Across 57 long Claude Code and Codex sessions, models drifted away from their stated coding persona early, and in-session compaction neither reliably reset nor amplified it
AspireAI community · arXiv
Frontier models are trained to hold stable programming-assistant personas, but those personas drift over long agentic coding sessions in production, where existing persona-stability work covers only scripted chat of tens of turns rather than thousands of tool-using steps with repeated compaction. ContextEcho provides a 25-probe identity suite, a snapshot-then-probe protocol that evaluates conversation state without perturbing the session, judge-based and judge-free metrics, and a corpus of 57 long agentic-coding sessions collected with consent and redacted for release across Claude Code and Codex CLI, used for a 23-model cross-provider study and for corpus-scale trajectory analysis. Persona drift showed up across providers and emerged early in real agentic interaction, while compaction neither reliably reset nor amplified it, and a single-shot anchor did restore the original persona, though placebo controls suggest that is partly a generic register-reset effect. The sharpest finding is a trade-off: drifted prefixes improved tool-call fidelity on same-task continuations, which the authors attribute to retained task context, but in tool-free chat the same prefixes broke output-format contracts and inflated output length. The limitation is the corpus size and its confinement to two coding agents.
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.