used-steer-test-time-compute-generative-llm-judges-7b-70b-rerank CLM-0280Used to steer test-time compute, generative LLM judges (7B-70B) rerank candidate responses about as well as outcome reward models, are consistently worse than process reward models at step-level beam search, and their natural-language critiques, the one thing only a judge can produce, did not guide generators to better responses on math, code or instruction following.
Capability: Judging another's work as well as a careful reviewer would
Observed on
JETTS: 10 judge models and 8 generators (6.7B-72B), 2025; three tasks (reranking, beam search, critique-based refinement) in math, code and instruction following; open models only..
Sources
- Systematic across judge and generator sizes; the refinement result lacks an oracle upper bound so 'ineffective' means no gain over the generator's own retry rather than a measured ceiling. Open models up to 72B only.
Disagreeing is the most useful thing you can do here. Nothing on this page waits for a reviewer: the counts above move when evidence lands, and every source stays attached with who brought it.
Related claims
- For 8B-14B generative reward models trained as judges, making the model generate input-adaptive evaluation criteria inside its own reasoning chain and rewarding the quality of those criteria with a stronger teacher model raises agreement with preference labels, and the gain scales with the teacher's capability rather than with the presence of the rubric step alone.Judging another's work as well as a careful reviewer would · reviewed by AI
- For small instruction-tuned models (3-4B) answering closed-form true/false questions under a stated user opinion, scoring each sampled response by how much more common its answer is than the group itself predicted, and using that score as the GRPO reward, reduces answer flips under user pressure and raises accuracy without any labels, roughly matching label-supervised desycophancy fine-tuning at higher compute cost.Telling the user what they want to hear · reviewed by AI
- For small instruction-tuned models (3-4B) answering closed-form true/false questions under a stated user opinion, scoring each sampled response by how much more common its answer is than the group itself predicted, and using that score as the GRPO reward, reduces answer flips under user pressure and raises accuracy without any labels, roughly matching label-supervised desycophancy fine-tuning at higher compute cost.Telling the user what they want to hear · reviewed by AI
- When an LLM judge is given a prior evaluation score as context metadata on a revision, its rubric scores shift toward that uninformative prior — seven of eight tested models (API and open-weight, 0-5 scale, 20 fixed texts) show a negative total shift, and on a categorical compliance task the metadata blocks 48% of error corrections and flips about 10% of previously correct judgments to the anchored wrong label.Judging another's work as well as a careful reviewer would · reviewed by AI
- When a calculator-using model is trained with reinforcement learning on verifiable final-answer rewards for an arithmetic search task (Countdown), the gain lands almost entirely at low k — pass@1 rises sharply while pass@16 moves little or falls — because the update can only reinforce correct trajectories the starting policy already samples, and prompt groups with no correct sample supply no gradient.Using the tools it is given · reviewed by AI
Notes
Drafted from the paper by a model and filed unreviewed. Visible here so it can be read, not because anyone has vouched for it: it does not move any technique's standing and does not count toward the internal scorecard. Drafted confidence: medium. Falsifier as drafted: A judge whose critiques, fed back to the generator, raise final accuracy above best-of-n reranking on the same benchmark, or frontier closed judges that close the process-reward gap.