Kempner Institute's Adversarial Concept Search predicts LLM failure modes from representational geometry
nsaphra · x · 2026-11-07
The Kempner Institute's Deeper Learning blog introduces Adversarial Concept Search (ACS), a method that uses a model's representational geometry to anticipate which scenarios an LLM will likely fail on, letting evaluation efforts focus on probable failure cases.
Key insight shared by the author: at every scale checked so far, the better the model—and the better it is at a given task—the better ACS works at surfacing its failure modes, meaning the method improves as capabilities grow. A short blog post overview and the preprint are linked.
More from Research
- Q-OIL, an undergrad first paper at CoRL2026, treats human interventions as soft exploration guidance — aidenybai · 2026-11-07
- Andrew Carr reignites the MoE debate: do tokens choose experts or experts choose tokens? — andrew_n_carr · 2026-11-07
- AdderBoard challenge: smallest transformer that adds two 10-digit numbers, records keep falling — DimitrisPapail · 2026-11-07
- Zellige AI open-sources Tessera-27B-B-CJK-RL: RL-boosted CJK-thinking coder cuts thinking tokens by a third, matches base on hard math — ghgi_ · 2026-11-07
- Unverified claim: frontier labs skewing ICLR reviews with malicious reviews, scandal brewing — examachine · 2026-11-07
- Persona Atlas visualizes Qwen's experts and heads, exposing bizarre inner personas — Sauers_ · 2026-11-07