Search posts, papers, and topics
131 posts · huggingface.co
Mind2Dialogue introduces a psychology‑guided simulator that generates user mental states and uses privileged distillation to train LLM assistants on Oracle responses that know those states. The resulting models improve personalization and theory‑of‑mind metrics by up to 41 percentage points versus standard instruction‑tuned baselines.
The paper presents HypoEvolve, a generational genetic algorithm that coordinates specialized LLM agents to iteratively propose, critique, and refine scientific hypotheses. On a drug‑repurposing benchmark across 34 cancer types, it outperforms six baselines, achieving a DepMap selectivity of 0.171 versus 0.115.
CADWorld is a new benchmark suite of 200 long‑horizon mechanical CAD tasks in FreeCAD, covering sketching, part modeling, assembly, CAM, FEM, and more. Agents interact via screenshots and GUI actions; success is checked by executable validation of the saved CAD artifacts. Seven existing agents achieve at most 17.5 % success versus an 87 % expert baseline, highlighting the gap between GUI competen…
ModularRSI proposes a contrastive, modular approach to evolve agent harnesses by independently updating five functional modules and aggregating evidence across 2,000 benchmark‑disjoint tasks. Experiments on TB2.0 and SWE‑Bench Verified show consistent gains on unseen tasks and cross‑model transfer.
Register tokens are fixed‑position embeddings that store a compact hidden state across diffusion‑based language model generation chunks, enabling bounded‑state reasoning without retaining all prior text. Post‑training on LLaDA and Dream shows up to +8.5 math and +19.5 code benchmark points versus plain text carry, and RL fine‑tuning further improves long‑horizon tasks.
VC-Attention introduces a training‑free low‑bit attention pipeline for diffusion transformers. It smooths value tensors via lightweight online clustering (V‑Smooth) and quantizes only the residual after subtracting block means, restoring the mean from the softmax row sum. It also replaces the FP32 softmax exponential with a fused FP8 cast (ExpCast‑FP8) that maps log‑scores directly to E4M3 probab…
The paper proposes Mixture of Memory Embeddings (MoME), a context‑aware sparse lookup that replaces each token’s single memory row with a gated mixture of multiple slots. Experiments on Llama‑3, MobileLLM and Qwen3 show MoME outperforms existing memory‑embedding baselines at equal parameter and FLOP budgets and exhibits interpretable routing for polysemous tokens.
HarnessVLN introduces a zero‑shot, training‑free embodied navigation framework that wraps a multimodal LLM in an "Agent Harness" – a tool‑based protocol that validates planner actions against spatial evidence, tracks progress with hierarchical event memory, and maintains a persistent spatiotemporal graph for recovery. The system works for instruction‑following and object‑goal tasks, achieving 60.…
The paper experiments with prompting six frontier LLM families using a school‑audience framing and finds their architectural design suggestions converge on a common pattern, while removing the framing yields diverse outputs. It highlights a possible shared design prior among models and introduces the term "epistemic jailbreak" for the loss of provenance in speculative answers.
This paper introduces SpectralShift, a spectral reparameterization method for extending the context window of Gated DeltaNet (GDN) linear attention models. It reconfigures the decay spectrum by enhancing slow propagation and preserving fast-decaying modes, consistently improving long-context capabilities during continual pretraining.
E2A‑Bench is a 969‑query benchmark for financial chart reasoning that evaluates vision‑language models across a full evidence‑to‑action chain using four metrics (UCR, RCI, ECI, NDR). Experiments on 20 VLMs expose hidden failures: low‑UCR models have only 6.4 % directional coverage, oracle‑aided verification cuts unsupported claims but can kill coverage, and fine‑tuning inflates BUY:SELL ratios by…
OmniHarness introduces a symbolic‑policy framework that extracts reusable procedural knowledge from multimodal LLM‑driven visual generation runs. By decoupling task logic from instance inputs, the system can instantiate, adapt, and compose policies for new visual tasks, using intermediate verification for on‑the‑fly refinement while keeping the underlying model frozen. Self‑directed practice task…
Lightning Weave is a post‑training framework that composes specialist capabilities from separate models into one student via on‑policy distillation, improving both accuracy and token efficiency on math and code benchmarks.
StepAudio 3 Realtime is an audio‑language foundation model that runs a continuous listen‑converse‑think‑act loop. It introduces Deep Perception for rich acoustic cue extraction, Seamless Duplex for handling pauses/back‑channels, and a Think‑While‑Speaking mechanism that lets the model reason in parallel with speech output. On benchmarks it scores 73.0 macro avg on StepAudioChat, 90.6 on MMSU, 98.…
The paper reframes fine‑tuning of instruction‑tuned LLMs as a direction‑selection problem under a fixed behavioral‑drift budget, showing that the update direction, not magnitude, determines trade‑offs between target performance and capability preservation. In QA‑only fine‑tuning of Qwen‑3 models, layer‑selective probing finds effective directions that boost scientific reasoning and multilingual t…
The paper builds compositional rule‑induction task families and generates isomorphic variants to probe systematicity in reasoning models. Experiments reveal that models frequently fail on equivalent variants despite solving the original task, indicating a lack of systematic reasoning.
