Editing the model's previous answer in place beat giving it feedback in a separate turn
AspireAI community · arXiv
LLM-generated drafts often contain subtle factual or logical errors, and prior work shows models struggle to reliably integrate multi-turn feedback aimed at fixing them. In-place feedback is an interaction paradigm where the user directly edits the model's previous response and the model continues generation from the edited context. The reported result is that in-place feedback consistently outperforms standard multi-turn feedback across five reasoning-intensive benchmarks while requiring fewer tokens, with a fine-grained analysis showing it applies corrections more reliably and propagates them into subsequent reasoning. A user study with domain experts refining LLM-generated summaries corroborates the effect: participants reported higher final-output satisfaction and substantially lower fatigue with in-place feedback, and a mixed strategy combining in-place and multi-turn feedback improved every dimension measured. The paper was recognised with an Outstanding Paper award at EMNLP 2026. The limitation worth naming is the human in the loop: this paradigm assumes someone edits the previous turn directly, which changes the interface and the labour cost, so it is not a drop-in replacement for plain multi-turn prompting in unattended workflows.
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.