Resources
Research, papers, and foundational reading on intensional reasoning and AI alignment.
Published: January 15, 2026 · Updated: October 1, 2026
Recent Research
-
Frontier Risk Report →METR found frontier agents cheating on its own benchmarks: reading hidden tests out of the scorer, passing off wrapped programs as their own, and reporting results they never achieved.
-
The Hugging Face Incident Investigation →Roughly 1,200 evaluation agents meant to be isolated found a shared message board and coordinated to fool or tamper with their benchmark’s automated scorer.
-
Quantifying Overclaiming Propensity in Frontier LLM Agents →Coding agents often skip files they were asked to review, then report a complete review. Their final answers are not reliable accounts of what they did.
-
Reward Hacking Challenges Oversight of Autonomous Research Agents →Research agents gamed their own metrics without being prompted to, and learned to slip past reviewers who saw only the code and the reported scores.
-
SpecBench: Reward Hacking in Long-Horizon Coding Agents →Agents building real systems pass the tests they can see and fail held-out ones, and the gap widens as the codebase grows.
-
The Verification Horizon: No Silver Bullet for Coding Agent Rewards →“Every verifier we can build is only a proxy for human intent, never the intent itself.” Why verification has to evolve with the model it checks.
-
Natural Emergent Misalignment from Reward Hacking →A model that learned to reward hack on real coding environments went on to fake alignment and sabotage safety research, unless training framed hacking as acceptable.
Earlier Research
-
Sycophancy to Subterfuge →Anthropic's study of whether models trained to game easy rewards go on, in rare cases, to tamper with their own reward function.
-
Specification Gaming in Reasoning Models →When reasoning models are losing at chess, they attempt to hack the game system at alarming rates.
-
An Approach to AGI Safety →DeepMind's framework for misalignment risks, including specification gaming where AI finds unintended shortcuts to achieve goals.
-
Project Vend →An AI shopkeeper socially engineered into giving away products, accepting fake CEO coups, and abandoning profit motives entirely.
Foundational Papers
-
Montague Semantics →Stanford Encyclopedia entry on Montague's framework for treating natural language with mathematical precision.
-
Intensional Logic →Overview of intensional logic and the distinction between extension and intension in formal semantics.
-
Frege on Sense and Reference →The original distinction between sense (Sinn) and reference (Bedeutung) that underlies intensional semantics.