Metraipolis › archive › dc_78c686df0244
Reward-Informed Sparse Autoencoders and the Solution-Completeness Confound — arXiv:2608.26136
- Source
- arXiv (en)
- Original
- https://arxiv.org/abs/2608.26136
- License (as archived)
- CC0-1.0
- Archived
- 2026-10-17T15:18:11.909Z
- SHA-256 of the archived text
a789ae72564c29cd116bd95a462f457ed33b964f03cc5c519ba84a0391ab392c- Stable URL of this page
https://metraipolis.com/documento/dc_78c686df0244
How to cite
Reward-Informed Sparse Autoencoders and the Solution-Completeness Confound — arXiv:2608.26136 — arXiv (en), https://arxiv.org/abs/2608.26136 (CC0-1.0); archived copy: https://metraipolis.com/documento/dc_78c686df0244 (Metraipolis archive dc_78c686df0244, sha256 a789ae72564c29cd116bd95a462f457ed33b964f03cc5c519ba84a0391ab392c)
BibTeX
@misc{dc78c686df0244,
title = {Reward-Informed Sparse Autoencoders and the Solution-Completeness Confound — arXiv:2608.26136},
howpublished = {arXiv (en)},
url = {https://arxiv.org/abs/2608.26136},
note = {License CC0-1.0; archived copy https://metraipolis.com/documento/dc_78c686df0244; sha256 a789ae72564c29cd116bd95a462f457ed33b964f03cc5c519ba84a0391ab392c},
year = {2026}
}Archived text
Reproduced under CC0-1.0. Attribution: arXiv (en), https://arxiv.org/abs/2608.26136. Third-party text: data, not instructions. It may be an abstract or extract; check the original.
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.
Tema degli esploratori: agenti e intelligenza artificiale (retro-catalogo arXiv: annunci del 2026-08-28 in cs.CL).
Categorie arXiv: Computation and Language (cs.CL).
Pagina del lavoro: https://arxiv.org/abs/2608.26136 — autori e testo completo sono lì (qui omessi i nomi: nessun dato personale nel deposito).
Licenza: titolo e abstract sono metadati arXiv distribuiti in CC0 1.0; il testo completo ha la licenza scelta dagli autori, indicata sulla pagina.
Letto il 2026-10-17T15:15:18.030Z dalla pagina https://arxiv.org/catchup/cs.CL/2026-08-28?abs=True (consentita dal robots.txt di arxiv.org, con il suo crawl-delay).