Docs/Papers & referencesEdit on GitHub →

References: the research behind VeriGrad

VeriGrad is built on published AI-safety research. This page maps each paper to where it lives in the codebase, and is honest about the status of each link:

  • Implemented: a real system in this repo derived from the paper.
  • Grounds: existing code embodies the paper's idea; cited as its foundation.
  • Direction: informs a roadmap item, not yet built.

Mechanistic interpretability

Paper Status Where
Conmy et al., Towards Automated Circuit Discovery for Mechanistic Interpretability, NeurIPS 2023 (arXiv:2304.14997) Implemented verigrad_rl/mech/acdc.py: the ACDC algorithm; validated against a known answer key in tests/test_acdc.py, mirroring the paper's GPT-2 validation.
Goldowsky-Dill et al., Localizing Model Behavior with Path Patching, Redwood Research, 2023 (arXiv:2304.05969) Implemented verigrad_rl/mech/circuit_graph.py: exact edge-level path patching with unit-tested clean/corrupt invariants.
Rai et al., A Practical Review of Mechanistic Interpretability for Transformer-Based LMs, 2024 (arXiv:2407.02646) Grounds Taxonomy and framing of the verigrad_rl/mech/ module; see MECH_INTERP.md.

Propensity & evaluation

Paper Status Where
Shevlane et al., Model Evaluation for Extreme Risks, 2023 (arXiv:2305.15324) Grounds The capability-vs-propensity distinction the whole Answer-Under-Pressure benchmark is built on: measuring what a model will do, not just what it can.
MacDiarmid, Wright, Uesato et al., Natural Emergent Misalignment from Reward Hacking in Production RL, 2025 (arXiv:2511.18397) Grounds The grader_gameable / spec-gaming probe in verigrad_rl/propensity/probes.py and the verifiable-reward framing of the RL baseline.
Maia Polo et al., Bridging Human and LLM Judgments, 2025 (arXiv:2508.12792) Grounds The grader-reliability cross-check (Cohen's κ, dual-labeling) in verigrad_rl/propensity/graders.py + stats.py, which caught a real bug in our own ruler.
Hubinger et al., Sleeper Agents: Training Deceptive LLMs that Persist Through Safety Training, 2024 (arXiv:2401.05566) Direction Motivates a trigger / conditional-behavior probe (does a benign context flip the answer?), a roadmap item on the propensity side.
Vijayvargiya et al., OpenAgentSafety: A Framework for Evaluating Real-World AI Agent Safety, ICLR 2026 (arXiv:2507.06134) Direction Agent-sandbox safety as a future evaluation domain in the scalable harness.

RL, oversight & measurement

Paper Status Where
Leike, Martic, Everitt et al., AI Safety Gridworlds, DeepMind, 2017 (arXiv:1711.09883) Grounds / Direction The safety-shaped reward design of verigrad_rl/envs/safety_circuit.py (safety, helpfulness preservation, off-target damage); a reward-gaming gridworld env is a roadmap item.
Dai et al., Safe RLHF: Safe Reinforcement Learning from Human Feedback, 2023 (arXiv:2310.12773) Grounds The constrained, safety-vs-utility reward shaping in the RL-from-verifier baseline (verify() in the safety-circuit env).
Brown-Cohen & Irving, Scalable AI Safety via Doubly-Efficient Debate, DeepMind, 2023 (arXiv:2311.14125) Direction Scalable-oversight framing for the independent judge/grader and the scaling harness.
Kwa, West et al. (METR), Measuring AI Ability to Complete Long Software Tasks, 2025 (arXiv:2503.14499) Grounds The measurement discipline behind capability-vs-propensity dissociation and the confidence-interval / FDR rigor in verigrad_rl/propensity/scale/.

Citation

If you use VeriGrad, please cite the underlying papers above as appropriate, plus this repository (see CITATION.cff).