292

Papers

10/10

Roadblocks Active

0

Connections

SpatialClaw gives a vision-language model a stateful Python environment preloaded with perception tools (depth estimation, segmentation, 3D reconstruction), letting it write and iteratively revise code rather than issuing one-shot commands. Evaluated across 20 spatial reasoning benchmarks without any task-specific training, it outperforms prior spatial agents by 11.2 percentage points (59.9% average accuracy). The key insight is that intermediate feedback between code steps — rather than single-pass execution — is what allows the model to correct its own spatial errors, which matters because spatial reasoning is a well-documented weak point of current multimodal AI.

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

ReadSaveConnections

Current AI agents trying to control professional software like CAD tools via simulated mouse clicks and screen reading fail almost completely — frontier models score near zero on CAD tasks with GUI-based interaction. ComAct replaces this with the Windows Component Object Model (COM) interface, turning software control into deterministic program synthesis where the agent generates API calls rather than visual actions. This matters because it eliminates the fragility of visual control (pixel changes, window state) and opens a path to reliable automation of complex professional workflows that GUI agents cannot yet handle.

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

ReadSaveConnections

Static web-search benchmarks can be gamed by models that simply memorize answers seen during training, making high scores misleading about actual retrieval ability. EvoBrowseComp addresses this by continuously generating questions from live post-January-2026 web content that no model could have memorized, using a three-agent pipeline to build 800 verified bilingual questions. The dataset is publicly released, giving the community a contamination-resistant way to measure whether search agents are genuinely reasoning or just recalling.

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

ReadSaveConnections

When AI systems use live web search to make product recommendations, injecting a single fake-product description into the retrieved results can fool tested models up to 27% of the time; polluting all three top results raises that to 73.8%. The FORGE benchmark tests this across 12 commercial and open-weight LLMs on 225 real products in 15 categories, with the benchmark code and frozen evidence bundles publicly released. This is a concrete, measurable robustness failure in retrieval-augmented systems that is already relevant to deployed products.

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

ReadSaveConnections

Vision-language-action models trained on household robot data fail to transfer to laboratory settings because labs have specialized instruments, transparent liquids, and rigid protocol sequences not found in consumer environments. LabVLA addresses this with a two-stage training recipe — first teaching the model an action vocabulary via token pretraining, then adding fine-grained continuous motion control via flow matching — and a synthetic data engine built on NVIDIA Isaac Sim covering 16 robot types. The practical implication is that scientific lab automation requires purpose-built training pipelines rather than transfer from general robotics data.

██████████ 0.8 embodied-ai Preprint

ReadSaveConnections

EurekAgent argues — and demonstrates — that the limiting factor in autonomous AI research agents is not how the agent reasons but how its execution environment is designed: isolated sandboxes, bounded permissions, and Git-based artifact sharing. The system achieves state-of-the-art results on mathematics and machine learning tasks, and found new record circle-packing configurations for under $11 in API costs. If the environment-first framing holds up to scrutiny, it reorients research investment away from agent architecture toward infrastructure design.

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

ReadSaveConnections

Most medical AI is English-only, leaving over a billion Indian-language speakers without reliable AI health tools. ArogyaSutra builds a multilingual medical reasoning system using an actor-critic multi-agent loop — one agent proposes diagnoses, another critiques and refines — combined with visual grounding tools (zoom, edge detection, depth analysis) for six imaging modalities across 21 clinical domains. The ArogyaBodha dataset covering 31 body systems in 8 languages is released publicly, providing a concrete resource for low-resource medical AI development.

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

ReadSaveConnections

MÖVE evaluates 39 LLMs across both performance tasks (summarization, question answering, topic extraction) and governance criteria (hallucination, energy use, transparency, constitutional value alignment) specifically for German-language government applications. The headline finding — no single model wins across all criteria, and model size alone predicts little — is a useful corrective for procurement decisions that rely on global leaderboard rankings. The benchmark also includes a self-evaluation of its own statistical reliability, which is uncommon and adds methodological credibility.

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

ReadSaveConnections

Most research agents treat a scientific paper as just its abstract and a flat list of citations, discarding figures, tables, equations, and the semantic structure of arguments. Agents-K1 builds a full pipeline that parses entire papers into structured knowledge graphs capturing entities, multimodal evidence, and typed relationships between ideas, then uses these graphs for multi-hop scientific reasoning. The approach targets a real bottleneck: agents that cannot read papers deeply cannot do meaningful literature synthesis or hypothesis generation.

██████████ 0.7 reasoning-reliability Preprint

ReadSaveConnections

Standard AI-driven CAD generation runs in a single pass, producing code that frequently fails to compile or violates geometric constraints. IterCAD makes the process iterative: a multimodal agent generates CAD code, runs it in a live sandbox, observes the result, and revises — trained with reinforcement learning that specifically rewards maintaining valid geometry prefixes even when later code fails. This closed-loop approach improves both code executability and geometric precision over open-loop baselines, pointing toward a general pattern where executable feedback loops matter more than bigger models for structured code generation.

██████████ 0.7 agent-tool-use Preprint

ReadSaveConnections