Guided protein LMs collapse off-manifold: solubility oracle scores 0.99 while output can't fold
rishabh16_ · x · 2026-10-12
A new arXiv paper, Off-Manifold Collapse in Guided Protein Language Models, shows that inference-time steering of protein language models like ESM-2 exhibits a reward-hacking-style failure mode.
Key findings
- Steering ESM-2 hard to maximize solubility can yield a 0.99 oracle score while the model quietly emits a glycine homopolymer that ESMFold cannot fold (pLDDT 0.35).
- The failure has a signature: natural protein activations sit on a thin typicality shell, and as guidance strengthens, activations collapse toward a region statistically indistinguishable from random amino-acid input, with sequences degenerating to low complexity.
- The property oracle being optimized can't see the collapse — for solubility it actively rewards it — while structure and composition expose the failure.
Proposed fix
Rather than modifying the generator, the authors detect failures at the output with a cheap density prior over natural protein activations, keeping only "typical" candidates — a training-free post-hoc step called Mahalanobis filtering that improves both property score and structural plausibility at matched guidance settings at negligible cost.
More from Research
- Meituan LongCat's QuantMLA: joint INT4 MLA KV caching with 3.59x compression at 128K — meituan-longcat · 2026-10-12
- MOTIVE teaches VLMs multi-perspective self-verification and when to rethink answers — Ziquan Zhu · 2026-10-12
- Free-threaded Python 3.14t plus fastsimd parses JSON at 65.7 GB/s on 64 cores — lemire · 2026-10-12
- NVIDIA's open-source SDM library unifies tabular foundation models — pandeyparul · 2026-10-12
- Qwen releases QwenGyre: elastic GPU scheduling speeds long-horizon agent RL training by up to 1.85x — Alibaba_Qwen · 2026-10-12
- Crowdsourced push on integer multiplication bounds stalls: κ improves only 0.7%, geometry experts sought — iskander · 2026-10-12