288

Papers

10/10

Roadblocks Active

0

Connections

Early reasoning mistakes in AI models compound into cascading failures — once a wrong intermediate conclusion is drawn, everything downstream is corrupted. This paper introduces MRPO, a reinforcement learning method that applies exponentially larger penalties to errors in early reasoning steps, reducing early-stage failures from 64% to 13% in medical visual question answering. The mechanism is general: it shows that where credit is assigned during training determines whether errors propagate or get caught, which matters for any AI system that must reason in sequential steps.

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

ReadSaveConnections

When AI agents work through long tasks, they face a memory crisis: either truncate history and lose crucial context, or keep everything and overflow their processing window. ECHO solves this by compressing completed steps into indexed memory records and learning which past evidence actually led to correct answers, enabling explicit credit attribution back to specific retrieved sources. On a hard web research benchmark, ECHO reaches 43.4% accuracy versus 28.9% for standard multi-turn RL, demonstrating that how an agent manages its own history is as important as how it reasons.

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

ReadSaveConnections

LLM security problems are not just about the model weights — vulnerabilities arise at every stage from data collection through deployment, and the attack surface expands dramatically when LLMs are connected to tools, memory, and retrieval systems. The paper maps how untrusted data can become executable instructions once an agent retrieves and acts on it, a threat that bypasses all model-level defenses. This is a timely systematization as agentic deployments move from research to production, and it identifies trust-boundary failures across prompts, outputs, tools, and user authority as the central unsolved problem.

██████████ 0.9 alignment-safety Preprint

ReadSaveConnections

AI research agents that simply ask themselves 'what went wrong?' after a failed experiment tend to either make superficial local fixes or abandon their entire approach — both of which are poor strategies. SAGE instead generates multiple competing causal hypotheses for the failure, scores them against observed evidence, and routes to a structured recovery, reducing hallucinated numbers in scientific reports by constraining outputs to actually measured values. This matters because autonomous research agents are becoming practical tools, and their failure-recovery behavior determines whether they produce trustworthy science or plausible-sounding fiction.

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

ReadSaveConnections

A key finding here is what the authors call an 'evaluation illusion': LLMs can produce fluent, clinically convincing explanations even when their underlying diagnosis is wrong, fooling both automated metrics and casual human review. Using progressive information masking — showing models less and less context — they expose a verbosity bias where GPT-4o-mini's diagnostic accuracy collapses from 95% to 32.5% under information scarcity. This technique of stress-testing models by deliberately degrading their inputs reveals brittle knowledge retrieval hidden behind persuasive text generation, a pattern likely present across many AI application domains beyond medicine.

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

ReadSaveConnections

ProtoPilot converts natural-language biology protocols into robot-executable code through a multi-agent pipeline that checks biological intent, parameter validity, and robotic SDK compatibility at each stage before actuation. On a benchmark of 294 tasks derived from 98 real laboratory protocols, it achieves a 90.2% expert-preference rate and 88.24% successful execution on physical Opentrons robots — compared to 32.35% for the baseline. The real-world wet-lab validation covering plasmid construction, mutagenesis, and DNA assembly is notable because most agent systems are tested only in simulation.

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

ReadSaveConnections

Teaching a vision-language model to search the web for images and text together requires solving two problems simultaneously: training efficiently when some rollouts take far longer than others, and verifying that retrieved visual evidence actually supports the answer. SimpleSearch-VL addresses both with Factorized Adaptive Rollout (which groups training samples to reduce wasted computation from slow outliers) and chain-of-thought evidence verification before committing to an answer. The result is a 15-16 percentage point improvement over strong baselines across six multimodal search benchmarks, trained in roughly one day on eight GPUs — suggesting the recipe is accessible beyond well-resourced labs.

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

ReadSaveConnections

Long-horizon AI agents don't gradually degrade as tasks get harder — they collapse sharply, like water turning to ice, once state complexity or task length crosses a narrow threshold. Using a deterministic puzzle environment with exact ground-truth world states at every step, the paper shows that an agent's internal representation of the world fails before its actions become invalid, meaning the agent is acting from a corrupted mental model rather than simply choosing bad moves. This phase-transition framing is practically useful: it implies that small increases in task complexity near the critical boundary can cause catastrophic — not incremental — performance drops.

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

ReadSaveConnections

This paper tests whether AI can manipulate what people or characters believe not through conversation but through physical action — moving objects, directing characters into rooms — requiring genuine planning about others' mental states rather than just fluent language. GPT-5 achieved roughly 80% success, outperforming human participants, but was less robust than humans across different task contexts, suggesting it exploits surface patterns rather than stable mental-state reasoning. The finding that all systems — human and AI — perform better at inducing true beliefs than false ones points to an asymmetry in how belief-state reasoning is learned that has direct implications for AI safety and deception risk.

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

ReadSaveConnections

When multiple AI agents collaborate on household tasks in simulation, communication between them is consistently beneficial — but the optimal collaboration structure (centralized vs. decentralized, how many agents) depends heavily on team size and individual model capability, meaning there is no universal answer. The benchmark tests 1–5 agent teams across 192 task instances with four communication protocols, finding that coordination overhead can erode the gains from collaboration if not managed. This is practically relevant as multi-agent systems move toward deployment: the paper provides a principled framework for deciding when to add agents versus when to rely on a single capable one.

██████████ 0.8 embodied-ai Preprint

ReadSaveConnections