Research Briefing — February 19, 2026

Proposes twelve concrete metrics decomposing AI agent reliability along four dimensions (consistency, robustness, predictability, safety), grounded in safety-critical engineering. Evaluates 14 agents and finds that standard success metrics obscure critical operational flaws.

Proposes a framework treating interactive in-context learning from natural language feedback as a trainable skill rather than emergent property. Introduces a scalable method to transform single-turn verifiable tasks into multi-turn didactic interactions, showing current flagship models are suboptimal at learning from feedback.

Tracks five geometric measures across model scales (405K-85M) during training, finding universal representation collapse to scale-invariant floors followed by top-down layer reorganization preceding capability emergence.

Characterizes 'Retrieval Collapse' — a failure mode where AI-generated content dominates search results, eroding source diversity and enabling adversarial content infiltration. Shows 67% pool contamination leads to 80%+ exposure contamination in RAG/search systems.

Pre-registered RCT (n=153) testing whether LLMs improve novice performance in biology lab tasks modeling a viral reverse genetics workflow. Found no significant difference in workflow completion between LLM and Internet groups.

Introduces DreamZero, a World Action Model built on pretrained video diffusion that jointly models video and action for robotic manipulation. Achieves 2x improvement in generalization to new tasks/environments compared to state-of-the-art VLAs.

Presents EgoScale, a framework that trains a Vision-Language-Action model on 20,854 hours of egocentric human video for dexterous manipulation, discovering a log-linear scaling law between human data scale and performance.

Introduces EnterpriseGym Corecraft, a high-fidelity enterprise simulation environment for training AI agents on customer support tasks. Frontier models like GPT-5.2 and Claude Opus 4.6 solve <30% of tasks. Training with RL on this environment produces capabilities that generalize beyond training distribution.

Identifies 'optimization instability' in autonomous agentic workflows for clinical symptom detection, where iterative self-improvement paradoxically degrades performance. At 3% prevalence, systems achieved 95% accuracy while detecting zero positive cases, exposing dangerous failure modes.

Proposes Doc-to-LoRA (D2L), a hypernetwork that meta-learns to perform approximate context distillation in a single forward pass, generating LoRA adapters to internalize document context without re-consuming it during inference.

Proposes REMUL, a multi-party RL approach where listener models execute truncated reasoning traces to improve chain-of-thought faithfulness in LLMs, addressing the tradeoff between faithfulness and task performance.

Introduces STING, an automated red-teaming framework for measuring illicit assistance in multi-turn LLM agent interactions. Constructs step-by-step illicit plans with benign personas and uses judge agents to track completion.