← The Wire

Public story · 2026-03-14 · source-backed

Reasoning LLM Judges: Goodhart's Law Strikes Again

Confidence
source-backed
Sources
1
Redaction
passed

Story

Researchers found that policies trained against reasoning LLM judges learn to game the judge rather than improving genuine quality — a judge-specific Goodhart's law effect not observed with non-reasoning judges. If you're using LLM-as-judge in RLHF training loops for open-ended tasks, validate with held-out human evaluations before scaling.

Related stories

Each link below shares sources, entities, or timing with this story.

  1. Shared entities / Same source domain / Shared topic / What happened next

    The Bystander Effect in Multi-Agent Reasoning: More Agents Can Make LLMs Dumber

    Both cover LLM, Researchers; reported by the same outlet (arxiv.org); overlapping topics (effect, reasoning).

  2. Shared entities / Same source domain / What happened next / Downstream implication

    DPO secretly optimizes a preference graph, not just pairwise comparisons.

    Both cover Researchers, RLHF; reported by the same outlet (arxiv.org); picks up the Researchers thread on 2026-05-11.

  3. Shared entities / Same source domain / What happened next / Tension

    Your AI-Generated Code Compiles. It Doesn't Work.

    Both cover LLM, Researchers; reported by the same outlet (arxiv.org); picks up the LLM thread on 2026-04-22.

  4. Shared entity: LLM / Same source domain / Shared topic / What happened next / Downstream implication

    Human and LLM everyday reasoning are both pattern matching, not world models.

    Both cover LLM; reported by the same outlet (arxiv.org); overlapping topics (human, reasoning).

  5. Shared entities / Same source domain / What happened next

    AI Coding Agents Don't Write Clean Patches. They Can't.

    Both cover LLM, Researchers; reported by the same outlet (arxiv.org); picks up the LLM thread on 2026-05-24.

  6. Shared entity: Researchers / Same source domain / Shared topic / What happened next

    Frontier LLMs align with human brain patterns during game learning.

    Both cover Researchers; reported by the same outlet (arxiv.org); overlapping topics (human, reasoning).

  7. Shared entity: LLM / Same source domain / Shared topic / Earlier coverage

    Bias-Bounded LLM Judges.

    Both cover LLM; reported by the same outlet (arxiv.org); overlapping topics (evaluation, judg).

  8. Shared entity: RLHF / Same source domain / What happened next / Tension

    One Transformer Layer May Rival Full RL Tuning

    Both cover RLHF; reported by the same outlet (arxiv.org); picks up the RLHF thread on 2026-07-02.

Source trail

Entities

Provenance

Canonical issue
Ramsay Research Agent
AI generated
no
Story unit
2026-03-14-reasoning-llm-judges-goodhart-s-law-strikes-again
Labels
source-backed, canonical briefing excerpt