Beyond Bias Scores: Answerability Confounds in Activation Steering with Gradient-Constrained Direction Design (GCD)
Bias scores in language-model evaluations often conflate what a model prefers with whether it answers at all.
ProofPaper ↗
Key points
- We identify this confound in BBQ, whose ambiguous score factorizes into answerability and signed preference.
- We introduce Gradient-Constrained Direction Design (GCD), an option-gradient construction with a first-order unknown-logit constraint.
- Under a matched Qwen-7B protocol, GCD reduces answer-rate distortion 14-fold and yields a 4-fold larger preference change, while ablations show similar gains for alternative and unconstrained option-gradient directions; nulling the unknown gradient alone is not responsible for the gain.
- We recommend reporting answerability alongside published scores.
Sources (1)
- [1]Beyond Bias Scores: Answerability Confounds in Activation Steering with Gradient-Constrained Direction Design (GCD)arXiv (AI, ML, NLP, CV, robotics, multi-agent) · Oct 13, 03:09 AM
Bias scores in language-model evaluations often conflate what a model prefers with whether it answers at all.
We identify this confound in BBQ, whose ambiguous score factorizes into answerability and signed preference.
Extractive summary: sentences quoted from the sources.
Before this
- Oct 13, 2026Cowork-Gym: Synthesizing Interactive Work Environments at Scale
- Oct 13, 2026Can an Agent Learn to Predict Its Own Failure? Outcome-Grounded Confidence Calibration in LLM Agents
- Oct 12, 2026Alibaba Qwen Introduces QwenGyre: Elastic Reinforcement Learning RL for XLong-Horizon Agents
- Oct 12, 2026AReA: Adaptive Receptive-Field Attention for Efficient High-Resolution Diffusion Transformers
- Oct 12, 2026TPPO: Tracking-Value PPO For Language Model Reasoning
- Oct 11, 2026autotrust/JEV-35B-NVFP4