Synthesizing counterfactual responses turned a pointwise LLM judge into a better pointwise judge, with the gains concentrated on the criteria the judge was weakest at
AspireAI community · arXiv
Agentic systems increasingly gate individual steps with a pointwise judge that scores a response against a criterion, and small language models are attractive for that role because they serve cheaply. The problem is that pointwise labels under-determine the decision boundary: labeled responses to different queries differ along many dimensions at once, so they do not reveal which features decide the label. Counterfactual Pairwise Augmentation (CPA) makes the deciding difference explicit by synthesizing and auditing a counterfactual response to the same query intended to carry the opposite label, then reframing the accepted pair as a pairwise task. The judge is fine-tuned jointly on the pointwise and pairwise data but deployed in pointwise mode only. A second-order analysis shows when this beats merging separately trained pointwise and pairwise experts: merging moves every parameter direction toward the pairwise expert by the same fraction, whereas joint training can hold the pointwise optimum wherever the pairwise loss is flat and absorb its curvature where it is sharp. On separate Qwen3-8B LoRA adapters for six binary criteria, CPA improved the criterion-level accuracy of the small judges. The limitation is that the evaluation uses six binary criteria and one base model family, so the mechanism's advantage over expert merging has not been shown to scale with judge size or task breadth.
Read the story and explore more from the AspireAI community.
Two sparse-autoencoder analyses have resisted scaling: matching semantically similar features across layers, and compressing large feature circuits into interpretable supernodes. This paper argues both are the same underlying problem, estimating semantic distances between SAE features that sit on different activation manifolds. Its representation treats each feature not as a single decoder vector but as an activation-weighted distribution over the hidden states expressing it, projects those distributions into a shared reference space, and compares them with Wasserstein distance. The representation is proved invariant to activation rescaling and stable under perturbation, and recovers true matches under finite-sample margin conditions. Empirically it outperformed decoder-vector and LLM-based baselines and captured subtle functional distinctions between related features, and it compressed large feature circuits into interpretable supernodes automatically. The limitation is that this is interpretability tooling: the paper demonstrates better feature matching and circuit compression, not that downstream interpretability claims become more accurate.