221

Papers

10/10

Roadblocks Active

3

Connections

Across 18 vision-language models, researchers tracked how model confidence evolves during chain-of-thought reasoning and found that models commit to early predictions and rarely revise them — a phenomenon they call 'answer inertia.' Even when visual evidence was sufficient to answer correctly, misleading text cues consistently overrode it. This matters because it shows that reasoning-trained models are more corrective but not fundamentally fixed: the bias toward text over vision persists and could cause systematic errors in safety-critical applications like medical imaging or autonomous navigation.

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

ReadSaveConnections

SGA-MCTS separates the expensive search for good reasoning paths (done once, offline via Monte Carlo Tree Search) from fast execution (done at inference using reusable 'State-Goal-Action' atoms distilled from successful trajectories). The atoms strip domain-specific surface details while preserving reusable causal logic, allowing frozen open-weight models to reportedly match frontier systems like GPT-5 without any fine-tuning. This is significant because it suggests capable reasoning agents may not require either large proprietary models or task-specific training — just a well-designed replay mechanism.

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

ReadSaveConnections

Using a new text-only benchmark (VRUBench) requiring models to predict observations after multi-step spatial rotations, the authors find both LLMs and VLMs score near chance while humans achieve 100%. Crucially, layer-wise probing shows the models do encode spatial orientation in their hidden states — the information is there — but head-wise causal analysis reveals a binding failure: the encoded orientation is not correctly linked to the predicted observation in final layers. This mechanistic distinction between representation failure and retrieval failure is valuable for interpretability because it identifies where to intervene, not just that models fail.

██████████ 0.9 interpretability Preprint

ReadSaveConnections

UniDoc-RL trains a single vision-language model agent via reinforcement learning to handle the full document question-answering pipeline — retrieving relevant documents, reranking candidates, and cropping precise visual regions — rather than chaining separate specialized models. A dense reward scheme provides stage-specific feedback (NDCG-based for retrieval, IoU-based for cropping) at each decision point, which avoids the sparse reward problem that makes end-to-end RL on sequential tasks unstable. The result matters because compound errors from chained retrieval-then-reasoning pipelines are a major practical failure mode, and a unified agent with intermediate rewards is a principled fix.

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

ReadSaveConnections

The MM-AQA benchmark (2,079 samples) tests whether vision-language models can recognize when a question is unanswerable — by converting answerable instances into unanswerable ones through structured visual or evidential transformations and validating them with human raters. Standard VLMs almost never voluntarily abstain; simple confidence baselines beat default prompting, and multi-agent systems improve abstention at a cost to accuracy on answerable questions. The finding that sequential multi-agent designs match more complex iterative ones points to miscalibration as the root cause rather than insufficient reasoning depth, which narrows the target for fixes.

██████████ 0.8 hallucination-grounding Preprint

ReadSaveConnections

LongAct observes that transformers processing long inputs develop sparse, high-magnitude activation patterns in query and key vectors, and uses this natural structure to selectively update only the most salient weights during reinforcement learning fine-tuning — skipping updates to weights that remain inactive on long contexts. This saliency-guided sparse update outperforms uniform gradient updates while reducing compute. It matters practically because training long-context RL agents is very expensive, and the finding that the model's own internal structure can guide efficient training suggests a general principle rather than a task-specific hack.

██████████ 0.8 long-context Preprint

ReadSaveConnections

SWE-TRACE addresses the challenge of training AI agents on long-horizon software engineering tasks (real bug fixes spanning many steps) by combining three components: curated oracle-verified trajectories filtered from 140K synthetic instances, a rubric-based process reward model that provides dense intermediate feedback beyond binary pass/fail, and heuristic test-time scaling that reuses the reward model to prune candidate actions without adding latency. Evaluated on SWE-bench Verified, the key insight is that sparse outcome rewards alone are insufficient for reliable long-horizon agents — rubric-structured intermediate signals are necessary. Confidence is limited because the experimental section is not visible in the preprint.

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

ReadSaveConnections

ADAPT introduces DynAfford, a benchmark built on the AI2-THOR simulator with 2,628 expert demonstrations across 57 scenes, designed to test whether embodied agents can track which objects they can actually interact with as environments change. All existing state-of-the-art planners (MOCA, FILM, SayCan, etc.) fail significantly under these dynamic conditions, but augmenting them with a plug-and-play ADAPT module — which uses a LoRA-finetuned VLM for affordance inference — recovers performance and outperforms GPT-4o on affordance judgment. This matters because real deployment environments constantly change, and agents that assume static affordances will fail reliably in practice.

██████████ 0.8 embodied-ai Preprint

ReadSaveConnections

RadAgent is trained with reinforcement learning to orchestrate 10 specialized chest CT analysis tools — segmentation, classification, measurement — guided by a clinician-reviewed diagnostic checklist, rather than generating radiology reports end-to-end from raw images. Compared to a baseline 3D vision-language model (CT-Chat), it improves macro-F1 by 36% and robustness under adversarial input perturbations by 42%, and uniquely produces traceable, step-by-step reasoning that the baseline cannot. The clinical significance is that tool-orchestrated agents may offer a practical path to explainable medical AI, where regulators and clinicians require reasoning traces, not just predictions.

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

ReadSaveConnections

IE-as-Cache repurposes information extraction as a dynamic working memory during multi-step reasoning: a query-driven extractor builds a compact, continuously updated cache of relevant facts that the reasoning agent consults at each inference step, filtering accumulated noise from long reasoning chains. The approach is model-agnostic and improves accuracy across multiple LLMs on multi-hop reasoning benchmarks. It matters because long chains-of-thought tend to drift as irrelevant context accumulates, and a structured cache that explicitly tracks extracted facts addresses this failure mode without modifying the underlying model.

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

ReadSaveConnections