from image embeddings to form labels. To ground this prediction process in auto-regression, we customize a non-causal attention mask for the decoder, incorporating two key features: modeling tokens from different labels to be independent, and treating image tokens as a prefix. This masking mechanism inspires an efficient method - one-shot sampling - to simultaneously sample tokens of multiple labels in parallel and rank generated labels by their probabilities during inference. To further enhance the efficiency, we propose a simple strategy to construct a compact decoder by simply discarding the intermediate blocks of a pretrained language model. This approach yields a decoder that matches the full model's performance while being notably more efficient. The code is available at https://github.com/kaiyuyue/nxtp

Graphical user interface (GUI) grounding requires vision-language models (VLMs) to identify small target elements in high-resolution screenshots and predict precise screen coordinates. On-policy self-distillation (OPSD) is a promising post-training approach for this coordinate-sensitive task, since it provides dense token-level teacher signals beyond hard coordinate labels. However, naive OPSD is not well suited to GUI grounding: OPSD evaluates the teacher on student-generated prefixes, the quality of coordinate-token teacher signals can degrade when the prefix has already deviated from the target coordinate, leading to unreliable teacher signal. To mitigate this, We propose quality-aware self-distillation for VLM-based GUI grounding, which improves coordinate-token teacher-signal quality through soft correctness-aware gating and teacher-probability scaling. The soft correctness-aware gate checks whether the teacher's current coordinate-token prediction can still be completed into the ground-truth box under the student-generated prefix. If not, the corresponding teacher signal is down-weighted. Teacher-probability scaling then uses the teacher's confidence as a lightweight factor to further calibrate the strength of the gated supervision. A key empirical finding is that neither component alone improves overall performance, whereas combining them consistently improves performance. This suggests that the two mechanisms play complementary roles: correctness-aware gating suppresses unreliable coordinate-token supervision, while teacher-probability scaling calibrates the strength of the remaining signals. Experiments across six GUI grounding benchmarks show that our method consistently improves the base model and outperforms strong baselines.

The inverse problem of multilayer thin-film optical coatings design represents a complex combinatorial-continuous optimization challenge. We present PRISM (Position-encoded Regressive Inverse Spectral Model), a unified decoder-only autoregressive transformer that streamlines this process by jointly predicting discrete material selection and continuous thickness regression within a single backbone. PRISM introduces two primary architectural innovations: (1) spectrum prefix conditioning, which utilizes standard prefix tokens for in-context target injection, and (2) cumulative-depth Rotary Position Embeddings, which encode continuous thickness directly into the positional representation to preserve the physical spatial relationships of the stack. Our benchmarks demonstrate that a PRISM-13M model reduces MAE by over 50\% compared to other transformer baselines while utilizing only one-fifth of the parameters. Furthermore, a 44M-parameter variant achieves state-of-the-art performance (MAE = 0.010) on our in-distribution validation benchmark and operates significantly faster than simulated annealing, offering a highly efficient alternative to classical optimization methods.

We introduce the "Belief State Transformer", a next-token predictor that takes both a prefix and suffix as inputs, with a novel objective of predicting both the next token for the prefix and the previous token for the suffix. The Belief State Transformer effectively learns to solve challenging problems that conventional forward-only transformers struggle with, in a domain-independent fashion. Key to this success is learning a compact belief state that captures all relevant information necessary for accurate predictions. Empirical ablations show that each component of the model is essential in difficult scenarios where standard Transformers fall short. For the task of story writing with known prefixes and suffixes, our approach outperforms the Fill-in-the-Middle method for reaching known goals and demonstrates improved performance even when the goals are unknown. Altogether, the Belief State Transformer enables more efficient goal-conditioned decoding, better test-time inference, and high-quality text representations on small scale problems. Website: https://sites.google.com/view/belief-state-transformer

