290

Papers

10/10

Roadblocks Active

0

Connections

Semia audits the safety of AI agent skills — the building blocks that let agents take actions in the world — by translating prose-defined skill descriptions into a structured logical representation and then running automated security checks. Tested on 13,728 real-world skills from public marketplaces, it found that more than half carry at least one critical risk, achieving 97.7% recall while outperforming both static analyzers and LLM-only auditors. This matters because AI agent ecosystems are scaling rapidly and conventional tools cannot read the natural-language conditions that govern when a skill fires, leaving a large and largely invisible attack surface.

██████████ 0.9 agent-tool-use Preprint

ReadSaveConnections

This paper introduces IVLR, a policy framework that generates an explicit chain of alternating text subgoals and visual keyframe predictions before a robot begins acting, then conditions closed-loop motor control on that pre-computed plan. On the LIBERO-Long manipulation benchmark, IVLR reaches 92.4% success versus 37.7% without any traces, and ablations show both the text and visual components are necessary — dropping either costs roughly 25 percentage points. The mechanism is important because it shows that structured intermediate representations, not just end-to-end learning, are the key lever for scaling robot manipulation to longer tasks.

██████████ 0.9 embodied-ai Preprint

ReadSaveConnections

AUTOMAT is a benchmark of 85 expert-curated scientific claims from real computational materials science publications, used to test whether LLM-based coding agents can actually reproduce research results by writing and running the required simulation workflows. Current best agents succeed only 54.1% of the time, with the hardest condition — reconstructing a workflow from the paper text alone — being the primary failure mode. This is a concrete quantification of a widely suspected problem: coding ability does not transfer cleanly to the messy, tool-heavy, domain-specific workflows that constitute real scientific computing.

██████████ 0.9 agent-tool-use Preprint

ReadSaveConnections

This paper diagnoses why robot policies built on vision-language-action models fail when the scene looks slightly different from training: they latch onto spurious visual correlations rather than task-relevant causes. The authors propose an Interventional Significance Score (ISS) that masks image regions and measures the causal impact on action outputs, plus a Nuisance Mass Ratio (NMR) that predicts out-of-distribution generalization without running the robot. The result is an interpretability toolkit that connects what the model looks at to whether it will actually generalize — a practical diagnostic for deployment safety.

██████████ 0.9 embodied-ai Preprint

ReadSaveConnections

Current video-generation-style world models can produce visually plausible futures but break down in physical consistency and long-horizon stability, which limits their usefulness for training robot controllers. This paper proposes encoding scene state as a latent phase space structured by Hamiltonian dynamics — borrowing from classical physics — so that the learned representations inherently respect conservation laws and support controllable, stable rollouts. The argument is that the bottleneck in world modeling has shifted from 'does it look real' to 'does it obey physics', and structured inductive biases are the most data-efficient path forward.

██████████ 0.9 embodied-ai Preprint

ReadSaveConnections

RunAgent converts natural-language plans into a structured agentic language with explicit control constructs (IF, GOTO, FORALL), then enforces step-by-step correctness using automatically derived constraints rather than hoping the LLM stays on track. It outperforms both raw LLM baselines and the state-of-the-art PlanGEN method on Natural-plan and SciBench datasets. The core insight is that reliability in multi-step agentic tasks requires deterministic execution scaffolding, not more capable language models alone.

██████████ 0.9 reasoning-reliability Preprint

ReadSaveConnections

RadLite fine-tunes two small language models (3–4 billion parameters) using LoRA on 162,000 labeled examples spanning nine radiology tasks, achieving accuracy gains of 53–89 percentage points over zero-shot baselines while running entirely on CPU hardware. Qwen2.5-3B excels at structured generation tasks while Qwen3-4B dominates extractive ones, and a task-routed ensemble of both achieves best overall performance. The practical significance is that hospital systems without GPU infrastructure can now deploy competitive radiology NLP without cloud dependency or large model licensing.

██████████ 0.9 efficiency-scaling Preprint

ReadSaveConnections

As retrieval-augmented generation (RAG) and agentic search increasingly make LLMs the primary consumers of retrieved documents rather than humans, the failure mode shifts: LLMs have fixed attention budgets and are uniquely harmed by noisy, unverifiable context. This survey argues that 'denoising' — maximizing the density of useful, verifiable evidence within the context window — is the central engineering challenge in modern information retrieval systems. It provides a taxonomy of denoising techniques across every stage of the RAG pipeline, offering a unifying frame for a fragmented literature.

██████████ 0.9 hallucination-grounding Preprint

ReadSaveConnections

Odysseus trains a vision-language model (Qwen3-VL-8B) to play Super Mario Land for over 100 consecutive turns using a PPO variant with a lightweight turn-level critic, which substantially improves training stability over critic-free methods like GRPO. Pretrained VLMs bring strong action priors that reduce sample complexity compared to training classical deep RL agents from scratch, showing that language-world knowledge transfers usefully to game-playing. This matters because 100+ turn decision-making is a prerequisite for real-world agentic deployment and most existing RL-for-VLM work stops at single or few-step tasks.

██████████ 0.9 multimodal-understanding Preprint

ReadSaveConnections

STARE attacks text-to-image models by treating their multi-step denoising process as an attack surface, using hierarchical reinforcement learning to concentrate harmful content into specific phases of image generation — a phenomenon the authors call Optimization-Induced Phase Alignment. This approach improves attack success rates by 68% over prior baselines, revealing that adversarial optimization fundamentally changes how toxicity is distributed across the generation trajectory. Understanding this attack mechanism is a prerequisite for designing defenses that intercept harm before the final image is produced.

██████████ 0.8 alignment-safety Preprint

ReadSaveConnections