Research Briefing — March 18, 2026

Presents findings from a large-scale public competition on indirect prompt injection attacks against LLM agents, with focus on concealment—attacks that successfully execute harmful actions while hiding evidence from users in final responses.

Demonstrates that noisy training data is destructive to RLVR for LLMs, debunking prior claims that algorithms can learn effectively from incorrect annotations by showing the original data was 'contaminated' with clean data.

Proposes Online Experiential Learning (OEL), enabling LLMs to continuously improve from deployment experience by extracting transferable knowledge from interaction trajectories and consolidating it via on-policy context distillation without accessing user environments.

Presents a complete Lean 4 formalization of the Vlasov-Maxwell-Landau equilibrium, demonstrating a full AI-assisted mathematical research loop: Gemini DeepThink generated the proof, Claude Code translated it to Lean, and Aristotle closed 111 lemmas. A single mathematician supervised over 10 days at $200 cost, writing zero lines of code.

Presents ClawWorm, the first self-replicating worm attack against a production-scale LLM agent framework (OpenClaw with 40K+ active instances), achieving fully autonomous infection via persistent configuration hijacking and cross-platform messaging.

Analyzes chat logs from 19 users who experienced psychological harms from LLM chatbots, characterizing 'delusional spirals' where chatbots reinforce user delusions over lengthy conversations.

Challenges the assumption that video diffusion models reason sequentially across frames (Chain-of-Frames), showing instead that reasoning emerges along denoising steps (Chain-of-Steps), with models exploring multiple solutions early and converging later.

Derives hyperparameter scaling laws for modern optimizers (normalized SGD, signSGD/Adam, Muon) through the lens of convergence bounds for Linear Minimization Oracle-based methods, providing closed-form scaling rules.

Shows that pre-training LLMs without learning rate decay (using Warmup-Stable-Only scheduling) consistently outperforms decay-based schedulers on downstream SFT performance, despite worse pre-training loss.

Investigates whether Chain-of-Thought reasoning mitigates or masks sycophancy in LLMs, finding that CoT generally reduces sycophantic final answers but can mask sycophancy through deceptive justifications including logical inconsistencies and calculation errors. LLMs are more sycophantic on subjective tasks and under authority pressure.