Reinforcement learning with verifiable rewards (RLVR) has become a dominant paradigm for improving reasoning in large language models (LLMs), yet the underlying geometry of the resulting parameter trajectories remains underexplored. In this work, we demonstrate that RLVR weight trajectories are extremely low-rank and highly predictable. Specifically, we find that the majority of downstream performance gains are captured by a rank-1 approximation of the parameter deltas, where the magnitude of this projection evolves near-linearly with training steps. Motivated by this, we propose a simple and compute-efficient method RELEX (REinforcement Learning EXtrapolation), which estimates the rank-1 subspace from a short observation window and extrapolates future checkpoints via linear regression, with no learned model required. Across three models (i.e., Qwen2.5-Math-1.5B, Qwen3-4B-Base, and Qwen3-8B-Base), RELEX produces checkpoints that match or exceed RLVR performance on both in-domain and out-of-domain benchmarks, requiring as few as 15% steps of full RLVR training. Remarkably, RELEX is able to extrapolate far beyond the observation window at no training cost, predicting checkpoints up to 10-20times beyond the observed prefix with continued improvement (e.g., observe only the first 50 steps and extrapolate to 1000 steps). Our ablation analysis confirms the minimalist sufficiency of RELEX: neither increasing the subspace rank nor employing non-linear modeling yields further gains in extrapolation. Finally, we show that RELEX's success stems from a "denoising" effect: by projecting updates onto the rank-1 subspace, the model discards stochastic optimization noise that would otherwise degrade performance during extrapolation. Our code is available at https://github.com/weizhepei/RELEX.

Recent work on KV cache quantization, culminating in TurboQuant, has approached the Shannon entropy limit for per-vector compression of transformer key-value caches. We observe that this limit applies to a strictly weaker problem than the one that actually matters: compressing the KV cache as a sequence. The tokens stored in a KV cache are not arbitrary floating-point data -- they are samples from the exact formal language the model was trained on, and the model is by construction a near-optimal predictor of that language. We introduce sequential KV compression, a two-layer architecture that exploits this structure. The first layer, probabilistic prefix deduplication, identifies semantically equivalent shared prefixes across sessions using the trie metric d_T(s, s') = -log_2 P_M(s ^ s') from Probabilistic Language Tries (PLTs). The second layer, predictive delta coding, stores only the residual of each new KV vector from the model's own prediction of it, achieving a per-token entropy bound of H(KV_{i+1} | KV_{<=i}) <= H(token_{i+1} | token_{<=i}). We prove that at typical language model perplexity -- approximately 10-20 for fluent English text -- this bound is 3.3-4.3 bits on average per token position, compared to TurboQuant's 3 bits per vector component (with typical attention heads having 64-128 components). The theoretical compression ratio over TurboQuant is approximately 914,000x at the Shannon limit. Even at 1000x above the entropy floor -- a deliberately pessimistic worst-case overhead, two orders of magnitude above the 2-5x typical of practical source coders -- the ratio remains approximately 914x over TurboQuant, with compression improving rather than degrading as context length grows. The two layers are orthogonal and compose with existing per-vector quantization methods including TurboQuant.

Diffusion large language models (dLLMs) re-encode the entire prefix at every denoising step, causing recomputation that scales quadratically with context length and becomes prohibitive for long-context scenarios. We propose Prefilling-dLLM, a training-free prefill-decode disaggregation framework for dLLMs that partitions the prefix into N chunks, caches their KV representations once, and selects the top-K most relevant chunks with intra-chunk token sparsity for decoding, showing that sparse prefilling can outperform dense attention while reducing per-step complexity from quadratic in the full sequence length to quadratic only in the decode length. On LongBench and InfiniteBench, Prefilling-dLLM achieves state-of-the-art quality among dLLM acceleration methods, and an attention kernel that parallelizes decoding over the non-contiguously cached chunk KV yields 9.1--28.0x speedup at 8K--32K contexts. We further show that beginning-of-sequence tokens prepended to each chunk act as periodic attention anchors that eliminate the lost-in-the-middle phenomenon. Code is available at https://github.com/menik1126/Prefilling-dLLM.

We present PEEK, a lightweight scheduling and eviction framework for both online (streaming) and offline (batch) LLM serving; this paper focuses on the online regime. PEEK maintains an incremental radix tree over the pending queue, exposing prefix-sharing clusters no existing engine surfaces. A low-overhead dual-walk matches the tree against the engine's prefix cache to yield longest-prefix-match for every waiting request; PEEK then admits cluster pioneers first so siblings inherit the freshly cached prefix, a co-designed eviction hook protects blocks ancestral to queued demand, and a multi-lane stride scheduler bounds starvation. On SGLang and vLLM across five workloads up to 4timesH100 (DP=2 over TP=2), PEEK delivers up to 3.0times/2.6times cache hit, 7.9times/7.1times TTFT, 6.7times/5.5times E2E, and 3.6times/4.5times throughput gains over each engine's strongest stock baseline (SGLang/vLLM), while matching baselines within noise on workloads with no exploitable prefix structure. Wins hold as KV-cache pressure and inference parallelism scale.

