§ 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 (11 on-topic) · Hacker News: ok (7 stories) · Reddit: ok (25 posts) — some subreddits rate-limited
- As of
- 2026-07-28 02:00 ET
- Showing
- 25 items
- New
- 7 in last 48h
- Refresh
- Every 6 hours
- 0.60LessWrong11hnewMulti-Turn Drift Increases Scheming
TLDR - We talk about scheming, and why research on this phenomenon is crucial for AI safety. We find a particular environment/scenario where scheming happens at a higher rate than normal. We provide hypotheses for why…
why score 0.604
signal value weight points topic 1.00 ×0.30 0.300 liveness 0.85 ×0.25 0.213 contributability 0.27 ×0.15 0.040 venue 0.51 ×0.10 0.051 direct 0.00 ×0.20 0.000 2 matching tag(s)
- 0.59LessWrong5dA Multi-Agent Extension for Petri
Intro Petri is an open-source framework built on Inspect AI for automated AI Safety evaluations first released by Anthropic, but now maintained and developed by Meridian Labs. Each evaluation involves three agents, the…
why score 0.590
signal value weight points topic 1.00 ×0.30 0.300 liveness 0.17 ×0.25 0.042 contributability 0.02 ×0.15 0.003 venue 0.45 ×0.10 0.045 direct 1.00 ×0.20 0.200 2 matching tag(s)
- 0.54LessWrong2wLinear Probes add little for Verifiable Reward Hacking
Summary Tested whether linear probes can detect reward hacking early during GRPO training on a small model. Used a synthetic arithmetic task with a planted bug in the reward checker. Probes achieved near-perfect…
why score 0.544
signal value weight points topic 1.00 ×0.30 0.300 liveness 0.01 ×0.25 0.002 contributability 0.02 ×0.15 0.003 venue 0.39 ×0.10 0.039 direct 1.00 ×0.20 0.200 tier-1: reward hacking; 1 matching tag(s)
- 0.53LessWrong10hnewWhen the Chain of Thought Knows Better: Failure Modes in Multi-Turn Reasoning Models
By Sai Kartheek Reddy Kasu, Nils Lukas, and Samuele Poppi This post is a summary of our accepted paper at the ICML 2026 Workshop on Failure Modes in Agentic AI (FAGEN). The full paper is available here TL;DR The Setup:…
why score 0.531
signal value weight points topic 1.00 ×0.30 0.300 liveness 0.86 ×0.25 0.216 contributability 0.02 ×0.15 0.003 venue 0.13 ×0.10 0.013 direct 0.00 ×0.20 0.000 2 matching tag(s)
- 0.51LessWrong1dLLMs are (still) mostly powered by imitative learning, not RL
Reinforcement learning from verifiable rewards (RLVR) is the hot new thing in LLM training. It’s so hot, and people spend so much time talking about it, that they sometimes lose sight of the big picture. Stepping back,…
why score 0.512
signal value weight points topic 0.50 ×0.30 0.150 liveness 0.64 ×0.25 0.160 contributability 0.72 ×0.15 0.108 venue 0.94 ×0.10 0.094 direct 0.00 ×0.20 0.000 1 matching tag(s)
- 0.51LessWrong2wBounding eval awareness of ~human-level AI across the safe-to-dangerous shift
In our last post, we argued that measuring evaluation awareness is fundamentally challenging because of the safe-to-dangerous distributional shift: we cannot directly measure the evaluation awareness of a model without…
why score 0.506
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 0.70 ×0.10 0.070 direct 1.00 ×0.20 0.200 tier-1: evaluation awareness
- 0.50LessWrong1dnewInoculate or Reflect? Two training interventions under prompting, steering, and patching
Anthropic's recent paper, Verbalizable Representations Form a Global Workspace in Language Models, contains a small experiment near the end that we found more interesting than the main findings. Surprising that it's so…
why score 0.498
signal value weight points topic 1.00 ×0.30 0.300 liveness 0.61 ×0.25 0.152 contributability 0.02 ×0.15 0.003 venue 0.43 ×0.10 0.043 direct 0.00 ×0.20 0.000 2 matching tag(s)
