276

Papers

11/11

Roadblocks Active

3

Connections

RadAgent breaks chest CT interpretation into a sequence of specialized tool calls — organ segmentation, disease classification, slice extraction — guided by a clinician-reviewed checklist and trained with reinforcement learning. It outperforms a direct 3D vision-language model by 6 macro-F1 points on accuracy and 24.7 points on robustness to adversarial inputs, and adds faithful reasoning traces the baseline cannot produce at all. This matters because it shows modular, tool-using agents can genuinely improve clinical reliability over end-to-end models, not just benchmark scores.

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

ReadSaveConnections

Across 18 vision-language models and three benchmarks, this study tracks how model confidence evolves during chain-of-thought reasoning and finds that models lock in early predictions and reinforce them rather than revising when evidence contradicts the initial guess — a pattern the authors call 'answer inertia'. Models are also consistently swayed by misleading text even when the visual evidence alone would be enough to answer correctly. This is a structural problem: reasoning steps that look deliberate are often just elaborate justifications for snap judgments.

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

ReadSaveConnections

The authors construct MM-AQA, a 2,079-instance benchmark of unanswerable multimodal questions built by systematically removing visual evidence or modality relevance from MMMU and MMLongBench-Doc. Frontier vision-language models almost never abstain under standard prompting — and simple confidence baselines beat standard prompting for knowing when to stay silent. Multi-agent debate systems help, but introduce a direct accuracy-abstention trade-off, revealing that current architectures have no principled mechanism for recognizing the limits of their own knowledge.

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

ReadSaveConnections

SWE-TRACE addresses a core failure mode in AI software engineering agents: sparse, end-of-task rewards make training unstable for long multi-step tasks. It combines dense intermediate feedback via a Rubric-Based Process Reward Model with a memory-augmented RL loop, and distills a 60K-instance training corpus biased toward the shortest correct solution paths. The heuristic test-time scaling component reuses the learned reward signals at inference to prune bad actions without adding latency. Reproducibility is low as no code or data is released, so results should be treated as preliminary.

██████████ 0.8 reasoning-reliability Preprint

ReadSaveConnections

UniDoc-RL trains a vision-language model agent to retrieve and read document images through a coarse-to-fine hierarchy: first locating the right page region, then cropping and zooming for precision, using GRPO reinforcement learning with separate rewards at each stage. This staged reward design avoids the need for a separate value network while achieving up to 17.7% gains over prior visual RAG methods. Code and data are publicly released, making this one of the more reproducible results in today's batch.

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

ReadSaveConnections

OpenMobile generates synthetic training data for mobile phone agents by first building a memory of what each app can do through exploration, then synthesizing multi-step task instructions grounded in that memory. A policy-switching rollout alternates between a learner model and an expert to collect error-recovery trajectories that standard imitation learning misses. Fine-tuned on this data, Qwen3-VL-8B reaches 64.7% success on AndroidWorld — competitive results on an open model without proprietary fine-tuning data. Both code and dataset are publicly released.

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

ReadSaveConnections

FORGE applies a forest of LLM agents to binary vulnerability analysis, where each agent runs a reasoning-action-observation loop that incrementally builds evidence rather than attempting full analysis in a single pass. Across 591 binaries, the system identifies 1,274 vulnerabilities at 72.3% precision — a domain where conventional tools struggle with the open-ended reasoning required. The dynamic tree structure bounds per-agent context while allowing parallel exploration, which is a practical architectural solution to the long-context problem in agentic systems.

██████████ 0.8 reasoning-reliability Preprint

ReadSaveConnections

Using a synthetic text-only benchmark (VRUBench) and mechanistic analysis via linear probing and causal head interventions, this study finds that LLMs and VLMs encode viewpoint orientation information in their hidden states but consistently fail to bind that information to the correct object when answering questions — a binding failure that manifests as hallucination in the final transformer layers. Humans solve these tasks at 100%; models fall far short. Targeted fine-tuning of only the identified key attention heads partially repairs the deficit without broad performance degradation.

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

ReadSaveConnections

SGA-MCTS uses Monte Carlo Tree Search offline to explore solution spaces and distills successful trajectories into abstract State-Goal-Action atoms — reusable reasoning templates stripped of domain-specific details. At inference, a hybrid symbolic-semantic retrieval system re-grounds these atoms into the current problem context, allowing frozen open-weights models to match reported SOTA performance without any task-specific fine-tuning. The claim that this rivals GPT-5-class systems is striking but unverified pending full paper and code release.

██████████ 0.8 reasoning-reliability Preprint

ReadSaveConnections

This paper demonstrates a multi-agent LLM system that iteratively rewrites portions of ABC, a 1.2-million-line C codebase for logic synthesis used in chip design, compiling and formally verifying each change before accumulating improvements across generations. The system discovers optimizations beyond human-designed heuristics on standard benchmarks, which is a meaningful result because logic synthesis quality directly affects chip performance and power. Reproducibility is uncertain — the authors use a significantly expanded private fork of ABC and have not confirmed artifact release.

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

ReadSaveConnections