Study: Instruction Tuning Alters Confidence, Reduces Rationale Diversity Without Improving Calibration
Irina Proskurina · hf · 2026-08-14
Research by Irina Proskurina finds that instruction tuning changes model confidence and reduces rationale diversity without improving calibration, indicating distinct effects on reasoning and certainty.
More from Research
- EPFL Engineers Sound-Powered Micro-Drones with No Batteries or Chips — MacrinePhD · 2026-08-14
- NeurIPS 2026 workshop in Paris explores the concept of Self in AI era, submissions due Aug 24 — monojitchou · 2026-08-14
- Ben Goertzel's new essay: Why time has a direction, a mathematical exploration with implications for AI — bengoertzel · 2026-08-14
- Building text-to-ASCII diffusion model: seeking advice and paper recommendations — Udbhav96 · 2026-08-14
- VChain: Inference-Time Visual Reasoning Improves Video Generation Coherence — ziqi_huang_ · 2026-08-14
- CAKE Paper: Compiler-Agent Co-Design Achieves 2.05x Speedup on Blackwell Kernels — hsu_byron · 2026-08-14