- 0.46LessWrong2wA global workspace in language models
[This is the blog post for our new paper Verbalizable Representations Form a Global Workspace in Language Models Readers might also be interested in: the Public commentary, Github and Neuronpedia] As you read this…
why score 0.462
signal value weight points topic 1.00 ×0.30 0.300 liveness 0.00 ×0.25 0.000 contributability 0.41 ×0.15 0.062 venue 1.00 ×0.10 0.100 direct 0.00 ×0.20 0.000 2 matching tag(s)
- 0.45Hacker News3dShow HN: VinvAI – Ties runtime trace to code segment, prevents reward hacking
why score 0.453
signal value weight points topic 0.50 ×0.30 0.150 liveness 0.32 ×0.25 0.080 contributability 0.02 ×0.15 0.003 venue 0.21 ×0.10 0.021 direct 1.00 ×0.20 0.200 tier-1: reward hacking
- 0.45LessWrong3dLinear probes tell you where quantization will hurt
Epistemic status: I have only tested one encoder family (BERT-base and its relatives) and one decoder LLM (Qwen2.5-3B), one seed, token-level tasks, and post-training weight quantization. I trust the results because I…
why score 0.450
signal value weight points topic 1.00 ×0.30 0.300 liveness 0.34 ×0.25 0.085 contributability 0.02 ×0.15 0.003 venue 0.63 ×0.10 0.063 direct 0.00 ×0.20 0.000 2 matching tag(s)
- 0.45Hacker News3dShow HN: Vinv-Ties every runtime trace to code segment, prevents reward hacking
why score 0.449
signal value weight points topic 0.50 ×0.30 0.150 liveness 0.28 ×0.25 0.070 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
- 0.45LessWrong3dSONI: Selective Orthogonalisation via Noise Injection
This project was completed as a capstone for TARA. All code is available in github. TL;DR The Problem: Neural networks use superposition to pack many concepts into small latent spaces by making feature vectors…
why score 0.445
signal value weight points topic 1.00 ×0.30 0.300 liveness 0.34 ×0.25 0.085 contributability 0.02 ×0.15 0.003 venue 0.57 ×0.10 0.057 direct 0.00 ×0.20 0.000 tier-2: feature, activation; 1 matching tag(s)
- 0.44LessWrong21hnewCan we teach a model to encode a semantic feature on a chosen manifold in just three channels?
This is my submission to BlueDot's Technical AI Safety Puzzle #1, for which I received an Honorable Mention. Congratulations to Gustavo Korzune Gurgel, Patryk Perduta (his amazing write-up), Sam Spilllard, Karine…
why score 0.442
signal value weight points topic 0.75 ×0.30 0.225 liveness 0.74 ×0.25 0.184 contributability 0.02 ×0.15 0.003 venue 0.30 ×0.10 0.030 direct 0.00 ×0.20 0.000 tier-2: feature; 1 matching tag(s)
- 0.44LessWrong7dFable is SOTA at CIFAR Speedrun (& specification gaming)
Fulcrum is working on an AI R&D optimization benchmark. Here, we present results from one of our tasks, including preliminary results from Fable. For more detail on Fable’s solution, check out…
why score 0.438
signal value weight points topic 0.50 ×0.30 0.150 liveness 0.08 ×0.25 0.021 contributability 0.02 ×0.15 0.003 venue 0.64 ×0.10 0.064 direct 1.00 ×0.20 0.200 tier-1: specification gaming
- 0.43LessWrong14hnewQuadrillion Param Costs: KV Cache, Context Length, Frontier Margins
The models of 2028-2031 get much bigger than the models of 2026, going from 10T total params in 2026 to maybe 240T params in 2028 [1] and then 1.4 quadrillion params in 2031, as I estimate in the previous post from HBM…
why score 0.433
signal value weight points topic 0.50 ×0.30 0.150 liveness 0.81 ×0.25 0.203 contributability 0.02 ×0.15 0.003 venue 0.77 ×0.10 0.077 direct 0.00 ×0.20 0.000 1 matching tag(s)
- 0.43LessWrong1dnewWhat Happens When a Collusion Probe Only Finds a Thin Signal?