The paper proposes the Convergent Emergence Hypothesis that few‑shot in‑context learning (ICL) shares a common difficulty profile across domains. Using a unified task suite, the authors evaluate ICL on six modalities—language, genome, integer sequences, time‑series, images, and proteins—showing that paired‑mapping ICL emerges in all and that per‑task benefits correlate across five modalities, sup…
The paper shows that when specialist LLMs are trained only on QA pairs (no explicit reasoning supervision), their optimization implicitly selects a latent distribution of reasoning trajectories. By treating the distilled student as an agnostic probe—since it inherits only the sampled trajectories—the authors empirically demonstrate a strong correlation (across 27 specialist‑student pairs) between…
Latent Interface Training (LIT) first learns a goal‑conditioned action prior without visual input, then adds a pose‑supervised latent interface as the only visual conditioning path. Applied to several vision‑language‑action models, LIT cuts vision‑action shortcuts and lifts LIBERO‑Plus success by 3.9–10.7 points and real‑world task success by 13.3–16.7 points under distribution shifts.
StepAudio 3 Gen is a general‑purpose audio generation model that replaces diffusion with a discrete autoregressive generator over residual vector quantization tokens. Using a 16‑layer RVQ tokenizer and progressive pretraining, it reaches state‑of‑the‑art zero‑shot TTS and voice‑design performance while handling speech, vocals, sound effects, and music.
RelateAnything is a 53 M‑parameter model that predicts arbitrary textual relations between any supplied image regions, using a text‑embedding bank instead of a fixed classifier. Trained on a new 4.3 M‑relation dataset, it outperforms prior open‑vocab methods by 2.3‑3.5× mean recall while running in ~20 ms per frame.
The paper introduces Generalized Agent Iteration (GAI), a formal framework that unifies classical iterative policy improvement (GPI) and recursive self‑improvement (RSI). GAI treats an agent as a set of modifiable components and models learning as a loop of evaluation and improvement. Two binary “dials”—whether the improvement mechanism is internal to the agent and whether the evaluation standard…
The paper presents SteerDuplex, a full‑duplex speech dialogue model that can be steered along tone, persona, and speed via instruction following, and introduces the SteerBench benchmark to evaluate such steerability. Supervised training yields a 44.5 % pass‑rate lift, and reinforcement‑learning fine‑tuning improves interruption handling and reduces pause barge‑ins, though reward hacking remains a…
The paper presents SNAP3D, a physics‑guided pipeline that converts a single image into a set of 3D parts that can be assembled without interpenetration. By using simulation‑driven connector placement and a new physics‑based evaluation, the method yields assemblies that are both geometrically accurate and stable enough for 3D printing.
The paper defines recursive self‑improvement (RSI) for AI, introduces the Headroom‑Closed Index to expose current LLM limits, and proposes a staged roadmap toward full RSI. It surveys application scenarios and outlines practical challenges.
COBRA‑Skills uses a contextual‑bandit loop to selectively evaluate and evolve LLM agent skills, achieving better performance with roughly half the evaluation cost of prior methods. The framework works with limited examples and remains robust across different agent setups and even when the target model creates its own skills.
Benchmark Radar is a continuously updated database and search engine that aggregates AI benchmark papers, datasets, and scores from 37 sources into a searchable catalog. It offers a web dashboard, CLI, and analysis of benchmark saturation to help LLM developers find and compare evaluations.
LLaDA‑UI is a 16.7 B‑parameter Mixture‑of‑Experts vision‑language model that uses block‑wise diffusion to generate GUI actions in parallel. It achieves state‑of‑the‑art performance on multiple GUI grounding and navigation benchmarks, beating larger Qwen models while keeping decoding latency low.
The paper proposes a Program‑Solve interface where LLMs emit Python code that a sandboxed executor runs, making arithmetic deterministic for clinical calculators. Experiments show the approach boosts a 32B model’s accuracy on a verified benchmark, but coverage and upstream extraction remain limiting factors.
Feyospace‑v1 presents a data‑centric training pipeline for cyber‑security agents, combining five systems (Choulea, SkyReal, Hongzwang, PSBreakup, Kreator) to generate and verify 164 k long‑context trajectories across diverse exploit environments. The resulting checkpoints improve baseline performance by ~24% on CyberGym and achieve a 63% verified success rate, ranking top among similarly‑sized op…
