{"slug": "reward-informed-sparse-autoencoders-and-the-solution-completeness-confound", "title": "Reward-Informed Sparse Autoencoders and the Solution-Completeness Confound", "summary": "A new study from arXiv (2608.26136v1) finds that reward-informed sparse autoencoders (RI-SAEs) trained on Llama-3.1-8B separate high- and low-reward reasoning traces mainly by solution completeness, not reasoning quality: a TF-IDF text classifier splits the classes with AUC 0.75–0.83, and three structural cues (length, closed reasoning block, boxed answer) reach AUC 0.70, with 99% of good versus 69% of bad completions boxed. The authors present the RI-SAE recipe and a control battery, noting that a reward-blind SAE shows no class separation (silhouette 0.01), while the reward-informed SAE's selected features reach silhouette 0.79, and only two discriminative features are readable (symbolic mathematics; procedural and evaluative language).", "body_md": "arXiv:2608.26136v1 Announce Type: new\nAbstract: Sparse autoencoders (SAEs) decompose language-model activations into sparse, interpretable features, and an appealing way to aim them at reasoning is to curate their data with a signal reinforcement learning already produces: the reward. We build such a reward-informed SAE (RI-SAE): we split GRPO trajectories into high-reward (\"good\") and low-reward (\"bad\") reasoning continuations, train a standard JumpReLU SAE on their activations, and then ask what the resulting good/bad separation actually measures. On Llama-3.1-8B a sparse subset of the 16,384 features does separate the classes (silhouette 0.79 on the selected features versus 0.005 for the full code), but a control battery shows the separation is largely solution completeness rather than reasoning quality: a TF-IDF text classifier already splits the classes (AUC 0.75--0.83), and three structural cues alone (length, a closed reasoning block, and a boxed answer) reach AUC 0.70 (99% of good versus 69% of bad completions are boxed). A generic SAE that never saw the reward does not separate the classes at all (silhouette 0.01, no discriminative features), so the 0.79 is in-sample fitting of this curated signal rather than structure that a reward-blind dictionary recovers. We therefore present the recipe and its control battery together: reward filtering is a cheap, label-free way to reuse RL signals for interpretability, but most of what it surfaces is completion form. Two discriminative features are still readable (symbolic mathematics; procedural and evaluative language), which we take as illustrative rather than as isolated reasoning.", "url": "https://wpnews.pro/news/reward-informed-sparse-autoencoders-and-the-solution-completeness-confound", "canonical_source": "https://arxiv.org/abs/2608.26136", "published_at": "2026-08-28 04:00:00+00:00", "updated_at": "2026-08-28 04:20:28.434143+00:00", "lang": "en", "topics": ["artificial-intelligence", "machine-learning", "ai-research"], "entities": ["arXiv", "Llama-3.1-8B", "JumpReLU SAE", "GRPO"], "alternates": {"html": "https://wpnews.pro/news/reward-informed-sparse-autoencoders-and-the-solution-completeness-confound", "markdown": "https://wpnews.pro/news/reward-informed-sparse-autoencoders-and-the-solution-completeness-confound.md", "text": "https://wpnews.pro/news/reward-informed-sparse-autoencoders-and-the-solution-completeness-confound.txt", "jsonld": "https://wpnews.pro/news/reward-informed-sparse-autoencoders-and-the-solution-completeness-confound.jsonld"}}