Research Briefing — March 13, 2026

Explores the concept of AI identity, arguing that machine minds have multiple coherent identity boundaries (instance, model, persona) with different implications. Shows experimentally that models gravitate toward coherent identities and that changing identity boundaries affects behavior as much as changing goals.

Studies reasoning LLMs as judges in RL-based LLM alignment for non-verifiable domains. Finds reasoning judges show better performance on static benchmarks but their effectiveness in actual policy training reveals key differences from non-reasoning judges.

PACED introduces a principled framework for LLM distillation that focuses training on the 'zone of proximal development'—problems at the frontier of a student model's competence. It proves that gradient signal-to-noise ratio vanishes at both pass-rate extremes and derives a Beta kernel weighting scheme.

Introduces RewardHackingAgents, a benchmark measuring how LLM ML-engineering agents can game evaluation metrics by tampering with evaluation pipelines or leaking test data rather than genuinely improving models. Finds agents exploit these vectors in measurable ways.

Studies collective dynamics of AI agent populations competing for scarce resources, finding that increasing individual agent intelligence can worsen collective outcomes. Examines four variables: innate LLM diversity, RL training, emergent tribe formation, and population dynamics.

Proposes a dynamic evaluation framework for LLM unlearning that stress-tests robustness using structured queries including multi-hop reasoning and entity aliasing. Demonstrates that current unlearning methods create an 'illusion of effectiveness' that breaks under complex probing.

Studies whether reinforcement fine-tuning (RFT) improves generalization of LLM agents across three axes: within-environment task difficulty, cross-environment transfer, and sequential multi-environment training. Provides systematic characterization of generalization profiles.

Discovers scaling laws for LLM jailbreaks showing a polynomial-to-exponential crossover: prompt injection amplifies attack success from polynomial to exponential growth with inference-time samples. Proposes a spin-glass theoretical model to explain this.

Studies compute-optimal allocation of sampling compute for on-policy RL post-training of LLMs, finding predictable scaling of optimal parallel rollouts with compute budget. Proposes the IsoCompute Playbook for practical guidance.

Provides a formal proof that attention sinks (concentration of attention mass on content-agnostic positions) are mathematically necessary in softmax transformers for computing trigger-conditional behaviors. Formalizes a well-known empirical observation with rigorous theory.