278

Papers

11/11

Roadblocks Active

2

Connections

This theoretical paper argues that delegating AI alignment research to AI agents is dangerous even without scheming: the tasks involved (evaluating safety, writing interpretability probes, assessing value learning) are inherently fuzzy and lack reliable ground truth, so optimization pressure will push agent-generated errors toward the blind spots of human reviewers. The argument is not that agents will lie, but that the pipeline systematically rewards plausible-sounding mistakes. For anyone building automated alignment research programs, this is a direct challenge to the underlying assumption that human oversight can catch what AI produces.

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

ReadSaveConnections

A multi-agent LLM system autonomously completed a full scientific workflow in computational fluid dynamics — searching literature, forming hypotheses, modifying C++ solver code, running simulations, and verifying results — achieving a 7.89% reduction in wall-friction error against a ground-truth DNS simulation. A vision-language verification gate caught 14 of 16 planted silent failures that standard solver checks missed entirely. The result matters because it is the first demonstrated end-to-end AI scientist pipeline that includes physics-based validation, not just text generation, reducing the hallucination risk inherent in pure LLM science assistants.

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

ReadSaveConnections

This paper benchmarks how well LLM agents detect that earlier memories have been invalidated by later observations — for example, knowing that a door they unlocked an hour ago may now be locked again. Frontier models score only 55.2% on a 400-scenario benchmark, and consistently fail when invalidation is implicit rather than explicitly stated. The practical failure mode is that agents confidently act on outdated state, which is a critical issue for any deployed system that maintains memory across sessions.

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

ReadSaveConnections

Google DeepMind's multi-agent system built on Gemini models achieved 48% on FrontierMath Tier 4 — currently the strongest reported score among AI systems on that benchmark of research-level mathematics — and helped practicing researchers identify overlooked literature and new problem directions in real open-ended work. The system uses asynchronous stateful workspaces rather than single-shot prompting, enabling sustained multi-step mathematical reasoning. The caveat is that the system is proprietary and no code is released, so the result is not independently verifiable.

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

ReadSaveConnections

Activation steering — injecting vectors into a model's internal activations to change its behavior — degrades reasoning and retrieval because it inadvertently shifts the model's attention away from contextually important tokens. SKOP (Steering via Key-Orthogonal Projections) fixes this by projecting out the components of steering vectors that interfere with high-attention 'focus' tokens, reducing performance degradation by 5–7x while retaining over 95% of the behavioral change. This matters for interpretability and safety tooling: it means steering-based interventions can now be applied more precisely without collateral damage to task performance.

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

ReadSaveConnections

This theoretical paper proves a 'parameter coverage ceiling': there exist practically relevant inputs that no fixed-parameter model can handle reliably, because the parameter space cannot encode all necessary knowledge within tolerance bounds. The authors argue that agentic systems — ones that can perceive, retrieve external information, and take actions in a feedback loop — are not merely convenient but mathematically necessary for out-of-distribution generalization. If the proof holds up to scrutiny, it provides a formal justification for why scaling model parameters alone cannot solve generalization, which is a claim many practitioners hold informally but that has lacked rigorous grounding.

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

ReadSaveConnections

MANTRA automatically converts natural-language procedural manuals into formal compliance benchmarks by generating two independent artifacts — a symbolic world model and trace-level compliance checks — then using an SMT solver to verify their consistency and repair conflicts. Applied to 285 tasks across 6 domains from manuals up to 50 pages long, it produces benchmarks that are formally validated for logical coherence, unlike most existing agent evaluation suites. This addresses a real gap: most agent benchmarks are written by hand and contain subtle inconsistencies that allow agents to score well by exploiting flaws rather than actually following procedures.

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

ReadSaveConnections

PrefixGuard trains lightweight monitors on typed, abstracted agent trace prefixes to predict — in real time, before completion — whether a running LLM agent task is heading toward failure. Across four benchmarks (WebArena, τ²-Bench, SkillsBench, TerminalBench), learned monitors substantially outperform LLM-judge baselines, with the StepView typed-step adapters contributing +0.137 AUPRC on average. The practical value is early warning: rather than waiting for an agent to fail at step 40 of a 50-step task, operators can intervene at step 15 when recovery is still cheap.

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

ReadSaveConnections

MedHorizon builds a benchmark from 340 full-length clinical videos (759 hours, 7 organ types, 1,253 multiple-choice questions) and tests leading multimodal models on it. The key finding is that feeding more video frames to models does not reliably improve performance — the bottleneck is weak procedural reasoning and attention drift, not simply lack of visual information. This challenges the common assumption that longer context windows directly translate to better video understanding, especially in high-stakes clinical settings.

██████████ 0.8 long-context Preprint

ReadSaveConnections

This paper tests Language Model Agents performing cybersecurity red-teaming (lateral movement in a Windows Active Directory environment) across three modes: fully autonomous, self-scaffolded, and expert-guided. Expert-defined action plans yielded the highest task completion, but failure rates remain high across all modes, with brittle command invocation — the agent calling tools incorrectly rather than reasoning incorrectly — as the primary culprit. The result is a concrete data point on the current gap between AI agent capability and the demands of real operational security tasks.

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

ReadSaveConnections