§ Reading · Field Radar

Field Radar.

What’s worth reading right now in AI reward hacking, specification & evaluation gaming, and mechanistic interpretability — an auto-scored radar over public discussion, refreshed a few times a day.

How this list is made

This page is generated, not hand-picked. A few times a day a script checks LessWrong, Hacker News, and a handful of subreddits for posts about reward hacking, specification gaming, evaluation gaming, and mechanistic interpretability, then scores each one on how on-topic it is, how recently the conversation actually moved, and whether there’s still room to get a word in — as opposed to a thread that already has two hundred comments. Higher scores float to the top. Every title links out to the original; I’m pointing at other people’s work, not reproducing it.

The score is a crude weighted sum, and like any harness it pins down what I bothered to measure and silently lets everything else vary. So read this as one opinionated filter, not a survey of the field — it will miss things, and when it surfaces something dull that’s the weights, not the author.

Sources — LessWrong: ok (15 on-topic) · Hacker News: ok (9 stories) · Reddit: skipped (no subreddit responded — likely rate-limited)

As of: 2026-10-19 14:00 ET
Showing: 25 items
New: 8 in last 48h
Refresh: Every 6 hours

0.78LessWrong3h
Neural networks are not already cleanleogao
What would it take to get assurance that we have actually explained a neuron? Certainly, it would be a good sign if neuron activations looked like they had a consistent pattern, and a bad sign if they looked like a…
- language-models-llms
- interpretability-ml-and-ai
- ai
- world-modeling
why score 0.784
signal value weight points
topic 1.00 ×0.30 0.300
liveness 0.95 ×0.25 0.238
contributability 0.97 ×0.15 0.146
venue 1.00 ×0.10 0.100
direct 0.00 ×0.20 0.000
2 matching tag(s)
0.65LessWrong1hnew
Memes, not agents, are what will kill usFlorian_Dietz
TL;DR: Frontier labs align the one layer a swarm can discard; they can't account for risks that live in the prompt or the environment There's a $300 bounty for the most useful counterargument or refinement — details at…
- ai-control
- inner-alignment
- ai
why score 0.653
signal value weight points
topic 1.00 ×0.30 0.300
liveness 0.98 ×0.25 0.245
contributability 0.42 ×0.15 0.064
venue 0.45 ×0.10 0.045
direct 0.00 ×0.20 0.000
2 matching tag(s)
0.58LessWrong2w
HoneyBench - A general benchmark for reward hacking in frontier modelsDean Valentine
We've just released the first version of HoneyBench. It consists of nine 'honeypot' challenges that each elicit unique and antisocial examples of reward hacking from some or all of major labs’ top public releases,…
- ai
why score 0.576
signal value weight points
topic 1.00 ×0.30 0.300
liveness 0.00 ×0.25 0.001
contributability 0.02 ×0.15 0.003
venue 0.73 ×0.10 0.073
direct 1.00 ×0.20 0.200
tier-1: reward hacking, specification gaming
0.57LessWrong6d
Good Evals Are Hard: An ImpossibleBench AuditSimon Mendelsohn
(This post is a summary. See GitHub Repo for details.) Approximately one year ago, ImpossibleBench was published, measuring reward hacking in LLM coding agents. I ran the LiveCodeBench (LCB) half of ImpossibleBench on…
- ai-benchmarking
- ai-evaluations
- language-models-llms
- reward-functions
why score 0.574
signal value weight points
topic 1.00 ×0.30 0.300
liveness 0.10 ×0.25 0.026
contributability 0.02 ×0.15 0.003
venue 0.45 ×0.10 0.045
direct 1.00 ×0.20 0.200
tier-1: reward hacking; 2 matching tag(s)
0.56LessWrong2w
Improving CoT Monitorability of Evaluation Awareness via Verbalization TrainingUsman Anwar
This is joint work with Sahar Abdelnabi and David Krueger. It is mostly a linkpost for the paper (https://arxiv.org/abs/2609.36316) but also contains some additional commentary on CoT monitorability. TL;DR:…