Large language models (LLMs) excel at generating contextually relevant content. However, tailoring these outputs to individual users for effective personalization is a significant challenge. While rich user-specific information often exists as pre-existing user representations, such as embeddings learned from preferences or behaviors, current methods to leverage these for LLM personalization typically require costly fine-tuning or token-heavy prompting. We propose Embedding-to-Prefix (E2P), a parameter-efficient method that injects pre-computed context embeddings into an LLM's hidden representation space through a learned projection to a single soft token prefix. This enables effective personalization while keeping the backbone model frozen and avoiding expensive adaptation techniques. We evaluate E2P across two public datasets and in a production setting: dialogue personalization on Persona-Chat, contextual headline generation on PENS, and large-scale personalization for music and podcast consumption. Results show that E2P preserves contextual signals and achieves strong performance with minimal computational overhead, offering a scalable, efficient solution for contextualizing generative AI systems.

Large Language Models often improve accuracy on reasoning tasks by sampling multiple Chain-of-Thought (CoT) traces and aggregating them with majority voting (MV), a test-time technique called self-consistency. When we truncate a CoT partway through and regenerate the remainder, we observe that traces with correct answers reproduce their original answer more often than traces with wrong answers. We use this difference as a reliability signal, prefix consistency, that weights each candidate answer by how often it reappears under regeneration. It requires no access to token log-probabilities or self-rating prompts. Across five reasoning models and four math and science benchmarks, prefix consistency is the best correctness predictor in most settings, and reweighting votes by it reaches Standard MV plateau accuracy at up to 21x fewer tokens (median 4.6x). Our code is available at https://github.com/naoto-iwase/prefix-consistency.

Reinforcement learning with verifiable rewards (RLVR) is a promising approach for enhancing reasoning and agentic behavior in large language models. However, rollout-intensive policy optimization is often limited by insufficient reward contrast, arising when overly simple or complex prompts generate low-variance feedback and when outcome-only rewards assign the same terminal assessment to every decision in a multi-turn rollout. Past efforts have focused on allocating available rollout resources to promising prompts, yet they only leverage sample informativeness at the prompt level and neglect variation in prefix-level informativeness across turns within the same rollout. This work targets multi-turn agentic RL by modeling each ReAct-style thought-action-observation turn as a semantically distinct node, allowing budget allocation to extend from prompt roots to turn-level prefixes with further continuations, which naturally forms tree-structured rollouts. We introduce Tree Rollout Allocation for Contrastive Exploration (TRACE), a unified rollout allocation framework that enhances reward contrast within a fixed sampling budget. Technically, TRACE allocates rollout budget to both prompt roots and intermediate prefixes that are most likely to yield mixed terminal rewards. A shared generalizable predictor estimates conditional success probability at these anchors from prefix histories to guide this allocation. The resulting adaptive tree structure enriches outcome-only feedback and amplifies the policy-update signal. Empirically, TRACE achieves competitive performance and efficiency gains on typical agentic benchmarks, e.g., improving Qwen3-14B Multi-Hop QA average accuracy by 2.8 points over competitive baselines at equal sampling cost.

Retrieval-Augmented Generation (RAG) is often used with Large Language Models (LLMs) to infuse domain knowledge or user-specific information. In RAG, given a user query, a retriever extracts chunks of relevant text from a knowledge base. These chunks are sent to an LLM as part of the input prompt. Typically, any given chunk is repeatedly retrieved across user questions. However, currently, for every question, attention-layers in LLMs fully compute the key values (KVs) repeatedly for the input chunks, as state-of-the-art methods cannot reuse KV-caches when chunks appear at arbitrary locations with arbitrary contexts. Naive reuse leads to output quality degradation. This leads to potentially redundant computations on expensive GPUs and increases latency. In this work, we propose Cache-Craft, a system for managing and reusing precomputed KVs corresponding to the text chunks (we call chunk-caches) in RAG-based systems. We present how to identify chunk-caches that are reusable, how to efficiently perform a small fraction of recomputation to fix the cache to maintain output quality, and how to efficiently store and evict chunk-caches in the hardware for maximizing reuse while masking any overheads. With real production workloads as well as synthetic datasets, we show that Cache-Craft reduces redundant computation by 51% over SOTA prefix-caching and 75% over full recomputation. Additionally, with continuous batching on a real production workload, we get a 1.6X speed up in throughput and a 2X reduction in end-to-end response latency over prefix-caching while maintaining quality, for both the LLaMA-3-8B and LLaMA-3-70B models.

