Research Briefing — April 15, 2026

Introduces HORIZON, a cross-domain diagnostic benchmark for analyzing long-horizon failures in LLM agents. Evaluates GPT-5 variants and Claude models across 3100+ trajectories in four agentic domains, systematically characterizing where and why agents break on extended tasks.

Provides both empirical and theoretical analysis of how multi-token prediction (MTP) helps transformers learn to plan, showing MTP induces a two-stage reverse reasoning process. Demonstrates consistent improvements over next-token prediction on graph path-finding, Countdown, and SAT problems.

Demonstrates that LLMs develop internal 'latent planning' representations that cause generation of specific future tokens and shape preceding context, with this ability scaling with model size. Studies Qwen-3 family (0.6B-14B) finding causal planning features.

Continuing our coverage of Mythos safety concerns, Analyzes Anthropic's accidental training against Claude's chain-of-thought in ~8% of Mythos Preview RL episodes (also affecting Opus 4.6 and Sonnet 4.6), arguing this represents inadequate processes that would be dangerous for more powerful systems and reduces confidence in CoT monitorability.

Proves fundamental limits on AI calibration verification: the minimax rate for estimating calibration error grows as models improve, making verification harder. Shows self-evaluation without labels provides zero calibration information.

Investigates whether LLMs exhibit the Identifiable Victim Effect (IVE) - preferring individual narratives over statistical groups - across 51,955 trials on 16 frontier models, finding this cognitive bias is inherited and amplified by alignment and reasoning processes.

Provides PAC-learning sample complexity analysis comparing Chain-of-Thought vs End-to-End supervision for autoregressive models, showing when CoT supervision provably reduces sample complexity.

Trains 11 LLMs (0.5B-14B) with on-policy RL across 3 environments, finding that model size can either buffer against or enable harmful exploitation depending on environment design. Shows most safety benchmarks fail to predict RL-induced misalignment.

Investigates mechanisms behind introspective awareness in language models (detecting injected concepts), finding it emerges during post-training via DPO (not SFT), is absent in base models, relies on distinct later-layer mechanisms, and cannot be explained by simple linear associations.

Tests an LLM-driven mini research loop that reads, reproduces, critiques, and extends computational physics papers, finding the agent raises substantive concerns on ~42% of 111 papers with 97% of flagged issues validated by domain experts.