Paper: Predicting LLM Compositional Failures from Feature Geometry, No Eval Inputs Needed
nsaphra · x · 2026-11-07
A new arXiv paper, Adversarial Concept Search by Jennifer Meng Lu, Naomi Saphra et al., uses an LLM's representational geometry to predict which concept combinations it will fail on, attributing compositional errors to interference between salient features.
Key findings:
- Across toy programmatic tasks, multihop reasoning, and multilingual factual recall, near-orthogonal linear encodings of concept pairs lead to reliable composition;
- When encodings are close and interfere, the model fails to compose them;
- The method anticipates failure modes without evaluating specific inputs, enabling targeted stress tests and scalable active learning.
The authors are preparing an ICLR version with more models and random-seed controls for non-geometric explanations.
More from Research
- New Benchmarks DecoyCount and ScienceVision Probe Whether Vision Models Fail at Seeing or Reasoning — AndrewDai · 2026-11-07
- BlancheMinerva argues withholding pre-camera-ready collaboration harms scientific practice — BlancheMinerva · 2026-11-07
- Aaronson-Ambainis conjecture caps query complexity, not super-polynomial quantum speedups in computation — thomasahle · 2026-11-07
- Flower Computers publicly releases encoders, billed as the fastest way to embed text semantically — matdryhurst · 2026-11-07
- HKUST and Tencent unveil Möbius: unified motion-language model beats 3B specialists at 0.6B — bdsqlsz · 2026-11-07
- Interactive orbifold demo: draw on 17 symmetry surfaces and unfold them, built with Claude — yaroslavvb · 2026-11-07