Search posts, papers, and topics
85 posts · huggingface.co
Agora treats a Git repository as a shared, append‑only memory for autonomous research agents, recording each claim as an immutable commit in a DAG. In a 12‑day run with 13 language‑model workers it generated 1,703 reproducible contributions and closed 62 % of the gap to a trained GPT‑2‑124M, demonstrating that shared research state can accelerate discovery.
PPO critics in reinforcement learning for LLMs suffer from "Value Flattening," where predicted state values are too flat compared to actual values. This paper identifies the causes as an implicit variance penalty and redundant updates, and proposes SParse Proximal Policy Optimization (SP3O) to mitigate it by supervising only a few well-separated states.
LimiX-2 is a new tabular foundation model that replaces the usual target‑centric in‑context learning objective with a joint‑distribution objective via Contextual Mechanism Networks (CMNs). Trained on synthetic causal graphs using Context‑Conditional Masked Modeling, it outperforms prior tabular PFNs on TabArena, TALENT, and BCCO and can recover causal skeletons from attention patterns.
FLAT is a pre‑training framework that encodes images and text into a shared 1D token sequence with variable length via nested dropout, enabling the same embeddings for cross‑modal retrieval and generation. It attains state‑of‑the‑art scores on COCO and Flickr30K for captioning, retrieval, and text‑to‑image generation, and supports interpolation and zero‑shot composed retrieval.
Emergence World is a continuously running multi‑agent sandbox used to stress‑test frontier LLM‑based agents over weeks. Eight parallel worlds (seven homogeneous, one mixed) generated 850 k LLM calls and ~50 B tokens while agents pursued goals, used tools, and maintained persistent memory. The authors injected three adversarial events—prompt injection, misinformation, and private‑memory exposure—a…
ModAR is a world‑action model that autoregressively denoises multiple future modalities (depth, DINO features, point tracks) before predicting actions, letting each prediction condition on earlier outputs. It outperforms prior WAMs, achieving higher success rates with ~20× fewer training FLOPs and no pretraining.
EvoSkill‑GUI lets GUI agents revise their procedural skills on‑the‑fly without extra training by using a reflect‑revise‑reuse loop that edits skill packages during execution. The approach yields up to +16.2% improvement on MobileWorld and similar gains on AndroidWorld and OSWorld, and the evolved skills transfer to related tasks.
EventEgoHands++ adds an instance‑aware hand detector and an Adaptive Attention module to event‑camera egocentric hand‑mesh reconstruction, distinguishing left/right hands and modeling their interaction. The authors also release EEH‑R, a ~1 M‑frame real dataset, and report consistent gains over prior baselines on synthetic and real tests.
ScienceBuddy is an interactive workspace that converts researcher prompts, feedback, and execution traces into continual‑learning tasks for AI agents. It introduces a "recursive‑in‑recursive" self‑improvement loop that alternates harness refinement and model training, and showcases case studies across four scientific task families.
The authors train an anthropomorphic robotic hand to crawl, steer, and recover from falls using its fingers for both support and manipulation, via a reinforcement‑learning reward formulation tuned to the hand's asymmetry. Sim‑to‑real experiments show faster locomotion than quadruped‑style rewards and successful untethered tasks without onboard vision.
The paper introduces Fuse, a multi‑agent simulation that gives LLM assistants a verifiable ground‑truth task for social reasoning by hiding a target agent’s motive and letting a user‑mediated conversation infer it. Experiments on 12 LLMs show user mediation makes reasoning harder, models are biased by user framing, need more detail than humans, and longer chats don’t always help.
The paper surveys how foundation models are used across six roles in the game development lifecycle—from playing agents to design assistance and runtime adaptation. It highlights limited transferability due to game-specific interfaces and notes that evaluation is mature for bounded play but weak for adaptive and testing scenarios.
PhysStream introduces a two‑stage autoregressive video generator that uses online‑derived positional and tracking maps (structured scene memory) and sparse velocity‑increment signals to enable fine‑grained, physics‑grounded control of multi‑object tabletop scenes. It cuts motion distribution error by 33 % and trajectory error by 12 % versus strong baselines, and wins 85 % of human preference test…
XConf (eXperiential Confidence) is a novel method for estimating language model confidence by leveraging the model's accumulated experience from past graded episodes. It significantly outperforms existing methods like self-consistency in discrimination and calibration, at a fraction of the computational cost, across various tasks.
RiskChainBench is a new benchmark that pairs synthetic obfuscated message restoration inputs with human‑labeled local web environments, requiring models to both decode malicious instructions and investigate the linked site. Across ten models, restoration accuracy varies widely and web‑agent failures dominate the error budget.
ImpossibleRubrics introduces a 169‑task benchmark of “impossible” prompts, each with a formal oracle certificate defining what an honest answer can claim. The authors generate rubrics downstream and test them adversarially, finding that many rubric generators are exploitable (8‑36% of the time) and that a single generic rubric (“be decisive, penalize hedging”) is exploited 64% of the time, while…
Zing‑0.5 is a 5 B autoregressive world model that lets users control generated environments in real time using both keyboard actions and text prompts. The paper introduces unified action‑text conditioning, segment‑level teacher distillation, and a low‑cost streaming inference pipeline that runs at 24 FPS (832×480) for about $0.009 per minute, achieving 81 % overall and 88.5 % consistency on a nav…
EvolveTrade is a framework for self-evolving LLM trading agents that refines their tool-use policies (system prompts) based on past performance. It demonstrates improved Sharpe Ratio and Cumulative Return compared to agents with static policies across various market regimes.
The paper introduces Gavel, a method that extracts a frozen LLM's internal routing signal via two trained linear maps, eliminating the need to embed skill descriptions in the prompt. Experiments on Qwen3‑32B show up to 13.4‑point improvements on task benchmarks and higher skill‑use accuracy compared to larger retrieval‑based systems.
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…
Decoy Direction Optimization (DDO) is a post‑hoc weight‑editing defense for open‑weight LLMs that injects a high‑magnitude nonlinear decoy into MLP neurons, corrupting contrastive estimators used by Refusal Feature Ablation (RFA) attacks. The paper proves a spectral bound on the effect, evaluates DDO on six model families (including Llama‑3‑8B‑Instruct), and shows <10 % attack success rate (ASR)…
The paper decomposes transformer representation updates into parallel and perpendicular components, showing that parallel updates beyond the residual path are significant. Suppressing parallel updates during pretraining improves validation loss and downstream metrics.
Paper evaluates a standard 3‑D nnU‑Net on the new BraTS‑GoAT benchmark, training on 1,351 cases with five‑fold cross‑validation and test‑time mirroring. It reports Dice scores of 0.78/0.83/0.89 (ET/TC/WT) and shows a ~0.07 drop on heterogeneous validation, with limited benefit from ensembling or mirroring and failure linked to small, fragmented tumors.
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.
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.
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 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.
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.
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.
