Can LLMs Judge Better Than They Generate? Evaluating Task Asymmetry, Mechanistic Interpretability and Transferability for In-Context QA
Sambaran Bandyopadhyay
Why It Matters
What makes this one worth your time
Understanding the limitations of self-evaluation in LLMs is crucial for improving automated evaluation pipelines and ensuring reliable AI-generated content.
The paper challenges the assumption that LLMs can evaluate their own outputs better than they generate them.
Summary
The paper investigates whether large language models (LLMs) are better at evaluating their own generated answers than generating them in a controlled in-context question-answering setting. It finds that generation accuracy often surpasses self-evaluation accuracy, except in multi-hop scenarios. The study uses attention analysis and LoRA fine-tuning to explore the asymmetry between generation and evaluation tasks, challenging assumptions in self-evaluation pipelines.
Key contributions
- Demonstrates that generation accuracy can exceed self-evaluation accuracy in LLMs.
- Provides attention analysis to explain task asymmetry in LLMs.
- Uses LoRA fine-tuning to confirm that asymmetry is not a training artifact.
Notable insights
- Evaluation by LLMs attends to context significantly less than generation does.
- LoRA fine-tuning shows that generation and evaluation tasks have distinct impacts on model behavior.
Possible limitations
- Not stated in the abstract
Abstract
arXiv:2606.28050v1 Announce Type: cross Abstract: LLM-as-a-Judge and self-evaluation pipelines implicitly assume that evaluation is easier than generation. We test this in a controlled in-context QA setting where a context passage is the sole information source and each model judges the answer it generated, removing the parametric-knowledge confound of open-domain comparisons. Across four benchmarks (SQuAD 2.0, DROP, HotpotQA, MuSiQue) and two models, evaluation is not uniformly easier: generation accuracy exceeds self-evaluation on three of four, with multi-hop MuSiQue the exception. Attention analysis reveals why: evaluation attends to context 3--5x less than generation does and barely reads the candidate answer. LoRA fine-tuning confirms the asymmetry is not a training artifact: generation fine-tuning induces over-acceptance and evaluation fine-tuning degrades generation. These findings challenge core assumptions in self-evaluation pipelines.