🧬 Science & Research Frontier

METHOD: Formalizes the Multi-Agent Exploration problem as a partially observable stochastic game (POSG), demonstrating that modern LLM agents exhibit myopic, polarized interaction patterns that lead to suboptimal coordination and higher regret. Introduces MACE (Multi-Agent Contextual Exploration), a lightweight framework that promotes structured peer selection.

KEY RESULT: LLM agents default to shallow, non-exploratory behavior — they talk to the same peers repeatedly and fail to discover diverse capabilities. MACE substantially improves exploration and downstream task performance across both contextual and parametric diversity settings. Theoretically proves that exploration value increases with agent diversity — meaning heterogeneous teams need more explicit exploration, not less.

METHOD: Controlled experiments prepending irrelevant context (including semantically meaningless pseudo-words — random character combinations) to benchmark questions. Measures both aggregate accuracy and per-example prediction stability across models, datasets, context types, context lengths, and test-time compute budgets.

KEY RESULT: Aggregate accuracy remains deceptively stable, but individual predictions flip — some correct answers become wrong, some wrong answers become correct — masking significant instability. Even pseudo-words trigger this effect. The phenomenon is consistent across model families and datasets but the affected examples are largely model-specific. Test-time compute and instruction-tuning modulate but don't eliminate the effect.

METHOD: Casts sycophancy/overconfidence as failures of internal incentive-compatibility (IC). Uses interchange interventions to causally identify low-rank, near-orthogonal report coordinates (answer, confidence, caveat). Introduces CRC clamp — a training-free mechanism that references the model's own output under a counterfactually incentive-neutralized context.

KEY RESULT: The two-pass CRC clamp achieves resist and update jointly at 1.00 (Wilson 95% CI [0.99, 1.00]) — a causal certificate that the model can simultaneously resist forbidden influences (pressure, prestige) while remaining responsive to genuine evidence. The single-pass deployable version is lossy (0.73/0.97). Mechanism reproduces across three model families and transfers to SycophancyEval. Global decoding shows a single-parameter resist/update tradeoff — resist-only training kills evidence-responsiveness.

A cognitive-neuroscience-inspired framework that parcels LLM internal features into functional units mapping to interpretable cognitive functions. Critical finding: major LLM failures — hallucination, bias, refusal failure, sycophancy — correspond to distinct disruptions in specific representational systems, yielding internal biomarkers for mechanism-guided detection and targeted intervention. The framework also improves prediction of human cortical responses during language comprehension, with strongest alignment in higher-order association cortex. This is the most ambitious bridge between AI interpretability and cognitive neuroscience to date — and it provides actionable internal signatures, not just descriptive maps.