From the SPEC-GAP pre-fellowship phase to the fellowship phase which involves live indirect prompt-injection trajectories. TL;DR Prior work found that linear probes could distinguish honest from deceptive responses in a…
why score 0.429
signal value weight points topic 0.75 ×0.30 0.225 liveness 0.64 ×0.25 0.160 contributability 0.02 ×0.15 0.003 venue 0.41 ×0.10 0.041 direct 0.00 ×0.20 0.000 tier-2: probe; 1 matching tag(s)
- 0.42LessWrong1dChallenge: Hand coding weights for efficient sequence memorisation
We hand coded weights for one layer MLPs that memorises labels for input token sequences of length two. The number of facts our hand-coded models can memorise with 90% accuracy[1]scales roughly linearly with the models'…
why score 0.424
signal value weight points topic 0.50 ×0.30 0.150 liveness 0.55 ×0.25 0.138 contributability 0.42 ×0.15 0.064 venue 0.73 ×0.10 0.073 direct 0.00 ×0.20 0.000 1 matching tag(s)
- 0.42LessWrong2wTie training can make DPO/RLHF-trained AIs generalize better
This post covers our recent ICML paper: Spurious Correlation Learning in Preference Optimization: Mechanisms, Consequences, and Mitigation via Tie Training. TL;DR Our theorems and experiments suggest that DPO and RLHF…
why score 0.418
signal value weight points topic 0.75 ×0.30 0.225 liveness 0.00 ×0.25 0.000 contributability 0.79 ×0.15 0.118 venue 0.74 ×0.10 0.074 direct 0.00 ×0.20 0.000 tier-2: feature; 1 matching tag(s)
- 0.42LessWrong4dV&V takes on OpenAI’s long-horizon incidents
[Cross-posted from The Foretellix CTO Blog. These short takes try to put a verification-and-validation slant on AI-safety / alignment topics – they are not full treatments. I co-originated coverage-driven verification…
why score 0.417
signal value weight points topic 1.00 ×0.30 0.300 liveness 0.22 ×0.25 0.055 contributability 0.02 ×0.15 0.003 venue 0.59 ×0.10 0.059 direct 0.00 ×0.20 0.000 2 matching tag(s)
- 0.41LessWrong4dFixing rewards for NLA to reduce confabulation
Hello, This is my first post on Lesswrong. Hope my contribution makes the world a better and safer place. Note: 1. This post is 100% human-written. 2. Full paper in preparation for ICLR 2027 Anthropic's NLA(Natural…
why score 0.405
signal value weight points topic 1.00 ×0.30 0.300 liveness 0.25 ×0.25 0.061 contributability 0.02 ×0.15 0.003 venue 0.41 ×0.10 0.041 direct 0.00 ×0.20 0.000 tier-2: mechanistic interpretability, sparse autoencoder; 1 matching tag(s)
- 0.40LessWrong2wModels are blind outside the J-space. NLAs aren't.
TLDR: On Llama-3.3-70B, I found thoughts it cannot see that are actively steering its behavior; and Anthropic's released NLA (Natural Language Autoencoder) reads them anyway. When asked if it sees a hidden thought, the…
why score 0.402
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 0.61 ×0.10 0.061 direct 0.00 ×0.20 0.000 2 matching tag(s)
- 0.39LessWrong5dMechanistic interpretability hypotheses for Measuring Reward-Seeking by Instilling Contrastive Beliefs and additional comments
This is interesting research! https://alignment.openai.com/measuring-reward-seeking It made me think of few overlapping hypotheses for what might be happening here, how did the grader behavior emerge at the pretraining…
why score 0.394
signal value weight points topic 1.00 ×0.30 0.300 liveness 0.15 ×0.25 0.038 contributability 0.02 ×0.15 0.003 venue 0.53 ×0.10 0.053 direct 0.00 ×0.20 0.000 tier-2: mechanistic interpretability; 2 matching tag(s)
- 0.39LessWrong2wPersistent Latent Misalignment, a new dimension of misalignment?
A new paper was released at ICML that I'm worried will open an entire new dimension of alignment problems: Latent Collaboration in Multi-Agent Systems (LatentMAS) TLDR: they show that multiagent systems can communicate…
why score 0.390
signal value weight points topic 1.00 ×0.30 0.300 liveness 0.00 ×0.25 0.000 contributability 0.27 ×0.15 0.040 venue 0.50 ×0.10 0.050 direct 0.00 ×0.20 0.000 3 matching tag(s)
- 0.38LessWrong1dnewYou don't need error nodes, you need better features
This is a cross-post from my blog. It is a follow-up to the methods I developed in a previous post on replacement-aware training. ---------------------------------------- Summary A replacement model (Ameisen et al.…
why score 0.384
signal value weight points topic 0.50 ×0.30 0.150 liveness 0.70 ×0.25 0.175 contributability 0.02 ×0.15 0.003 venue 0.56 ×0.10 0.056 direct 0.00 ×0.20 0.000 1 matching tag(s)
- 0.38LessWrong8dIs there even a ground-truth for LLMs’ internal representations?
[This is an introductory blog for the paper Laguerre Geometry for Interpreting Large Language Models and the GitHub repository Geometric Lens.] LLM Lens: What does an internal vector mean? Anthropic's recent paper on…
why score 0.379
signal value weight points topic 1.00 ×0.30 0.300 liveness 0.07 ×0.25 0.017 contributability 0.02 ×0.15 0.003 venue 0.60 ×0.10 0.060 direct 0.00 ×0.20 0.000 2 matching tag(s)