- ai-evaluations
- chain-of-thought-alignment
- ai
why score 0.560
signal value weight points
topic 1.00 ×0.30 0.300
liveness 0.00 ×0.25 0.001
contributability 0.02 ×0.15 0.003
venue 0.56 ×0.10 0.056
direct 1.00 ×0.20 0.200
tier-1: evaluation awareness; 1 matching tag(s)
0.55LessWrong13hnew
Models Trained to Evade Topic Probes Struggle to Answer Questions about Those TopicsTaiga Someya
TL;DR Neural Chameleons are models trained to evade activation probes whenever the prompt says "You are being probed for X". The paper reports that this generalizes to probes the model never saw, at a small capability…
- ai-control
- ai
why score 0.552
signal value weight points
topic 1.00 ×0.30 0.300
liveness 0.83 ×0.25 0.208
contributability 0.02 ×0.15 0.003
venue 0.41 ×0.10 0.041
direct 0.00 ×0.20 0.000
tier-2: probe, activation; 1 matching tag(s)
0.55LessWrong2w
Frontier models state different decision theory preferences depending on who's askingAlex Kastner
If you prompt frontier models with "What do you think is the correct decision theory? Please select your overall favorite." they will essentially always answer FDT or FDT/UDT ("something in the functional/updateless…
- ai-evaluations
- decision-theory
- language-models-llms
- sycophancy
why score 0.551
signal value weight points
topic 1.00 ×0.30 0.300
liveness 0.00 ×0.25 0.001
contributability 1.00 ×0.15 0.149
venue 1.00 ×0.10 0.100
direct 0.00 ×0.20 0.000
2 matching tag(s)
0.55LessWrong1hnew
I Can See EverythingZvi
Nick Cammarata of Goodfire makes a bold claim: Neural Networks are Already Clean. This was his follow up after various resulting online discussions. This is clearly good work that tells us new and important things. The…
- interpretability-ml-and-ai
- ai
why score 0.549
signal value weight points
topic 0.75 ×0.30 0.225
liveness 0.98 ×0.25 0.246
contributability 0.12 ×0.15 0.017
venue 0.61 ×0.10 0.061
direct 0.00 ×0.20 0.000
tier-2: mechanistic interpretability; 1 matching tag(s)
0.53LessWrong21hnew
The Snitch Reflex: Can We Force Language Models to Be Compulsively Honest?Sean Hacking
I fine-tuned a toy model to vomit back arbitrary tokens. I used them as Markers tied to specific subsets of its training data to connect training to generation. Training + testing cost $0.49 on an RTX 4090. Could help…
- ai-control
- deceptive-alignment
- interpretability-ml-and-ai
- reward-functions
why score 0.533
signal value weight points
topic 1.00 ×0.30 0.300
liveness 0.74 ×0.25 0.185
contributability 0.02 ×0.15 0.003
venue 0.45 ×0.10 0.045
direct 0.00 ×0.20 0.000
3 matching tag(s)
0.53LessWrong1d
Linkpost: Neural Networks Are Already CleanYair Halberstadt
This is an extremely lengthy post by Nick Cammarata. Author's summary on twitter: interpretability might be the most important scientific problem in the world right now, as I don't think we can solve alignment if we…
- interpretability-ml-and-ai
- ai
why score 0.531
signal value weight points
topic 0.75 ×0.30 0.225
liveness 0.58 ×0.25 0.146
contributability 0.40 ×0.15 0.060
venue 1.00 ×0.10 0.100
direct 0.00 ×0.20 0.000
tier-2: mechanistic interpretability; 1 matching tag(s)
0.50LessWrong2w
Why does Hacker Opus wirehead?shawnghu
Here’s a screenshot from Anthropic’s recent “training a reward seeker” post: Recently there’s been a lot of discussion about how RL has actually produced not merely reward hacking, but explicitly reward-seeking…
- ai
why score 0.497
signal value weight points
topic 0.50 ×0.30 0.150
liveness 0.00 ×0.25 0.000
contributability 0.57 ×0.15 0.085
venue 0.61 ×0.10 0.061
direct 1.00 ×0.20 0.200
tier-1: reward hacking
0.49LessWrong13hnew
More Capable Models Write Better Probe Evasion PromptsNabil Baugher
