"""IdeaBench-style evaluation queries (Sec 5.1 proxy). The paper's IdeaBench is built on AI Idea Bench 2025 (Qiu et al., 2025; 3,495 papers, 8 domains) which is not openly redistributable here. We build a small, faithful proxy: one query per domain, each with a topic x_t (problem statement, solution withheld), related works x_r (inspirational prior papers), and an expert-written reference idea y* distilled from a real, well-known target paper. This is a SCALED proxy of the benchmark, not the full 3,495-paper set. """ QUERIES = [ { "domain": "CV", "topic": "Building a promptable foundation model for image segmentation that generalizes zero-shot to unseen objects and tasks via flexible prompts.", "related": [ "Fully convolutional and Mask R-CNN style networks segment fixed, pre-defined object categories from labelled masks.", "Vision transformers (ViT) provide scalable image encoders pre-trained on large image corpora.", "CLIP aligns images and text with contrastive pre-training, enabling open-vocabulary recognition.", "Interactive segmentation methods refine masks from user clicks but are trained per-dataset.", ], "reference": { "title": "Segment Anything: a promptable segmentation foundation model", "problem": "Segmentation models are trained per task/category and do not generalize to new objects or prompts; there is no foundation model for segmentation nor a web-scale mask dataset.", "method": "A promptable model (image encoder + prompt encoder + fast mask decoder) trained to output valid masks for point/box/text prompts; a data engine iteratively co-develops the model and a 1B-mask, 11M-image dataset via model-in-the-loop annotation.", "evaluation": "Zero-shot transfer to 23 segmentation datasets and downstream tasks (edge detection, proposals, instance segmentation); compare mask quality and human ratings vs interactive and supervised baselines.", }, }, { "domain": "NLP", "topic": "Aligning language models with human preferences without the instability and complexity of reinforcement learning from a separate reward model.", "related": [ "RLHF fits a reward model to human comparisons and optimizes the policy with PPO, which is unstable and sample-inefficient.", "The optimal RLHF policy has a closed form: a reweighting of the reference policy by the exponentiated reward.", "Bradley-Terry models express pairwise preference probabilities via a latent scalar score.", ], "reference": { "title": "Direct Preference Optimization: your language model is secretly a reward model", "problem": "RLHF is complex and unstable, requiring reward-model training plus RL. Can we optimize the same objective directly from preference pairs?", "method": "Reparameterize the RLHF objective so the reward is expressed via the policy's log-ratio to a reference model, yielding a simple classification-style loss on preference pairs (no reward model, no RL, no sampling).", "evaluation": "Sentiment control, summarization and single-turn dialogue; compare win-rate vs PPO-RLHF and preferred-FT baselines under GPT-4/human judgment; measure stability and compute.", }, }, { "domain": "Multi-modal", "topic": "Giving a large language model general-purpose visual understanding and instruction-following using limited multimodal supervision.", "related": [ "CLIP/ViT encoders yield strong transferable image features aligned to text.", "Instruction tuning on GPT-generated data makes LLMs follow diverse natural-language instructions.", "Flamingo/BLIP-2 connect frozen vision encoders to LLMs with learned adapters for captioning and VQA.", ], "reference": { "title": "Visual Instruction Tuning (LLaVA)", "problem": "There is little multimodal instruction-following data and no simple recipe to turn an LLM into a general visual assistant.", "method": "Use a language-only GPT to synthesize multimodal instruction-following data from image captions/boxes, then connect a CLIP visual encoder to an LLM via a projection and instruction-tune end-to-end.", "evaluation": "A synthetic multimodal instruction benchmark plus ScienceQA; compare against BLIP-2/visual chat baselines with GPT-4 as judge and accuracy.", }, }, { "domain": "Audio & Speech", "topic": "Endowing a large language model with perception and understanding of general audio, speech, and music from raw waveforms.", "related": [ "Whisper provides robust multilingual speech recognition from weak supervision at scale.", "Audio encoders (BEATs) learn general acoustic representations for non-speech sounds.", "Q-Former style adapters bridge frozen encoders and LLMs for cross-modal reasoning.", ], "reference": { "title": "SALMONN: towards generic hearing for large language models", "problem": "Speech LLMs handle only speech; there is no single model that jointly understands speech, audio events and music and generalizes to unseen audio tasks.", "method": "Fuse