Research Briefing — July 1, 2026

FLARE-AI audits 12 existing AI flaw-reporting systems, identifies five recurring design challenges (discoverability, scope, information collection, coordination, guidance), and proposes standardized triage-ready reporting infrastructure. Addresses the fragmented ecosystem for reporting deployed AI system failures.

This paper builds a controlled student-teacher protocol to disentangle whether multi-turn natural-language feedback actually improves LLM agents versus gains from resampling, format fixes, or extra test-time compute. It evaluates thirteen open-weight models as both students and teachers across math, coding, and reasoning benchmarks, isolating the true causal contribution of feedback.

BayesBench introduces simulation environments to test whether LLM belief updates across multi-turn conversations match a rational Bayesian reasoner as evidence accumulates, rather than only scoring final answers. It probes Bayesian estimation and progressively harder inference tasks.

This work applies the autoresearch paradigm to discover convex relaxations yielding certified lower bounds, using a coding agent to propose tightening constraints and a theory agent to verify and search for counterexamples, with bounds certified via rigorous interval arithmetic. It complements prior LLM-agent work finding extremal upper bounds.

This paper audits AI-generated video detection benchmarks, showing a trivial clip-length classifier reaches near-perfect AUC under unaudited protocols and that most published evaluations omit standard controls. It introduces a six-control audited protocol and the VidAudit toolkit to expose and correct confounds.

This paper argues that for expressive generalist robot policies, adapting via reinforcement learning over language prompts is more effective than optimizing directly over actions, since language modulation can elicit skills already latent in the policy. From Sergey Levine's group, it offers a promising route to adapt VLA models to long-horizon, out-of-distribution tasks.

RoPoLL analyzes the LLM jury under the Huber contamination model, proving that consensus averaging suffers unbounded bias when a single judge fails in typical ways like sycophancy or safety refusal. It replaces mean aggregation with a tuning-free robust geometric-median estimator.

HealthAgentBench is a Microsoft Research benchmark of 54 agentic healthcare tasks across 7 categories, each with its own environment replicating end-to-end clinical workflows requiring multi-step exploration of raw health data. It reports task success rate as a holistic measure of frontier agent capability in healthcare.

This paper shows that standard linear probing plus steering can overstate what VLMs actually ground in images for spatial reasoning, and introduces a blank-image ablation arbiter that distinguishes grounded, prior-based, and other regimes. It reveals probing conflates vision-dependent and vision-independent knowledge.

This position and modeling paper argues agents should help non-expert users construct preferences, not merely elicit them via clarifying questions, drawing on the Search-Experience-Credence framework from information economics. It introduces CoPref, a model of how users form preferences based on agent dialog actions.

Sweeps Qwen3 models, optimizers, datasets, and batch sizes to characterize emergent misalignment, finding optimizer choice produces a 7x spread in misalignment while model scale (1B-235B) has negligible effect. Provides a systematic account of training-choice sensitivity for this safety phenomenon.

This paper shows that Scene Coordinate Regression models, assumed privacy-preserving because scenes are only implicitly stored in weights, can leak 3D training geometry via a query-based reconstruction attack. It demonstrates the privacy assumption is false under varying levels of model access.