285

Papers

10/10

Roadblocks Active

0

Connections

An AI system called AlphaProof Nexus pairs a large language model with the Lean formal proof checker to autonomously search for and verify mathematical proofs, solving 9 of 353 open Erdős problems and 44 of 492 OEIS conjectures. The key mechanism is that Lean rejects incorrect proofs unconditionally, so the LLM's role is to generate candidate proof sketches while the verifier enforces correctness — eliminating the hallucination problem for this domain. This is the most direct demonstration to date that AI can contribute genuinely new results to professional mathematics, though the core pipeline remains proprietary.

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

ReadSaveConnections

Applying reinforcement learning fine-tuning (GRPO) to a small 4B-parameter model inside a live Microsoft Excel environment nearly doubles its task success rate on SpreadsheetBench — from 12.0% to 23.4% Pass@1 — compared to prompting general-purpose LLMs. The training data is collected automatically by an agent that scrapes online forums for real spreadsheet problems, making the pipeline self-sustaining without expensive human annotation. This shows that domain-specific RL in a real execution environment is far more effective than scale or prompting alone for tool-use agents.

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

ReadSaveConnections

AgroTools evaluates 13 multimodal LLMs on 539 agricultural questions requiring real tool use — selecting the right tool, generating valid arguments, recovering from execution errors, and synthesising a final answer — across 1,097 domain images and 14 executable tools. The benchmark exposes that current models fail not just at final answers but at intermediate steps like argument generation and execution recovery, which purely outcome-based benchmarks would miss entirely. This dual-view process-plus-outcome evaluation methodology is reusable beyond agriculture for any tool-augmented agent setting.

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

ReadSaveConnections

SpaceDG synthesises nine types of realistic visual degradation (motion blur, low light, adverse weather, lens distortion, compression artefacts, etc.) directly into 3D scene renderings of ~1,000 indoor environments, producing ~1 million VQA pairs to test whether multimodal models can reason spatially under real-world imaging conditions. Testing 25 state-of-the-art models reveals a consistent and substantial performance drop under degradation that clean-image benchmarks completely hide. A fine-tuned 8B model trained on degraded data partially recovers this gap, suggesting the problem is addressable with targeted data rather than larger models.

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

ReadSaveConnections

SegCompass uses a Sparse Autoencoder to project both the model's chain-of-thought reasoning tokens and the visual image tokens into the same high-dimensional sparse concept space, then selects salient concepts to drive a segmentation mask decoder. Because sparse activations are human-inspectable by design, this directly links what the model is 'thinking' to what it is drawing in the output mask, achieving interpretability without sacrificing accuracy. The strong correlation found between concept quality and mask accuracy suggests interpretability and performance are complementary here, not in tension.

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

ReadSaveConnections

DecentMem gives each agent in a multi-agent system its own dual-pool memory — one pool for exploiting known-good strategies, one for exploring new ones — updated continuously via LLM-as-a-judge feedback without any central coordinator. This architecture eliminates the privacy, bottleneck, and homogenisation problems of shared memory while still guaranteeing that every useful experience can eventually propagate across agents, with cumulative regret provably matching the O(log T) lower bound of stochastic bandits. Empirically it delivers up to 23.8% accuracy gains over centralised baselines and 52.5% over no-memory agents across diverse benchmarks.

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

ReadSaveConnections

AADvark is reported as the first agentic system that can generate 3D CAD assemblies containing moving parts — pistons, pendulums, scissors — with multiple degrees of freedom, a capability previous agent-design systems lacked entirely. The system uses an iterative visual feedback loop to compensate for LLMs' known weaknesses in spatial reasoning, letting the model self-correct geometry errors by inspecting rendered outputs. This extends agentic CAD generation from static objects to functional mechanisms, which is a prerequisite for AI-assisted mechanical engineering.

██████████ 0.9 agent-tool-use Peer-reviewed

ReadSaveConnections

Pre-VLA adds a lightweight verification head to a robot's Vision-Language-Action model that predicts, before executing an action chunk, whether that action is likely to be safe and successful — raising closed-loop task success from 30.79% to 37.62% across four LIBERO robotics suites. The verifier runs in 183.9 ms per chunk, making real-time preemptive rejection practical without slowing the robot. This approach moves robot safety from post-hoc detection toward anticipatory prevention, which is important as VLA models are deployed on physical hardware where failures are costly.

██████████ 0.8 embodied-ai Preprint

ReadSaveConnections

ChronoMedKG distils 13 million raw LLM extractions down to 460,497 high-confidence triples by requiring each disease–symptom association to carry a temporal component (onset window, progression stage), a traceable PubMed citation, and a multi-model consensus credibility score. The resulting graph covers 13,431 diseases and provides temporal grounding for 1,657 rare diseases absent from Orphanet, achieving 92.7% agreement with the Orphadata reference. Adding temporal structure to biomedical knowledge graphs is important because many clinical reasoning errors stem from conflating which symptoms appear early versus late in a disease course.

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

ReadSaveConnections

AwareVLN introduces a structured self-awareness module into a vision-language navigation agent that explicitly reasons about where the agent currently is and how far it has progressed toward the goal — using only monocular RGB images, without depth sensors or 3D maps. The model selectively activates this structured reasoning at key navigation decision points rather than every step, reducing overhead while improving performance above prior state-of-the-art methods on Habitat simulator benchmarks. Demonstrating that explicit state-awareness reasoning (rather than implicit learned representations) improves navigation is useful for deploying agents in unstructured real-world environments.

██████████ 0.8 embodied-ai Preprint

ReadSaveConnections