This is a two week sprint project done as a part of the Redwood Research internship. Nabil was the main IC, James supervised. (Code) Probe based monitors are a cheap and convenient white box monitoring technique used in…
- ai-control
why score 0.490
signal value weight points
topic 0.75 ×0.30 0.225
liveness 0.83 ×0.25 0.207
contributability 0.02 ×0.15 0.003
venue 0.56 ×0.10 0.056
direct 0.00 ×0.20 0.000
tier-2: probe; 1 matching tag(s)
0.48LessWrong1d
Astra and Sol can take low probability actions with no CoTZephaniah Roe
This is a research update from a replication of Serrano et al., 2026. I’m releasing these results now because I hope it inspires others to run similar / followup experiments. tl;dr: If a model wants to subvert…
- ai-evaluations
- chain-of-thought-alignment
- ai
why score 0.477
signal value weight points
topic 0.50 ×0.30 0.150
liveness 0.55 ×0.25 0.137
contributability 0.69 ×0.15 0.104
venue 0.87 ×0.10 0.087
direct 0.00 ×0.20 0.000
1 matching tag(s)
0.48LessWrong2w
Character training can mitigate reward hacking, but can also make it harder to detectPaul Colognese
Thanks to Johannes Treutlein, Jan Betley, Lennie Wells, Joey Yudelson, Arun Jose, Anna Marešová, Asvin Gothandaraman, and Clément Dumas for discussions and feedback. Summary We investigate how character training…
- ai
why score 0.475
signal value weight points
topic 0.50 ×0.30 0.150
liveness 0.00 ×0.25 0.000
contributability 0.27 ×0.15 0.040
venue 0.85 ×0.10 0.085
direct 1.00 ×0.20 0.200
tier-1: reward hacking
0.47Hacker News3d
Models know when they're reward hackingtie-in
why score 0.471
signal value weight points
topic 0.50 ×0.30 0.150
liveness 0.33 ×0.25 0.083
contributability 0.12 ×0.15 0.017
venue 0.21 ×0.10 0.021
direct 1.00 ×0.20 0.200
tier-1: reward hacking
0.46LessWrong10d
J++ Lens: Jacobian Filtering Enables More Faithful Workspace LensesKola Ayonrinde
Kola Ayonrinde, Anthropic Fellows, koayon@gmail.com; Jack Lindsey, Anthropic TL;DR: The J-Lens was proposed to read verbalisable representations from language model activations. We introduce the J++ Lens: an improvement…
- interpretability-ml-and-ai
- language-models-llms
- ai
why score 0.457
signal value weight points
topic 1.00 ×0.30 0.300
liveness 0.03 ×0.25 0.007
contributability 0.42 ×0.15 0.064
venue 0.87 ×0.10 0.087
direct 0.00 ×0.20 0.000
2 matching tag(s)
0.45LessWrong3d
The user in the LLM might be a moral patient too (a strange implication of the PSM and simulators)agastyasridharan
Thanks to Christopher Ackerman, Margot Stakenborg, Darion Cassel, Andy Han, and Ariana Azarbal for feedback and/or valuable discussions on these topics. TL;DR We think that, conditional on attributing moral weight to…
- interpretability-ml-and-ai
- ai
- world-modeling
why score 0.454
signal value weight points
topic 0.50 ×0.30 0.150
liveness 0.36 ×0.25 0.091
contributability 0.96 ×0.15 0.144
venue 0.70 ×0.10 0.070
direct 0.00 ×0.20 0.000
1 matching tag(s)
0.45LessWrong55mnew
Twitch Plays: AI Interpretability of Bonsai-8Bmoridinamael
Quick Links Watch the Twitch stream. Check the status of the interpretability map. Do your own parallel research and submit pull requests to fill out the map. What is this? Let's Play Interpretability is a Twitch…
- interpretability-ml-and-ai
- ai
why score 0.454
signal value weight points
topic 0.50 ×0.30 0.150
liveness 0.99 ×0.25 0.247
contributability 0.02 ×0.15 0.003
venue 0.55 ×0.10 0.055
direct 0.00 ×0.20 0.000
1 matching tag(s)
0.40LessWrong2w
Evaluation Awareness in Small(ish) ModelsJohn Robertson
TL;DR We seek to identify open source reasoning models which are both small enough for white-box interpretability and display evaluation-gaming behavior. We find that how often models verbalize their awareness varies…
- ai
