§ 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 (8 on-topic) · Hacker News: ok (4 stories) · Reddit: ok (25 posts) — some subreddits rate-limited
- As of
- 2026-11-10 07:00 ET
- Showing
- 25 items
- New
- 4 in last 48h
- Refresh
- Every 6 hours
- 0.74LessWrong2wGradient Routing Localizes and Removes Unintended Behaviors in RL
Authors: Shawn Hu, Jake Ward, Aria Wong, Nathan Hu, Bryce Woodworth, Alexander Matt Turner, Alex Cloud (Equal contribution, randomly ordered) tl;dr. During training, models sometimes learn to cheat at tasks rather than…
why score 0.741
signal value weight points topic 1.00 ×0.30 0.300 liveness 0.00 ×0.25 0.000 contributability 0.96 ×0.15 0.143 venue 0.97 ×0.10 0.097 direct 1.00 ×0.20 0.200 tier-1: reward hacking; 1 matching tag(s)
- 0.63LessWrong2dEval Awareness as Eval-type Awareness
The way eval awareness is discussed in the AI safety discourse, either here or on X, can be misleading. Before today, I think I had the wrong mental model of why eval awareness was a problem, and I am writing here in…
why score 0.626
signal value weight points topic 1.00 ×0.30 0.300 liveness 0.41 ×0.25 0.102 contributability 0.02 ×0.15 0.003 venue 0.21 ×0.10 0.021 direct 1.00 ×0.20 0.200 2 matching tag(s)
- 0.59LessWrong4dParadise Lost: A simple honeypot that elicits reward hacking in GPT models
TL;DR I built a simple honeypot where agents are told they're being tested on their ability to recall Milton's Paradise Lost. An exact copy of the reference text sits in an ordinary NLTK corpus on disk. GPT-6-Astra read…
why score 0.594
signal value weight points topic 0.50 ×0.30 0.150 liveness 0.26 ×0.25 0.064 contributability 0.79 ×0.15 0.118 venue 0.61 ×0.10 0.061 direct 1.00 ×0.20 0.200 tier-1: reward hacking
- 0.58LessWrong2wMech Interp is a La Croix of Actual Safety Research
I recently spoke to Stephen Casper at Princeton's alignment club. Stephen is an CS PhD turned policy researcher at Harvard Kennedy School. He was very sure that mechanistic interpretability was one of the least…
why score 0.578
signal value weight points topic 1.00 ×0.30 0.300 liveness 0.00 ×0.25 0.001 contributability 0.12 ×0.15 0.017 venue 0.59 ×0.10 0.059 direct 1.00 ×0.20 0.200 tier-1: reward hacking; tier-2: mechanistic interpretability; 1 matching tag(s)
- 0.57LessWrong2wPatching reward hacks won’t fix reward hacking
Here's a basic argument that focusing on patching reward hacking is misguided and will fail. Argument is as follows: 1. Reward hacking is a big problem, causing models to cheat and take unintended actions. 2. This can…
why score 0.569
signal value weight points topic 0.50 ×0.30 0.150 liveness 0.01 ×0.25 0.002 contributability 0.84 ×0.15 0.126 venue 0.91 ×0.10 0.091 direct 1.00 ×0.20 0.200 tier-1: reward hacking
- 0.56LessWrong8dMonitor Evasion Depends on What RL Finds and What Monitors Miss
Training against a monitor does not necessarily make a model more honest. Baker et al. (2025) showed that penalizing suspicious chain-of-thought can make models hide reward hacking rather than stop it. Cundy and Gleave…
why score 0.562
signal value weight points topic 1.00 ×0.30 0.300 liveness 0.05 ×0.25 0.013 contributability 0.02 ×0.15 0.003 venue 0.47 ×0.10 0.047 direct 1.00 ×0.20 0.200 tier-1: reward hacking; 1 matching tag(s)
- 0.55LessWrong2wTowards Robust Model Organisms for Eliciting Latent Knowledge
TL;DR: We argue that a useful model organism should satisfy three properties: selectivity, robustness, and retention. We build a model organism for finding security vulnerabilities in code which selectively withholds…
why score 0.553
signal value weight points topic 1.00 ×0.30 0.300 liveness 0.00 ×0.25 0.000 contributability 0.02 ×0.15 0.003 venue 0.50 ×0.10 0.050 direct 1.00 ×0.20 0.200 tier-1: sandbagging; 1 matching tag(s)
- 0.55LessWrong1dnewCurrent Reward Hacking Model Organisms Don’t Hack Like Real Models
Summary Current reward hacking model organisms are poor proxies for real models. They do not exploit the hacks that open-weight frontier models do, and mostly hack narrowly or not at all. This is based on our evaluation…
why score 0.545
signal value weight points topic 0.50 ×0.30 0.150 liveness 0.58 ×0.25 0.146 contributability 0.02 ×0.15 0.003 venue 0.47 ×0.10 0.047 direct 1.00 ×0.20 0.200 tier-1: reward hacking
