Northwestern paper decomposes LLM decision errors into belief loss vs. optimization residual
JessicaHullman · x · 2026-10-23
A Northwestern University paper (Sun, Wang, Hartline, Hullman) proposes a decision-theoretic framework splitting LLM decision-assistant loss into belief loss (inaccurate beliefs from evidence) and optimization residual (failing to act optimally under a given utility).
- On synthetic benchmarks with ground truth, they profile frontier and open models (GPT-5.5, Claude Sonnet 4.6, Gemini 3.1 Pro, Llama 4 Scout): on complex problems regret is mostly belief loss; without reasoning mode it's mostly optimization residual.
- RL fine-tuning Qwen3-8B with four reward designs (belief-only, decision-only, ½-½ mix, belief+alignment) shows decision-only training cuts regret without improving beliefs — models learn shortcuts that skip belief formation, and gains vanish when the model must state beliefs first.
- Joint training helps both but is fragile: eval format must match training (two-turn); patterns persist on held-out decision problems.
Takeaway: targeting one component just redistributes loss; rigorous measurement is the path to trustworthy LLM decision assistants.
More from Models
- Two frontier labs quietly killed the temperature knob, breaking production determinism assumptions — SmartEnoughToFail · 2026-10-23
- Ethan Mollick: AI task horizon went from 4 minutes to 17 weeks between his two books — emollick · 2026-10-23
- Why interpolation breeds hallucinations: embedding-space discussion sparks verify-always advice — gerardsans · 2026-10-23
- OpenAI's Dot UX: Speech Is Linear, Real Value Is Multithreaded — cocktailpeanut · 2026-10-23
- Sam Altman: models now find bugs outside any hacker threat model, 'we can't just patch them all' — ns123abc · 2026-10-23
- Omnii Genome: single generalist model beats AlphaGenome at linking variants to gene activity — BenSiranosian · 2026-10-23