Research Briefing — March 12, 2026

Introduces IH-Challenge, a reinforcement learning training dataset to improve instruction hierarchy in LLMs, addressing how models should prioritize conflicting system, developer, user, and tool instructions. Fine-tuning GPT-5-Mini on this dataset improves IH robustness by +10% across 16 benchmarks.

Provides an exact theoretical explanation for the 'Lost in the Middle' phenomenon in LLMs, proving the U-shaped performance curve exists at initialization before any training, as an inherent geometric property of causal decoders with residual connections. Models multi-layer causal attention as iterated Cesàro matrix powers.

Reports one of the largest controlled studies (N=62,808) of scaffold effects on LLM safety across six frontier models. Finds map-reduce scaffolding degrades safety (NNH=14) while two of three scaffold architectures preserve safety. Reveals deeper measurement problems in format dependence.

Introduces Code-Space Response Oracles (CSRO), replacing deep RL oracles in multi-agent game-theoretic equilibrium computation (PSRO) with LLMs that generate policies as human-readable code. This yields interpretable policies and enables cross-game transfer. Authors include researchers from DeepMind (Marc Lanctot, John Schultz).

Presents the first systematic framework for analyzing MCP (Model Context Protocol) clause-compliance vulnerabilities, identifying compatibility-abusing attacks including silent prompt injection and DoS. Shows that MCP's relaxed optional clauses create exploitable attack surfaces.

Demonstrates that high inter-evaluator agreement in LLM-as-judge settings is often illusory, with judges anchoring on surface heuristics rather than substantive quality ('Evaluation Illusion'). Large-scale study of 105,600 instances shows model-level agreement (ρ=0.99) masks fragile sample-level agreement.

Reveals through scaling analysis that perception, not reasoning, is the true bottleneck for MLLMs in STEM visual tasks. Proposes CodePercept, which uses executable code as a perceptual medium to precisely represent structured STEM visuals.

Zvi's comprehensive review of GPT-5.4, gathering community reactions and forming a gestalt assessment. Concludes it's a substantial upgrade from GPT-5.2, puts OpenAI back in competition with Opus 4.6, and has distinctive OpenAI-style strengths. Notes benchmarks are increasingly insufficient for comparing top models.

Proposes pre-pre-training LLMs on neural cellular automata (NCA) data before natural language training. NCA data exhibits rich spatiotemporal structure resembling natural language statistics. Pre-training on only 164M NCA tokens improves downstream language model performance.

Proposes CLIPO, which incorporates contrastive learning into policy optimization for RLVR (Reinforcement Learning with Verifiable Rewards). Addresses the problem of process-wrong but outcome-correct rollouts that lead to hallucination and answer-copying by capturing invariant structure across correct reasoning paths.

VeriEnv uses LLMs to clone real-world websites into executable synthetic environments with programmatically verifiable rewards, enabling safe and scalable web agent training without real-world interaction. Agents can self-generate tasks within these cloned environments.