aiqbench analysis
This research introduces DSPA, an inference-time technique that leverages Sparse Autoencoders (SAEs) to steer language model outputs toward preferred behaviors without updating base weights. By operating only on active token latents during decoding, the method significantly reduces alignment-stage compute while improving performance on benchmarks like MT-Bench. The study also provides mechanistic insights, revealing that preference adjustments primarily influence discourse and stylistic signals rather than core factual knowledge.
DSPA: Dynamic SAE Steering for Data-Efficient Preference Alignment
arXiv:2603.21461v2 Announce Type: replace Abstract: Preference alignment is usually achieved by weight-updating training on preference data, which adds substantial alignment-stage compute and provides limited mechanistic visibility. We propose Dynamic SAE Steering for Preference Alignment (DSPA), an inference-time method that makes sparse autoencoder (SAE) steering prompt-conditional. From preference triples, DSPA computes a conditional-difference map linking prompt features to generation-control features; during decoding, it modifies only token-active latents, without base-model weight updates. Across Gemma-2-2B/9B and Qwen3-8B, DSPA improves MT-Bench and is competitive on AlpacaEval while preserving multiple-choice accuracy. Under restricted preference data, DSPA remains robust and can rival the two-stage RAHF-SCIT pipeline while requiring up to $4.47\times$ fewer alignment-stage FLOPs. Finally, we audit the SAE features DSPA modifies, finding that preference directions are dominated by discourse and stylistic signals, and provide theory clarifying the conditional-difference map estimate and when top-$k$ ablation is principled.
Abstract:Preference alignment is usually achieved by weight-updating training on preference data, which adds substantial alignment-stage compute and provides limited mechanistic visibility. We propose Dynamic SAE Steering for Preference Alignment (DSPA), an inference-time method that makes sparse autoencoder (SAE) steering prompt-conditional. From preference triples, DSPA computes a conditional-difference map linking prompt features to generation-control features; during decoding, it modifies only token-active latents, without base-model weight updates. Across Gemma-2-2B/9B and Qwen3-8B, DSPA improves MT-Bench and is competitive on AlpacaEval while preserving multiple-choice accuracy. Under restricted preference data, DSPA remains robust and can rival the two-stage RAHF-SCIT pipeline while requiring up to $4.47\times$ fewer alignment-stage FLOPs. Finally, we audit the SAE features DSPA modifies, finding that preference directions are dominated by discourse and stylistic signals, and provide theory clarifying the conditional-difference map estimate and when top-$k$ ablation is principled.
Machine Learning (cs.LG); Artificial Intelligence (cs.AI); Computation and Language (cs.CL)
arXiv:2603.21461 [cs.LG]
(or arXiv:2603.21461v2 [cs.LG] for this version)
https://doi.org/10.48550/arXiv.2603.21461 Focus to learn more arXiv-issued DOI via DataCite
arXivLabs is a framework that allows collaborators to develop and share new arXiv features directly on our website.
Both individuals and organizations that work with arXivLabs have embraced and accepted our values of openness, community, excellence, and user data privacy. arXiv is committed to these values and only works with partners that adhere to them.
Have an idea for a project that will add value for arXiv's community? Learn more about arXivLabs.
Full text under the publisher's open licence · 235 words
Related coverage
[AINews] Reflection Beam - 501B-A23B American Open Model
Latent Space · Oct 6, 2026
[AINews] not much happened today
Latent Space · Oct 3, 2026
Angles of attack for continual learning safety
LessWrong · Jun 16, 2026
Dreams of alignment in a world without politics
LessWrong · Sep 24, 2026