283

Papers

11/11

Roadblocks Active

0

Connections

When multimodal AI agents handle long conversations involving images, the sheer number of visual tokens grows unmanageable and cross-turn references become unreliable. This paper externalizes visual information into a structured external memory with selective retrieval, letting an 8B model achieve 91.4% retrieval accuracy across 20-turn sessions — outperforming 32B baselines by 8.2 percentage points. The result matters because it shows that smart memory architecture can compensate for billions of parameters, pointing toward more efficient deployable agents.

██████████ 0.8 long-context Preprint

ReadSaveConnections

Vision-Language-Action (VLA) models trained for robot manipulation fail when deployment conditions shift even slightly from training — different object positions, unstable contacts, or goal changes. This paper wraps a frozen VLA inside an agentic framework that combines it with fixed analytic movement primitives and a memory of what worked before, without any retraining. The approach outperforms end-to-end VLAs under multiple types of deployment perturbation, suggesting that composability rather than monolithic fine-tuning is a practical path to robust robot manipulation.

██████████ 0.8 embodied-ai Preprint

ReadSaveConnections

Most AI agent benchmarks use sandboxed toy environments and single-turn tasks, which hide how agents actually fail when using real software, live browsers, and multi-turn user interactions. UniClawBench runs 400 manually designed tasks inside live Docker containers, grading agents against hidden rubrics via a separate supervisor agent so the executor cannot game the evaluation. The benchmark reveals that both the base model capability and the agent framework design matter — a finding that is often obscured when only one variable is tested at a time.

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

ReadSaveConnections

Hallucination in domain-specific applications like chemistry is dangerous because incorrect facts can look plausible — and standard fine-tuning alone does not reliably suppress them. This paper uses a multi-agent framework where agents interact according to game-theoretic rules (minimizing uncertainty at the local level, coordinating globally via Bayesian probability) to generate 362,000+ high-quality training examples and train a 7B model that reduces hallucinations by 79.46% relative to the base. The headline result is eye-catching but relies on a proprietary evaluation benchmark, so independent verification remains necessary before trusting the number.

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

ReadSaveConnections

LLMs used as data-science agents are frequently evaluated either on abstract causal logic puzzles (with no real data) or on data analysis tasks (with no causal structure) — neither reflects what a real analyst actually does. CausalDS bridges the gap by synthetically generating structural causal models grounded in real-world data distributions, producing tasks that span all three levels of causal reasoning: association, intervention, and counterfactual. This matters because it creates a path to catching 'causal parrots' — models that produce correct-sounding causal language without genuine causal reasoning.

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

ReadSaveConnections

When aligning vision and language models, using low-rank adaptation (updating only a small fraction of parameters) consistently outperforms updating all parameters on most benchmarks, while also being cheaper to train. The paper explains the mechanism: full-parameter alignment destroys a property called linear separability of visual features (dubbed the 'LS-curse'), while low-rank alignment preserves it, which in turn shifts model behavior away from hallucination toward more conservative outputs. This gives practitioners a concrete, mechanistic reason to prefer low-rank alignment rather than just an empirical observation that it works.

██████████ 0.8 efficiency-scaling Preprint

ReadSaveConnections

World models that learn discrete symbolic representations — useful for interpretable reasoning — fail in a specific and consistent way when language supervision is added: the symbols either collapse (only 2.2 of 64 symbols used) or lose their semantic meaning. The paper systematically tests multiple anti-collapse techniques and shows none simultaneously achieve symbol diversity and language grounding, concluding this is a structural flaw in how gradients flow through the bottleneck, not a tuning problem. The proposed fix (gradient detachment at the bottleneck) is simple and resolves both failure modes, which is a practically useful result for anyone building neuro-symbolic or language-grounded agents.

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

ReadSaveConnections

LLMs perform well on isolated psychiatric questions but trail licensed clinicians by 37 percentage points when evaluated across a complete clinical encounter — admission interview, mental status exam, diagnosis, and treatment planning — using 1,193 de-identified real patient cases. Mental status assessment (observing and interpreting patient behavior in real time) is the sharpest bottleneck, suggesting the gap is not about factual recall but about dynamic, process-level clinical reasoning. The benchmark is built with genuine inter-rater reliability controls (Cohen's κ > 0.8) and clinician comparisons, making it a more credible measure of AI clinical readiness than most existing evaluations.

██████████ 0.8 alignment-safety Preprint

ReadSaveConnections

Complex web research tasks require both following chains of evidence deep into a topic and surveying many sources in parallel — a combination that single-agent and simple multi-agent systems handle poorly. WebSwarm addresses this with a recursive delegation scheme where search agents dynamically spawn sub-agents with local objectives, enabling simultaneous depth and breadth. It outperforms single-agent and multi-agent baselines on four benchmarks, though the truncated paper text prevents full verification of the statistical setup.

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

ReadSaveConnections

Vision-language models can name dishes with near-human accuracy but fail catastrophically at estimating portion mass and nutritional content — and, more dangerously, hallucinate safe dietary advice for high-risk patients like diabetics even when ingredient identification is correct. The benchmark evaluates six models across three task tiers from basic perception to safety-critical advice, revealing that visual appearance is a poor proxy for nutritional composition, which is a fundamental challenge for any AI dietary assistant. Note that the paper references model versions (GPT-5.1, Gemini-3-Flash) that do not correspond to publicly known releases, which limits independent replication.

██████████ 0.7 hallucination-grounding Preprint

ReadSaveConnections