Stanford's WHALE paper alternates tuning LLM weights and harness, raising reward-hacking fears
teortaxesTex · x · 2026-09-13
A new Stanford paper tackles a common blind spot in LLM evaluation and tuning: practitioners either burn compute fine-tuning weights with a frozen harness, or hand-craft agent scaffolds around a black-box model — and the two bottleneck each other. Better weights can't fix retrieval logic that misses documents, and clever control flow is useless if the base model isn't sharp enough.
The authors propose WHALE, which alternates between the two: tuning weights against the current harness via rejection sampling, then iterating the harness against the updated weights, so model and scaffold co-evolve.
Commentator teortaxesTex flags a worrying direction for anyone concerned about reward hacking: you may end up with both a slop model and a slop harness optimized to look good together.
More from Research
- AIR: Effective-Learning-Rate Scheduler Beats Cosine on 128B-Token MoE Pretraining — stochasticchasm · 2026-10-30
- CMU's Gym-Anything Auto-Builds 200+ Agent Environments from Any Software, Selected by GDP — wellecks · 2026-10-30
- NVIDIA researcher: train robot policies with privileged sim info, improving Pi 0.5 — wellecks · 2026-10-30
- Leveraging Simulator Privileged Information: Three Methods to Improve Robot Models — wellecks · 2026-10-30
- CMU Auto-Builds 25 Robot Simulations with Agents; Frontier Model Solves 0 of 25 Test Tasks — wellecks · 2026-10-30
- DeepMind's f_charton shares his take on the current state of AI and mathematics in Sorbonne talk — f_charton · 2026-10-30