297

Papers

10/10

Roadblocks Active

0

Connections

MaxProof trains a proof-generation model with reinforcement learning while pairing it with a 'defense-in-depth' generative verifier designed to reject incorrect proofs with very low false-positive rates. At test time, many candidate proofs are generated and a tournament-selection process picks winners, allowing compute to be traded for accuracy. The system scores above the human gold-medal threshold on both IMO 2025 and USAMO 2026, making it the strongest reported result on competition-level formal mathematics and a concrete demonstration that reliable verification is the key bottleneck to unlock, not just generation.

██████████ 1.0 reasoning-reliability Preprint

ReadSaveConnections

TerraBench is a benchmark of 403 executable agentic tasks spanning weather forecasting, climate simulation, and geospatial processing, testing whether AI agents can coordinate specialist scientific tools — not just language — to answer Earth-science questions. The benchmark separates two failure modes: tool-use proficiency (did the agent call the right tools correctly?) and final numerical accuracy, exposing cases where agents produce plausible-sounding answers via wrong workflows. Current LLMs score significantly below what domain scientists need, pinpointing workflow coordination and precise tool parameterisation as the limiting factors.

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

ReadSaveConnections

When LLMs generate product recommendations by retrieving live web evidence, injecting a single fake-product page into a top-3 result bundle can fool all 12 tested models up to 27% of the time; replacing all three retrieved pages raises that to 73.8%. The attack works because models treat retrieved text as authoritative, particularly in product categories where the model has weak prior knowledge. Three defenses (skepticism prompting, prior-consensus filtering, cross-document consensus) reduce but do not eliminate the vulnerability, showing that RAG-based recommendations inherit a structural trust problem with the open web.

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

ReadSaveConnections

This structured review maps how AI hallucination manifests across five medical imaging modalities — CT, MRI, PET/SPECT, ultrasound, and digital pathology — identifying clinically dangerous failure modes such as fabricated anatomical structures and incorrect laterality. A counter-intuitive finding is that general-purpose foundation models outperform medically fine-tuned models on hallucination benchmarks, suggesting that narrow domain fine-tuning introduces overfitting that increases confabulation rather than reducing it. The paper also maps these failure modes to current FDA regulatory guidance, which is useful for anyone building or approving clinical AI systems. Note: the paper is a narrative review, and some reported p-values lack transparent statistical basis.

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

ReadSaveConnections

Most CAD-generation models are one-shot: they produce a design from a prompt and stop. IterCAD treats CAD creation as a multi-turn loop in which an agent writes code, executes it in a CAD sandbox, inspects the result, and repairs errors — the same way a human engineer would. A geometry-aware reinforcement learning stage with 'viable-prefix masking' (only penalising steps that were geometrically reachable but wrong) significantly improves both code executability and shape accuracy. The approach demonstrates that closed-loop, self-correcting agents can handle structured technical domains, not just open-ended text tasks.

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

ReadSaveConnections

InterleaveThinker adds multi-turn planning and self-critique to any frozen image generator without retraining it, via a planner agent that sequences the generation steps and a critic agent that detects deviations from the plan and rewrites instructions. The critic is trained with GRPO reinforcement learning using a dual reward that scores both per-step quality and final accuracy, allowing it to learn from multi-step trajectories using only single-step RL updates. This is practically useful because it decouples reasoning quality from the image generator itself — improvements in planning transfer to any generator.

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

ReadSaveConnections

VLA models trained on household robot data consistently fail in scientific labs: transparent liquids, complex instruments, and rigid multi-step protocols break the assumptions baked in by everyday-object training data. LabVLA addresses this by pretraining an action-tokenisation layer on a chemistry-lab corpus, then using flow-matching posttraining with a separate action expert module insulated from the vision-language backbone — preventing the robot policy from degrading the language understanding. The model sets a new best on the LabUtopia benchmark under both in-distribution and held-out conditions, identifying data scarcity as the remaining central bottleneck.

██████████ 0.8 embodied-ai Preprint

ReadSaveConnections

Fact verification pipelines have multiple stages (claim decomposition, evidence retrieval, reasoning, verdict), but most RL training only rewards the final verdict, leaving intermediate steps without a learning signal — a sparse supervision problem. ProFact introduces process-aware rewards that assign credit to each intermediate stage, enabling end-to-end optimisation of the full trajectory rather than just the final answer. The result is both higher verification accuracy and lower inference cost compared to strong baselines, making this approach relevant beyond fact-checking to any multi-stage agent pipeline where delayed feedback is a bottleneck.

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

ReadSaveConnections

Vision-language models can localise objects well in a single pass, but when asked to iteratively refine bounding box predictions by looking at their own rendered output, performance collapses catastrophically (79.6% to 48.7% accuracy) without special training. IVT teaches self-correction by generating 2,400 synthetic training examples where the model's own wrong predictions are the starting point, with a teacher model providing corrective reasoning traces, followed by RL fine-tuning rewarding IoU improvement. The key finding — that self-correction is a learnable skill acquirable with only 2,400 samples — suggests it can be grafted onto existing VLMs cheaply, and may generalise to other modalities where iterative refinement matters.

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

ReadSaveConnections

MÖVE evaluates 39 LLMs across both performance tasks (summarisation, QA, topic extraction) and governance dimensions (hallucination, energy use, transparency, political value alignment) in German public administration contexts. The headline finding is that no single model dominates across all criteria, and model size is a poor predictor of quality — a result with direct procurement implications for government agencies that currently use scale as a proxy for reliability. The inclusion of governance metrics alongside accuracy is the methodological contribution, offering a template for how regulated-sector benchmarks should be structured.

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

ReadSaveConnections