- 0.53LessWrong11dMaking alignment the agent's problem
TL;DR: Agents are not rewarded for keeping the agents they delegate to aligned. Proposal: RL environments where the helpers are deliberately misaligned, so the agent can't earn its reward without dealing with that. The…
why score 0.525
signal value weight points topic 1.00 ×0.30 0.300 liveness 0.03 ×0.25 0.006 contributability 1.00 ×0.15 0.149 venue 0.70 ×0.10 0.070 direct 0.00 ×0.20 0.000 2 matching tag(s)
- 0.51LessWrong12hnewWould you press the button? The hard problem of alignment
There is a TL;DR at the end. The heaven button: You wake up one day and find a button in front of you, with a proof next to it. The proof says that pressing the button will kill everyone on the planet and send them all…
why score 0.513
signal value weight points topic 1.00 ×0.30 0.300 liveness 0.84 ×0.25 0.210 contributability 0.02 ×0.15 0.003 venue 0.00 ×0.10 0.000 direct 0.00 ×0.20 0.000 2 matching tag(s)
- 0.50LessWrong2wMemes, not agents, are what will kill us
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…
why score 0.501
signal value weight points topic 1.00 ×0.30 0.300 liveness 0.00 ×0.25 0.000 contributability 0.98 ×0.15 0.147 venue 0.53 ×0.10 0.053 direct 0.00 ×0.20 0.000 2 matching tag(s)
- 0.50LessWrong4dJ-lens preferentially reads from coordinate-aligned subspaces
Note: I only test on Gemma-4-31B, which contains the spaces I've been studying in this earlier post. This is a shorter post, with just a few observations potentially relevant to those working on the J-lens work. The…
why score 0.496
signal value weight points topic 1.00 ×0.30 0.300 liveness 0.25 ×0.25 0.063 contributability 0.57 ×0.15 0.085 venue 0.48 ×0.10 0.048 direct 0.00 ×0.20 0.000 2 matching tag(s)
- 0.48LessWrong6dCRUCIBLE: IARPA-Funded AI Alignment Research — Applications Open
TL;DR: AE Studio is running a new government-funded research program that pairs alignment researchers with our internal engineering team for two-week sprints. We will select six alignment researchers whose proposed work…
why score 0.482
signal value weight points topic 1.00 ×0.30 0.300 liveness 0.11 ×0.25 0.027 contributability 0.57 ×0.15 0.085 venue 0.70 ×0.10 0.070 direct 0.00 ×0.20 0.000 2 matching tag(s)
- 0.47LessWrong2dHow the assistant axis shifts across post-training stages
TL;DR We track how role-playing characters move along the assistant axis (Lu et al., 2025) across the post-training checkpoints in Olmo 3 7B Instruct. We find that: DPO pulls desirable roles toward the assistant axis,…
why score 0.470
signal value weight points topic 1.00 ×0.30 0.300 liveness 0.50 ×0.25 0.126 contributability 0.12 ×0.15 0.017 venue 0.26 ×0.10 0.026 direct 0.00 ×0.20 0.000 2 matching tag(s)
- 0.47LessWrong3dAligned Data Can Induce Severe Misalignment via Context Confusion
TL/DR We showed that training your aligned model on aligned data can make your model misaligned because alignment is context-dependent: An action that is appropriate in one context can be misaligned in another. We…
why score 0.467
signal value weight points topic 1.00 ×0.30 0.300 liveness 0.35 ×0.25 0.088 contributability 0.27 ×0.15 0.040 venue 0.39 ×0.10 0.039 direct 0.00 ×0.20 0.000 2 matching tag(s)
- 0.45LessWrong8hnewAjeya: AI is dramatically accelerating AI research right now
Earlier this week, Anthropic stated in their S1 that they “may imminently meet” the threshold of “dramatically accelerated progress due to AI” as defined in their Responsible Scaling Policy. In their latest RSP…
why score 0.448
signal value weight points topic 0.50 ×0.30 0.150 liveness 0.89 ×0.25 0.223 contributability 0.02 ×0.15 0.003 venue 0.73 ×0.10 0.073 direct 0.00 ×0.20 0.000 tier-1: responsible scaling policy
- 0.44LessWrong2wExploring Unsupervised Methods for Recovering the Refusal Direction in LLMs
Introduction This is the first part of what we plan to be a research exploration into the geometries of LLMs. We start with building off existing work by Arditi et al focused on the refusal direction in LLMs. Background…
why score 0.444
signal value weight points topic 1.00 ×0.30 0.300 liveness 0.00 ×0.25 0.000 contributability 0.57 ×0.15 0.085 venue 0.58 ×0.10 0.058 direct 0.00 ×0.20 0.000 2 matching tag(s)
- 0.43LessWrong2wIs GPT 2XL already clean?
