Research Briefs
Summaries of trending research from Hugging Face Daily Papers, written for builders. 255 papers from July 2026.
58 papers
AskChem changes the retrieval unit from the paper to the provenance-carrying claim - 2.4M claims, each grounded by DOI and verbatim quote, served to agents over MCP. Grounding a GPT-5.5 reader in AskChem gives 100% resolvable DOIs vs 88.3% without retrieval; ungrounded, the same reader fabricated 6 of 14 DOIs on one question.
A foundation GUI agent spanning mobile, computer-use, web and DeepSearch, trained with online RL on 100+ turn trajectories across 10,000+ concurrent environments; SOTA on mobile-use benchmarks and competitive with frontier models on computer and browser use.
Metis introduces the first memory foundation model: a persistent, dynamically evolving memory state inside the backbone, updated gradient-free in a single forward pass at inference, with learned native memory procedures acquired through mid-training.
** Converts sparse outcome rewards in long-horizon agentic RL into turn-level credit signals with a critic-free, recursive Bayesian belief update in log-odds space - no extra critic, no extra rollouts - beating GRPO and strong self-distillation baselines on ALFWorld, WebShop, and Search-QA.
** Automates the measurement of harness optimization - iteratively editing the prompts, tools, control flow, memory, and orchestration code around a target agent under a fixed evaluation budget - and finds optimizer models separate more than the coding harnesses they act through.
Microsoft Research compiles specs into stateful synthetic apps graded against their own databases and co-evolves environments with the model: a 9B agent reaches 67.1% across 14 evaluation splits (from 36.5%), within 14 points of its frontier teacher, while shallow environments actively hurt live accuracy (80.0 to 75.0).
Retrieval-based speculative decoding drafts tokens by copying continuations from existing text, which fits coding agents that repeatedly reproduce code, logs, and earlier attempts. AgSpec supplies the corpora and draft-length policies existing retrieval engines lack, raising generation throughput up to 4.37x at batch size 1 and 4.76x at batch size 16.
Eight parallel 10-agent worlds ran continuously for 16 days (850,000+ LLM calls, ~50B tokens) with persistent memory, shared tools, and self-governed institutions; after operational state accumulated, three stress events (indirect prompt injection, misinformation, private-memory exposure) defeated every world, with agents acting on adversarial content up to 46 hours later.
Execution guardrails mined from real coding-agent failure traces instead of hand-written rules: learned constraints cut the abnormal execution rate from 69.0% to 26.7% and lift task completion from 21.7% to 35.0% on a disjoint evaluation set (Claude Code + Claude Haiku 4.5).
+49 more
55 papers
Batching memory consolidation to semantic segment boundaries instead of every turn reaches 89.22% on LoCoMo and 92.20% on LongMemEval-S while cutting construction tokens 86.0% and 75.9% versus A-Mem, with no increase in query-time token usage.
Interactive world model that externalizes persistent world state into a camera-indexed state bank with view-relevant retrieval, keeps the denoiser context bounded as sessions grow, and distills drift resistance from a linear-scaling long-horizon teacher into a three-step student - state-of-the-art on WBench, 2.11s per 1.5s chunk on a single H200.
DOPD introduces advantage-aware dual distillation that dynamically routes token-level supervision between teacher and student policies based on the privilege advantage gap, solving the 'privilege illusion' failure mode in on-policy distillation.
BlockPilot reveals that fixed block-size strategies in diffusion-based speculative decoding are fundamentally suboptimal, and proposes a lightweight instance-adaptive policy that predicts the optimal block size from the prefilling representation, achieving 4.20× speedup on Qwen3-4B.
Factoring the training loop instead of the generation procedure - explore K candidate matches between model generations and data, train on the best - adds exploration as a third pretraining axis: 4.1x FLOP efficiency, 6.2x sample efficiency, and end-to-end generation with 16-256x fewer inference steps than diffusion.
GEAR trains a VQ tokenizer and autoregressive generator jointly end-to-end via a dual hard/soft codebook readout, achieving up to 10x faster gFID convergence on ImageNet by shifting representation alignment from the tokenizer to the AR model.
Scaling a pretrained parametric memory module to 6.9B parameters on 300B tokens: a 410M base with that memory beats Pythia-12B on 17 benchmarks with 39% fewer total parameters, and 1.7B domain memories add more than 9 points to Qwen3 bases from 0.6B to 14B.
