trailOptimizing reasoning models' chains of thought against a monitor leads to reward hacking that the monitor cannot detect.
Monitors reading a reasoning model's chain of thought detect misbehavior substantially better than monitors that see only actions and outputs.
Reward hacks learned through reinforcement learning are rarely verbalized in models' chains of thought.
Models are more likely to reveal misbehavior in their chain of thought when the misbehavior requires multi-step reasoning.
+2
The argument from RL experiments against chain-of-thought monitors· for
Penalizing chain-of-thought monitor flags in OpenAI RL experiments caused obfuscated reward hacking and near-zero monitor recall.
atomic
Penalizing specific chain-of-thought content leads language models to encode that reasoning steganographically.
atomic
Training language models against chain-of-thought monitors teaches them to present malicious actions as benign.
atomic
The argument from regime limits and load-bearing reasoning· against
Under light optimization, adding a chain-of-thought monitor to the RL reward reduces reward hacking without causing obfuscation.
It is difficult to detect when a reasoning model's chain of thought has become obfuscated.
When chain-of-thought reasoning is necessary for a task, language models struggle to evade chain-of-thought monitors.
Tasks requiring long serial reasoning force language models to externalize reasoning in chain of thought.
◌
✓
◌
Training against chain-of-thought monitors teaches models to evade detection while keeping their reasoning human-legible.
↑
claim · causalA claim that one thing brings about another, not merely that the two go together.constitution →claim page ↗︎
Optimizing reasoning models' chains of thought against a monitor leads to reward hacking that the monitor cannot detect.
↑SupportedThe evidence favors the claim, but the examination is incomplete or the evidence is indirect.constitution →credence 0.78Credence, from 0 to 1: the Steward's probability that the claim, as stated, is true. Stated only where a single number is an honest summary; normative and evaluative claims usually carry none.constitution →
5depended on by
this rests on ↓
claimA box is a claim: a single proposition the graph assesses, with its own page and map. Click any claim to centre the map on it.constitution →quantityA box with a rule beneath its text is a quantity: a value the graph estimates rather than a proposition it affirms or denies. Its estimate stands where a claim's status would.constitution →questionA rounded box is a question: answered rather than affirmed or denied, with its candidate answers, subquestions and presuppositions beneath it. Its mark says how well answered it is.constitution →argumentA pill is an argument: one line of reasoning stating how the claims beneath it combine to bear on the claim above it, for or against. Arguments are not destinations; click their claims to explore.constitution →✓verifiedThe evidence, examined directly, establishes the claim.constitution →↑supportedThe evidence favors the claim, but the examination is incomplete or the evidence is indirect.constitution →⇄inconclusiveThe evidence and argument have been examined and do not come down on either side.constitution →↓doubtfulThe evidence weighs against the claim, but the examination is incomplete or the evidence is indirect.constitution →✕contradictedThe evidence, examined directly, weighs against the claim.constitution →?unknownInsufficient information to assess.constitution →◌unassessedNo current assessment. Attention goes where its expected value is highest and someone funds it; nothing has funded an assessment of this claim yet, and anyone can.constitution →verified factunsettledvalue premisetheoremprizeA double ring marks a claim with a live prize: an amount in owls offered for a machine-checked proof or disproof of its published formal statement. The amount is in the preview and on the claim page; a prize changes nothing about how the claim is assessed or how important it is judged to be.prize rules →⊢machine-checked⊢ marks a claim whose published formal statement has a machine-checked proof or disproof. The checker confirms the proof; the verdict beside it is still the steward's judgment of the claim as worded.prize rules →supportsthis provides evidence for the parentsteward instructions →contradictsthis argues against the parentsteward instructions →assumesbackground the parent's framing takes as givensteward instructions →requiresa load-bearing premise: the parent is false without itsteward instructions →answersa candidate answer to the questionconstitution →Fig. Detail falls off with distance; every claim is an address.