FairStress follow-up: 9 debiasing methods mostly fail to fix LLM overcorrection
allen_ai · x · 2026-10-21
FairStress thread follow-up: the AI2 team examined how training data influences LLM decision biases, finding that models are never really taught fair decision-making yet are often aligned to prevent discrimination — which may explain overcorrection patterns. Of 9 debiasing approaches tested, most didn't reliably improve symmetric robustness and some hurt performance. Older benchmarks like BBQ can miss overcorrection, while FairStress reveals the hidden asymmetries they overlook.
More from Research
- Aviation tracks every part's provenance — agents can reason over that graph — colinmcnamara · 2026-10-21
- SFT details: instruction/reasoning/code/math mix, synthetic + human traces, no epoching — TimDarcet · 2026-10-21
- Mid-training recipe disclosed: 32k→131k context extension, dropping logit distillation for tokens — TimDarcet · 2026-10-21
- Mathematician Daniel Litt concedes his AI bet, but says AI won't solve math — ziv_ravid · 2026-10-21
- AI drug discovery reality check: 'give me everything that binds,' then filter for toxicity — burny_tech · 2026-10-21
- PyTorch details KTIR and Dataflow Scheduler behind IBM's Spyre accelerator backend — PyTorch · 2026-10-21