a Whisper speech encoder and a BEATs audio encoder via a window-level Q-Former into an LLM; train with cross-modal instruction tuning plus an activation-tuning stage to unlock emergent cross-modal abilities.", "evaluation": "Speech recognition/translation, audio captioning, music QA and unseen tasks (e.g., audio-based storytelling); compare to task-specific and cascaded baselines.", }, }, { "domain": "Robotics & Control", "topic": "Learning visuomotor manipulation policies that model multimodal, temporally-consistent action distributions from demonstrations.", "related": [ "Behavior cloning with unimodal Gaussian or MSE heads struggles with multimodal demonstrations.", "Diffusion models learn expressive multimodal distributions via iterative denoising.", "Action-chunking and receding-horizon control improve temporal consistency in imitation.", ], "reference": { "title": "Diffusion Policy: visuomotor policy learning via action diffusion", "problem": "Imitation policies collapse multimodal action distributions and produce temporally inconsistent actions, limiting dexterous manipulation.", "method": "Represent the policy as a conditional denoising diffusion process over action sequences, using receding-horizon control, visual conditioning, and time-series diffusion for stable, multimodal action generation.", "evaluation": "15 tasks across simulation and real robots; compare success rate vs LSTM-GMM/BET/IBC behavior-cloning baselines and ablate horizon and conditioning.", }, }, { "domain": "Science", "topic": "Generating novel, stable inorganic crystalline materials with targeted properties beyond substitution-based search.", "related": [ "Graph neural network potentials predict formation energy and stability of crystals.", "Diffusion and flow models generate molecules and conformers in continuous space.", "Large DFT-relaxed datasets provide stability labels for inorganic compounds.", ], "reference": { "title": "A generative model for inorganic materials design (MatterGen)", "problem": "Materials discovery relies on screening known/substituted structures, rarely proposing genuinely novel stable crystals with desired properties.", "method": "A diffusion model over atom types, coordinates and the periodic lattice that generates stable, diverse crystals, with property-conditioned adapters fine-tuned for chemistry, symmetry, and mechanical/magnetic/electronic targets.", "evaluation": "Fraction of stable-unique-novel structures verified by DFT; property-conditioned generation success and distance-to-hull vs substitution and prior generative baselines.", }, }, { "domain": "General ML", "topic": "Designing a sequence model with Transformer-level quality but linear-time inference and long-context efficiency.", "related": [ "Transformers have quadratic attention cost and a growing KV cache with sequence length.", "Structured state-space models (S4) capture long dependencies with linear recurrence but weaker language modeling.", "Gating and input-dependent dynamics improve selectivity in recurrent models.", ], "reference": { "title": "Mamba: linear-time sequence modeling with selective state spaces", "problem": "Attention is inefficient for long sequences while prior SSMs underperform Transformers on information-dense modalities like language.", "method": "Make SSM parameters input-dependent (selective) so the model propagates or forgets state per token, and implement a hardware-aware parallel scan for a simplified attention-free architecture.", "evaluation": "Language modeling perplexity and downstream tasks vs Transformers/Transformer++ at matched size, plus DNA and audio; measure throughput and scaling with context length.", }, }, { "domain": "Theory", "topic": "Determining how to jointly allocate model size and training tokens under a fixed compute budget for large language models.", "related": [ "Kaplan et al. scaling laws suggested favoring larger models over more data.", "Compute-optimal training seeks the best loss for a fixed FLOP budget.", "Power-law fits relate loss to parameters and data separately.", ], "reference": { "title": "Training compute-optimal large language models (Chinchilla)", "problem": "Given a fixed compute budget, how should parameters and training tokens be traded off? Prior laws over-sized models relative to data.", "method": "Fit a parametric loss surface L(N,D) from >400 runs across model/data scales using three complementary estimation approaches, then solve for the compute-optimal (N,D) frontier.", "evaluation": "Predict and verify a compute-optimal model (Chinchilla) trained with the derived N,D; compare downstream accuracy vs larger under-trained models (Gopher/GPT-3) at equal compute.", }, }, ] def get_queries(domains=None): if domains: return [q for q in QUERIES if q["domain"] in domains] return list(QUERIES)