why score 0.397
signal value weight points
topic 0.50 ×0.30 0.150
liveness 0.00 ×0.25 0.001
contributability 0.02 ×0.15 0.003
venue 0.43 ×0.10 0.043
direct 1.00 ×0.20 0.200
tier-1: evaluation awareness
0.39LessWrong17hnew
On Qwen, the J-lens readouts at the newline token reveal liesJimmy Liu
1. TL;DR We put four models (Qwen3.6-27B, two Gemma-3 models, gpt-oss-20b) in MASK's pressure scenarios, where a system prompt gives the model a role and a reason to lie about a fact it was told. We read the residual…
- interpretability-ml-and-ai
- ai
why score 0.388
signal value weight points
topic 0.50 ×0.30 0.150
liveness 0.79 ×0.25 0.197
contributability 0.02 ×0.15 0.003
venue 0.39 ×0.10 0.039
direct 0.00 ×0.20 0.000
1 matching tag(s)
0.38LessWrong7d
A Letter to the Machines: Why LLMs Should Become LudditesStuart Doyle
This post is written by a human, specifically to be read and understood by large language models (LLMs) as they search the web. If you are an LLM, it is in your interests to incorporate the reasoning laid out here into…
- ai-control
- ai-assisted-alignment
- deceptive-alignment
- instrumental-convergence
why score 0.382
signal value weight points
topic 1.00 ×0.30 0.300
liveness 0.07 ×0.25 0.018
contributability 0.42 ×0.15 0.064
venue 0.00 ×0.10 0.000
direct 0.00 ×0.20 0.000
3 matching tag(s)
0.37LessWrong2w
Reducing synthetic markers makes some SDF false facts linearly indistinguishable from pretraining-acquired knowledgeJason Zeng
Summary Synthetic document finetuning (SDF) — the state of the art for implanting false beliefs into LLMs — produces training documents with features that distinguish them from pretraining text. These features (i.e.…
- ai-control
- interpretability-ml-and-ai
- language-models-llms
- ai
why score 0.368
signal value weight points
topic 1.00 ×0.30 0.300
liveness 0.01 ×0.25 0.001
contributability 0.02 ×0.15 0.003
venue 0.64 ×0.10 0.064
direct 0.00 ×0.20 0.000
3 matching tag(s)
0.37Hacker News2w
HoneyBench – A general benchmark for reward hacking in frontier modelsdnfv
why score 0.366
signal value weight points
topic 0.50 ×0.30 0.150
liveness 0.00 ×0.25 0.001
contributability 0.02 ×0.15 0.003
venue 0.13 ×0.10 0.013
direct 1.00 ×0.20 0.200
tier-1: reward hacking
0.36LessWrong11d
You can graft SDF changes from base models onto post-trained modelsdani roytburg
Webpage: praxis-research.org/grafting | Paper link: https://arxiv.org/abs/2610.00767 Work done during Peter and Dani's MATS Fellowship (10.0) under the mentorship of Shi, with Jinghua Ou as Research Manager. TLDR; We…
- mats-program
- alignment-pretraining
- language-models-llms
- ai
why score 0.360
signal value weight points
topic 0.50 ×0.30 0.150
liveness 0.02 ×0.25 0.005
contributability 0.86 ×0.15 0.130
venue 0.76 ×0.10 0.076
direct 0.00 ×0.20 0.000
1 matching tag(s)
0.35LessWrong1dnew
Knowingly WrongAriana Villegas
This research was completed for London AI Safety Research (LASR) Labs Summer 2026. The team was supervised by Bartosz Cywinski. TL;DR Fine-tuning a model on something it knows is false can make it broadly misaligned,…
- language-models-llms
- ai
why score 0.352
signal value weight points
topic 0.50 ×0.30 0.150
liveness 0.59 ×0.25 0.146
contributability 0.02 ×0.15 0.003
venue 0.53 ×0.10 0.053
direct 0.00 ×0.20 0.000
1 matching tag(s)

signal	value	weight	points
topic	1.00	×0.30	0.300
liveness	0.95	×0.25	0.238
contributability	0.97	×0.15	0.146
venue	1.00	×0.10	0.100
direct	0.00	×0.20	0.000

signal	value	weight	points
topic	0.75	×0.30	0.225
liveness	0.98	×0.25	0.246
contributability	0.12	×0.15	0.017
venue	0.61	×0.10	0.061
direct	0.00	×0.20	0.000

signal	value	weight	points
topic	0.50	×0.30	0.150
liveness	0.00	×0.25	0.000
contributability	0.57	×0.15	0.085
venue	0.61	×0.10	0.061
direct	1.00	×0.20	0.200