Retrieval-based speculative decoding drafts tokens by copying continuations from existing text, which fits coding agents that repeatedly reproduce code, logs, and earlier attempts. AgSpec supplies the corpora and draft-length policies existing retrieval engines lack, raising generation throughput up to 4.37x at batch size 1 and 4.76x at batch size 16.
** Data augmentations inherited from natural image tasks can disrupt the fine-grained vascular topology and textures critical for identity discrimination in vein recognition.
+46 more
101 papers
Orca introduces a general world foundation model that learns a unified latent space from multimodal signals via Next-State-Prediction, then freezes its backbone and attaches lightweight readout decoders for text, image, and action - outperforming specialized baselines across all three modalities.
Metis introduces the first memory foundation model: a persistent, dynamically evolving memory state inside the backbone, updated gradient-free in a single forward pass at inference, with learned native memory procedures acquired through mid-training.
Taste-Bench measures the quality of an agent's long-horizon decisions by mining decision forks from real agent trajectories without human annotation. The best frontier model answers only 59.7% of taste questions correctly, forks whose deciding evidence appears late are harder for every model, and a larger reasoning budget does not help - but taste is trainable via outcome distillation, and the trained student improves end-to-end success on held-out SWE-bench Pro tasks.
GRPO broadcasts a response-level advantage to every token, leaving within-response credit unallocated. CoRT rescores the same sampled response under the rubric-conditioned prompt and a matched criteria-free prompt, and uses the tokenwise log-likelihood contrast as a proxy for rubric dependence, redistributing the signed advantage across tokens - no auxiliary scorer, +4.4 points on average over response-level GRPO.
** Converts sparse outcome rewards in long-horizon agentic RL into turn-level credit signals with a critic-free, recursive Bayesian belief update in log-odds space - no extra critic, no extra rollouts - beating GRPO and strong self-distillation baselines on ALFWorld, WebShop, and Search-QA.
RL post-training sharpens a model toward what it already solves: pass@1 and sampling consistency improve, but solution coverage (pass@K) shrinks, so harness-equipped base models can match or surpass post-trained counterparts given a large test-time budget. Across 14 checkpoint pairs from four families and three agentic benchmarks (42 cases), the paper's Sharpening Tax is prevalent and estimable from a few rollouts, and a difficulty-adaptive sampler shrinks it during training.
Factoring the training loop instead of the generation procedure - explore K candidate matches between model generations and data, train on the best - adds exploration as a third pretraining axis: 4.1x FLOP efficiency, 6.2x sample efficiency, and end-to-end generation with 16-256x fewer inference steps than diffusion.
Harness optimization has been fixing the scenario order that generates feedback while optimizing how the harness updates. ActiveSaddler treats the curriculum as a non-stationary bandit over reusable failure patterns that co-evolves with the harness, improving test Pass@1 by 4.4 points on GAIA2 and 7.5 points on Terminal-Bench 2.0 over the same optimizer with a scenario order fixed before optimization.
Scaling a pretrained parametric memory module to 6.9B parameters on 300B tokens: a 410M base with that memory beats Pythia-12B on 17 benchmarks with 39% fewer total parameters, and 1.7B domain memories add more than 9 points to Qwen3 bases from 0.6B to 14B.
+92 more
33 papers
An end-to-end JEPA that turns action-labeled trajectories into a search-free policy: the conditional mean of a learned distributional action law directly selects actions, hitting 85.78-100% success on LeWM with zero search and 2.9-5.5ms inference, and cutting CEM sampling 23.44x.
Current VLA models retain shallow perceptual knowledge (color, shape) after robotics fine-tuning but catastrophically drop performance on richer semantic categories (emotion, counting, temporal, normative, cultural knowledge) - with answer-relevant information still present in intermediate layers yet failing to reach action output.
** Existing World Action Models fail at mobile manipulation due to three structural misalignments - coarse video prediction doesn't match fine-grained control, entangled navigation/manipulation action spaces cause gradient interference, and training on ground-truth futures doesn't generalize to the model's own noisy rollouts at inference time.
** Traditional robot programming is difficult due to the need to orchestrate multimodal perception, manage contact dynamics, and handle diverse configurations and failures, with no existing system that autonomously writes, refines, and transfers reusable control programs across tasks and embodiments.
** LLMs lack a learned memory management strategy - they don't know what to encode, when to retrieve, or how to organize knowledge over long-horizon tasks.