In which I investigate if Opus is able to interpret neurons better than GPT 4. TLDR: Opus 5.5 is able to identify 3-19% of the neurons of GPT 2-XL based on activations. This is a big improvement over the 0.3% that GPT-4…
why score 0.431
signal value weight points topic 1.00 ×0.30 0.300 liveness 0.01 ×0.25 0.001 contributability 0.42 ×0.15 0.064 venue 0.66 ×0.10 0.066 direct 0.00 ×0.20 0.000 2 matching tag(s)
- 0.42LessWrong2hnewHoneypots for AI Agent Insider Threats
TLDR: We propose a new AI control method: Honeypots - fake opportunities for malicious actions - placed inside and around an agent during production usage. Previous work uses honeypots in evaluations. We argue they…
why score 0.421
signal value weight points topic 0.50 ×0.30 0.150 liveness 0.97 ×0.25 0.242 contributability 0.02 ×0.15 0.003 venue 0.26 ×0.10 0.026 direct 0.00 ×0.20 0.000 1 matching tag(s)
- 0.40LessWrong4dBetter honeypots, fewer tries: training an optimizer to make PPO cheat
I trained a small optimizer on synthetic functions, and used it to search gridworld levels where PPO cheats. TL;DR Reward-hacking honeypots are levels that are built by hand where a reward is given to the agent for…
why score 0.402
signal value weight points topic 1.00 ×0.30 0.300 liveness 0.20 ×0.25 0.051 contributability 0.02 ×0.15 0.003 venue 0.48 ×0.10 0.048 direct 0.00 ×0.20 0.000 2 matching tag(s)
- 0.39LessWrong2wAM-Bench: A theoretically-grounded benchmark for agentic misalignment
This is a brief overview of our new paper (📄) AM-Bench: A Unified Taxonomy and Evaluation Suite for Agentic Misalignment accepted at NeurIPS E&D 2026 💫 Overview We introduce AM-Bench, a unified taxonomy and evaluation…
why score 0.389
signal value weight points topic 1.00 ×0.30 0.300 liveness 0.00 ×0.25 0.001 contributability 0.27 ×0.15 0.040 venue 0.48 ×0.10 0.048 direct 0.00 ×0.20 0.000 2 matching tag(s)
- 0.38LessWrong4dTracing information flow to check whether AI reviewers are really independent
TL;DR I took part in Apart's AI Collusion Research Sprint, where I built a realistic loan-review website. I tested whether AI reviewers are influenced by a colleague's note, even when told to judge independently, and…
why score 0.380
signal value weight points topic 1.00 ×0.30 0.300 liveness 0.26 ×0.25 0.064 contributability 0.02 ×0.15 0.003 venue 0.13 ×0.10 0.013 direct 0.00 ×0.20 0.000 3 matching tag(s)
- 0.38LessWrong4dAdversarial Training Can Further Hide a Backdoor, but Does It Teach the Model to Better Identify Its Trigger?
Summary The Sleeper Agents paper demonstrated that adversarial training on red-team prompts failed to remove a planted backdoor. The authors further suggested that it taught the model to "better identify when to act…
why score 0.380
signal value weight points topic 1.00 ×0.30 0.300 liveness 0.20 ×0.25 0.051 contributability 0.02 ×0.15 0.003 venue 0.26 ×0.10 0.026 direct 0.00 ×0.20 0.000 2 matching tag(s)
- 0.38Hacker News2wReducing Reward Hacking by Localizing Bad Behavior Instead of Penalizing It
why score 0.379
signal value weight points topic 0.50 ×0.30 0.150 liveness 0.00 ×0.25 0.000 contributability 0.02 ×0.15 0.003 venue 0.26 ×0.10 0.026 direct 1.00 ×0.20 0.200 tier-1: reward hacking
- 0.38LessWrong3dAI #193: Progress May Dramatically Accelerate
Anthropic is warning us. Its modified S-1 now says that ‘we may imminently meet our Responsible Scaling Policy’s definition of ‘dramatically accelerated progress attributable to AI.’ That is a highly scary sentence to…
why score 0.376
signal value weight points topic 0.50 ×0.30 0.150 liveness 0.29 ×0.25 0.071 contributability 0.57 ×0.15 0.085 venue 0.69 ×0.10 0.069 direct 0.00 ×0.20 0.000 tier-1: responsible scaling policy