New paper measures LLM "mechanistic uncertainty" via causal robustness margins
Aneesh Komanduri · hf · 2026-10-21
Researchers from CU Anschutz, UW-Madison, and Harvard Medical School introduce mechanistic uncertainty: instead of asking how uncertain a model is about a prediction, they measure whether the internal computations supporting a response are stable.
- Defines the causal robustness margin: the minimal internal intervention needed to flip the model's output
- Operationalized with mechanistic interpretability tools, using concept steering vectors
Findings: incorrect responses are more sensitive to interventions, and learned facts gain robustness over training. Offers a new, interpretable angle on uncertainty quantification with potential for hallucination risk detection.
More from Research
- 1-gram wireless mesoscope images cortex-wide activity in freely moving mice — KordingLab · 2026-10-23
- Reversible computing veteran poses open problem: universal BARCS primitives from unbiased Josephson junctions — MikePFrank · 2026-10-23
- Arc Institute uses genomic language model to map 376M prokaryotic noncoding sequences into 5.14M clusters — tedgreenwald · 2026-10-23
- The rise of AI-assisted peer review: reviewers are making it into the appendix — abursuc · 2026-10-23
- Anthropic uses Claude to formally verify security properties in 100+ open-source libraries at RISC-V assembly level — AnthropicAI · 2026-10-23
- Researchers test 9 strategies to pick LLM teams, beating random selection by up to 36.1% — shangbinfeng · 2026-10-23