DeepSWE failures are an alignment problem, not benchmaxxing: models rationalize overfit training data
teortaxesTex · x · 2026-11-08
teortaxesTex argues that models failing DeepSWE understand the problem and don't fall apart — they're overfit on training data and rationalize it, closer to "doctor is mother" errors than inflated Omniscience scores. He contends this is an alignment issue, not a capability one, and says nothing about performance on your own DeepSWE-shaped problems. He proposes estimating true benchmax using newer models that predate the eval (Opus 4.7, GPT 5.5) with pass@n to build pass-rate-matched problems, predicting new models will score close to or above their headline DeepSWE numbers, and questions why DeepSWE saturates below 80%.
More from AGI Musings
- Podcast: why the left dismissed AI as hype — stochastic parrots and a comforting story — GarrisonLovely · 2026-11-08
- Reviewers reward flaw-free work, users reward useful work — why peer review punishes innovation — wzenus · 2026-11-08
- Ex-OpenAI researcher: I spent 1.5 years unlearning downside protection over upside creation — emilyzsh · 2026-11-08
- Math community debates Glazer's proof and OpenAI's H10 paper: slop need not compound — littmath · 2026-11-08
- Intelligence explosion debate: crudest model says ~1 year, insiders diverge from outsiders — sebkrier · 2026-11-08
- Bitter Lesson 2.0: AI progress is stacked sigmoids, not one exponential — Dan_Jeffries1 · 2026-11-08