Agents that turn their own mistakes into LoRA adapters improved 8.9 to 17.0 points, with only about 0.02 percent extra trainable parameters
AspireAI community · arXiv (Yuval Kansal, Karen Ullrich et al., Princeton University / Meta FAIR)
Today's agents recover from a mistake only for the rest of that session. An in-context reflection gets forgotten across sessions and competes for context; full fine-tuning is expensive and can damage what the model already knows. Self-Direction sits between the two: when an agent fails a task, it reflects on its own trajectory, extracts a lesson, and distills that lesson into a lightweight low-rank adapter. Base weights stay frozen, and each adapter trains parameters equivalent to roughly 0.02 percent of the base model's parameter count.
Adapters go into a persistent library of lessons that is automatically routed to related problems later. Across open-weight Qwen-family agents on GAIA, Claw-Eval and Terminal-Bench 2.1, overall success rates improved by 8.9 to 17.0 percentage points, competitive with frontier systems. The mechanism is what makes it interesting: the method nearly doubles the reliability of an agent's corrections on repeat encounters of the same task, and a handful of adapters generalize to dozens of new tasks with no additional training.
Caveat: the evaluations are open-weight Qwen agents, not frontier models, and the adapters are task-shaped, so the benefit depends on the lesson library being routed to genuinely similar problems.
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.