277

Papers

10/10

Roadblocks Active

3

Connections

RadAgent breaks CT scan interpretation into discrete steps, routing each to one of 10 specialized analysis tools guided by a clinician-reviewed checklist, then trained end-to-end with reinforcement learning using a composite reward. Compared to a direct-prediction 3D vision-language model baseline, it improves macro-F1 by 6 points (36% relative) and adversarial robustness by 24.7 points (42% relative), while achieving measurable output faithfulness the baseline scores zero on. This matters because it demonstrates that tool decomposition — rather than larger models — can simultaneously improve accuracy, robustness, and interpretability in high-stakes medical AI.

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

ReadSaveConnections

Across 18 vision-language models and three public benchmarks, confidence tracking over chain-of-thought steps reveals that models commit to an answer early and reinforce it rather than revising it — a pattern the authors call answer inertia. Even reasoning-trained models are consistently steered by misleading textual cues even when visual evidence alone is sufficient to answer correctly. This is a safety-relevant finding because models can produce reasoning traces that look corrective while actually following spurious correlations established in their first few tokens.

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

ReadSaveConnections

MM-AQA is a 2,079-sample benchmark built by transforming answerable questions into unanswerable ones along two axes — whether the visual modality is necessary, and whether sufficient evidence exists — to test whether vision-language models know when to abstain. Under standard prompting, frontier VLMs almost never abstain, and simple confidence baselines outperform them on abstention decisions; multi-agent systems improve abstention but trade off answer accuracy. The practical implication is that deployed multimodal AI systems currently lack reliable self-knowledge about their own uncertainty, which is a prerequisite for safe deployment.

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

ReadSaveConnections

SWE-TRACE addresses the difficulty of training software engineering agents on long-horizon tasks through three stages: filtering 140K candidate trajectories to 60K token-efficient examples via LLM-guided oracle verification, then training with reinforcement learning using a rubric-based process reward model (PRM) that provides dense step-level feedback rather than sparse outcome signals. Critically, the trained PRM is then reused at inference time to prune low-value action candidates, converting training infrastructure into a test-time scaling mechanism. This dual-use pattern — where the same reward model improves both training stability and inference efficiency — is a generalizable architectural insight for agentic RL.

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

ReadSaveConnections

OpenMobile generates training data for mobile UI agents by first building a global memory map of the environment through exploration, then synthesizing trajectories using a policy-switching strategy that alternates between the learner model and an expert — deliberately capturing error-recovery behavior that pure imitation learning misses. Agents trained on this data reach 51.7% (Qwen2.5-VL) and 64.7% (Qwen3-VL) success rates on AndroidWorld, with gains attributed to broad functional coverage rather than benchmark-specific tuning. The error-recovery data collection mechanism is a practical solution to a known gap in behavior cloning for GUI agents.

██████████ 0.8 agent-tool-use Preprint

ReadSaveConnections

VRUBench is a text-only benchmark requiring models to predict what they would observe after a sequence of rotational steps — a task humans solve with 100% accuracy but on which both LLMs and VLMs fail substantially. Layer-wise probing shows models do encode viewpoint position information in hidden states, but cannot bind that position to the corresponding observation content; hallucination emerges specifically in the final layers when processing these spatial tasks. The targeted finding — that the bottleneck is positional binding rather than spatial encoding — suggests that fine-tuning only the identified attention heads responsible for this binding is more effective than full fine-tuning.

██████████ 0.8 interpretability Preprint

ReadSaveConnections

FORGE applies a Dynamic Forest of Agents architecture to binary vulnerability analysis, decomposing tasks across parallel agents while bounding each agent's context window to prevent the performance degradation that occurs when long analysis traces overwhelm a single LLM context. Reasoning-action-observation feedback loops let agents incrementally build evidence across multiple tool calls rather than attempting single-pass analysis. The system identifies 1,274 vulnerabilities across 591 unique binaries at 72.3% precision, demonstrating that context management through task decomposition is a practical lever for agentic reliability in technically complex domains.

██████████ 0.8 agent-tool-use Preprint

ReadSaveConnections

Standard Vision-Language-Action models predict actions directly, but planning over long action-space horizons is intractable because the probability of producing a feasible trajectory decays exponentially with horizon length. The World-Value-Action (WVA) model instead learns structured latent representations of future trajectories conditioned on visual observations and language instructions, then uses latent-space inference to reshape the search distribution toward feasible regions before committing to actions. This implicit planning mechanism addresses a fundamental scaling limitation of direct action prediction for robotic tasks requiring extended decision sequences.

██████████ 0.8 embodied-ai Preprint

ReadSaveConnections

This paper presents the first system where LLM agents iteratively rewrite the source code of ABC — a widely used logic synthesis tool in chip design — operating across the full integrated codebase while preserving the single-binary execution model. Through continuous feedback loops comparing synthesized circuit quality before and after code modifications, the system discovers optimizations that exceed human-designed heuristics. This is a concrete demonstration that LLM agents can improve specialized engineering software autonomously, with implications for how AI might accelerate tool development in other technically constrained domains.

██████████ 0.8 agent-tool-use Preprint

ReadSaveConnections

The Data Intelligence Layer (DIL) treats LLMs, the web, and users as first-class queryable data sources alongside traditional relational databases, defining query interfaces for each and allowing a data planner to transform natural-language user requests into executable query plans that unify relational operators with multi-modal retrieval. The architecture addresses the practical limitation that real user information needs typically require iterative, multi-source access patterns that no single SQL query can satisfy. While the paper is a system demonstration without controlled benchmarks, the unified query planning abstraction is a substantive architectural contribution to multi-source AI application design.

██████████ 0.8 multimodal-understanding Preprint

ReadSaveConnections