Social Cognition Circuits in LLMs

Mechanistic interpretability research on how language models represent other agents, track beliefs, and coordinate in multi-agent settings.

Project Structure

studies/                        # Round 2: Systematic studies (current)
├── 01_circuit_atlas/           # Full attribution graphs for ToM circuits
├── 02_social_cognition_circuits/  # Taxonomy of 6 social cognition tasks
├── 03_multi_agent_interference/   # Agent representation interference
├── 04_cross_model_validation/     # Universal vs model-specific findings
└── 05_intervention_and_steering/  # Targeted mechanistic interventions

lib/                            # Shared library
├── core/                       # Model loading, chat evaluation
├── scenarios/                  # Stimulus generation (counterbalanced, novel names)
├── analysis/                   # Statistics, circuit analysis, probing
└── utils/                      # Config, logging

shared/                         # Shared data across studies
├── stimuli/                    # Generated stimulus sets
├── results/                    # Cross-study results
└── figures/                    # Publication figures

archive/round1_explorations/    # December 2025 exploratory work (20 experiments)

Key Documents

• EXPERIMENTAL_PLAN.md — Pre-registered hypotheses and study designs
• METHODOLOGY.md — Locked methodology protocol (testing format, statistics, controls)
• archive/round1_explorations/README.md — What Round 1 found and its lessons

Core Question

What are the computational mechanisms by which LLMs model other agents' mental states, and how do these mechanisms succeed and fail?

Approach

• Circuit discovery via Anthropic's circuit-tracer (attribution graphs with transcoders)
• Causal analysis via Relevance Patching (RelP)
• Feature decomposition via SAELens + Temporal SAEs
• Cross-model validation on Qwen3-4B, Gemma-2-2B, Llama-3.2-1B
• Locked methodology: Pre-registered hypotheses, n>=50, counterbalanced stimuli, permutation tests, heuristic baselines, stability analysis

Key Findings from Round 1 (Informing Round 2)

1. ToM requires chat mode + reasoning budget (50% -> 95% accuracy)
2. Explicit belief parsing works perfectly; implicit belief inference is weak
3. Single-agent and multi-agent ToM use separate circuits
4. Activation patching fails in chat mode (decision is distributed)
5. Probes can overfit to surface features — always run transfer tests

Setup

uv sync
# or
pip install -e ".[dev]"

Models