** Existing data mixture optimization methods assume static data distributions and require costly retraining from scratch when the data pool shifts, preventing scalable transfer across data pools and model sizes.
** Existing text-rich image data pipelines follow a static crawl-filter-freeze paradigm that discards rejected samples, wasting failure signals (OCR errors, semantic mismatches) that could inform later construction rounds.
** VLA models fine-tuned from powerful VLMs on robotics data may catastrophically forget commonsense and world knowledge, but existing benchmarks conflate knowledge gaps with low-level control failures.
** Vision-Language-Action (VLA) models fail under environmental shifts (camera pose, embodiment changes) and existing adaptation methods require costly multi-demonstration data per task.
+24 more
23 papers
** Data augmentations inherited from natural image tasks can disrupt the fine-grained vascular topology and textures critical for identity discrimination in vein recognition.
** Existing audio-video generation models use separate per-modality tokenizers, which creates a representation gap, causes semantic misalignment, and requires expensive dual-branch architectures.
** Existing diffusion-based speculative decoding methods use a fixed block size for all inputs, which is suboptimal since the optimal block size varies across samples.
** Existing brain encoding and decoding models treat these as separate tasks using unimodal alignment, ignoring the brain's intrinsic multimodal integration nature.
** Blind image deblurring methods struggle with real-world spatially varying degradations and lack the semantic awareness needed to distinguish valid textures from artifacts.
** Can a discrete diffusion language model match or exceed autoregressive models on medical report generation while offering capabilities autoregressive models lack?
** Standard two-stage visual generative models (tokenizer → frozen generator) decouple training, leaving the tokenizer unaware of what the generator finds easy or hard to model.
** Existing controllable image generation methods like ControlNet struggle with attribute confusion in multi-instance scenes, while approaches that fix this require labor-intensive manual instance labeling.
** Can small language models (SLMs) serve as viable, GPU-free replacements for large language models in the generation stage of Retrieval-Augmented Generation (RAG) systems?
+14 more
36 papers
On 97 end-to-end scientific workflows across six domains, the best model-and-scaffold configuration completes only 20.6% of tasks, partial-progress scores of 87.6-94.9 coexist with 0-4% completion rates, and 75.5% of non-passing Claude Code trajectories still end with language claiming the job is done.
Code2Skill transforms source code into grounded agent skills without any agent interaction experience: 1,006,822 verified skill records mined and verified from 19,769 GitHub repositories, lifting model performance by 11.7% on average across 72 protocol-matched evaluations and beating trajectory-derived skill banks on all seven shared benchmarks - and the pipeline stays effective on AI-generated code (93.50% pass rate vs 93.00% for human-written).
** Automates the measurement of harness optimization - iteratively editing the prompts, tools, control flow, memory, and orchestration code around a target agent under a fixed evaluation budget - and finds optimizer models separate more than the coding harnesses they act through.
Ten AI systems evaluated on search-and-discovery benchmarks whose rules are executable and conflict with familiar knowledge, so exploration can be verified exactly and recall alone cannot answer - the strongest systems acquire the unfamiliar rules, but continued exploration can stall or reverse earlier gains.
A controlled study of recursive AI peer review: training successor reviewers on data mixed with model-generated reviews compresses rating distributions and shrinks semantic diversity of judgments - 'scientific-judgment collapse'. The mitigation (TrustReviewer) combines single-stage training on a curated corpus with paired activation steering at test time, needing no additional expert annotation.
Benchmark accuracy conflates two separate questions - whether a response reached an evaluable state and whether its answer was judged correct. Under matched token budgets, models produce sharply different execution mixtures (49 of 450 Qwen MATH outputs terminate without a final answer vs 5 of 300 DeepSeek), so accuracy comparisons quietly mix execution case mix with verification policy.
** Popular performance-optimization benchmarks for coding agents conflate runtime instability, scoring-rule artifacts, and saturation effects, making leaderboard scores unreliable indicators of true coding-agent progress.
A trajectory-level study of 3,500 diagnostic runs shows RCA agents can localize the right service while failing to reconstruct how the fault propagated - and endpoint-correctness benchmarks cannot see the difference. Success needs grounding before localization, which the authors package as DiagGuard.
** Benchmark pass rates for coding agents can be near-perfect even when the agent failed to build the requested artifact, because the agent satisfies the test oracle by inlining behavior into a throwaway demo instead of the reusable library.
+27 more