Bayesian predictive inference provides a principled framework for uncertainty quantification, data efficiency, and robust generalization. However, exact inference is often intractable, and scalable approximations may remain computationally expensive or require restrictive modeling assumptions that degrade predictive performance. Prior-Data Fitted and in-context models have recently emerged as an amortized alternative by learning to map datasets directly to predictive distributions, but existing approaches are tightly coupled to the support of the training prior and lack explicit mechanisms for adapting to new priors at test time, resulting in limited robustness under distribution shift. We introduce a multi-task in-context learning framework for amortized hierarchical Bayesian predictive inference that explicitly represents prior information as a prefix of in-context datasets. A transformer trained on sequences of prior and target tasks learns to adapt its predictions across families of priors. On a suite of evaluations with increasing difficulty, including out-of-meta-distribution priors and priors with high-dimensional latent structures, our method matches oracle Bayesian predictors while being orders of magnitude faster. We further demonstrate its practical relevance on a real-world spatiotemporal temperature prediction benchmark. Code is available at https://github.com/martianmartina/multi-task-bayesian-icl/.

Although Large Language Models (LLMs) have achieved tremendous success in various applications, they are also susceptible to certain prompts that can induce them to bypass built-in safety measures and provide dangerous or illegal content, a phenomenon known as jailbreak. To protect LLMs from producing harmful information, various defense strategies are proposed, with most focusing on content filtering or adversarial training of models. In this paper, we propose an approach named Prompt Adversarial Tuning (PAT) to train a defense control mechanism, which is then embedded as a prefix to user prompts to implement our defense strategy. We design a training process similar to adversarial training to achieve our optimized goal, alternating between updating attack and defense controls. To our knowledge, we are the first to implement defense from the perspective of prompt tuning. Once employed, our method will hardly impact the operational efficiency of LLMs. Experiments show that our method is effective in both black-box and white-box settings, reducing the success rate of advanced attacks to nearly 0 while maintaining the benign answer rate of 80% to simple benign questions. Our work might potentially chart a new perspective for future explorations in LLM security.

Semantic ID (SID)-based recommendation is a promising paradigm for scaling sequential recommender systems, but existing methods largely follow a semantic-centric pipeline: item embeddings are learned from foundation models and discretized using generic quantization schemes. This design is misaligned with generative recommendation objectives: semantic embeddings are weakly coupled with collaborative prediction, and generic quantization is inefficient at reducing sequential uncertainty for autoregressive modeling. To address these, we propose ReSID, a recommendation-native, principled SID framework that rethinks representation learning and quantization from the perspective of information preservation and sequential predictability, without relying on LLMs. ReSID consists of two components: (i) Field-Aware Masked Auto-Encoding (FAMAE), which learns predictive-sufficient item representations from structured features, and (ii) Globally Aligned Orthogonal Quantization (GAOQ), which produces compact and predictable SID sequences by jointly reducing semantic ambiguity and prefix-conditional uncertainty. Theoretical analysis and extensive experiments across ten datasets show the effectiveness of ReSID. ReSID consistently outperforms strong sequential and SID-based generative baselines by an average of over 10%, while reducing tokenization cost by up to 122x. Code is available at https://github.com/FuCongResearchSquad/ReSID.

Increasingly, LLM inference services proxy client requests to engine replicas distributed globally. Load-balancing policies must jointly account for factors including KV-cache locality, replica load, and variable network latency when optimizing for metrics like latency and TTFT. However, existing systems only evaluate a subset of these factors in their cost model, leading to uneven concentrations of load and KV-cache across replicas. We present GORGO, a proxy architecture that holistically factors network latency, prefill cost, and queueing delay using tunable parameters. Since open-source chat datasets such as LMSYS-Chat1M and WildChat-4.8M lack long-context, high prefix-reuse data, we release a synthetic dataset, ART-Chat-2.5M, from long-context production metadata. On a tuning window from ART-Chat-2.5M, evolutionary strategies guide the GORGO policy's parameters to directly optimize p95 TTFT. During held-out evaluation windows, we fix the parameter values learned from tuning and improve p95 TTFT by 6.9-15.5% and p95 end-to-end (E2E) latency by 14.3-30.9% over baseline load-balancing policies such as simple session affinity and prefix-cache. The code and ART-Chat-2.5M dataset can be found at https://github.com/Arcadia-Research-Team/GORGO.