SAEs as Stethoscopes: Interpretability-Guided Layer Selection for Task Vector Model Editing
This paper evaluates a Sparse Autoencoder (SAE)-guided model editing pipeline for mathematical reasoning on Gemma-3-4B-IT, finding that projecting task vectors onto SAE feature subspaces discards ~97% of modification energy due to geometric misalignment between activation-space SAE directions and weight-space task vectors. The authors reframe SAEs as diagnostic tools ('stethoscopes') rather than intervention filters ('scalpels'), using SAE-derived specificity scores to identify which layers to inject unfiltered task vectors into. This approach improves Number Theory accuracy from 29.6% to 39.4% on Minerva Math (p=0.0007), with 5 of 7 math subjects significantly improved and none degraded. The method is fully deterministic and adds no inference cost.
Related guides (3)
Related events (8)
SAERL: Using Sparse Autoencoders to Guide LLM Reinforcement Learning Data Engineering
SAERL is a post-training data engineering framework that uses Sparse Autoencoders (SAEs) — a mechanistic interpretability tool — to extract intrinsic model signals for controlling data diversity, difficulty, and quality during RL fine-tuning. The framework applies SAE-space clustering for batch diversity, a difficulty proxy for curriculum ordering, and a quality probe for data filtering. On Qwen2.5-Math-1.5B with GRPO, SAERL achieves 3% average accuracy improvement and reaches target accuracy with 20% fewer training steps. SAE representations transfer across model families and scales, suggesting broad applicability as a lightweight data engineering tool.
Study finds SAE unstable features reflect reproducible subspaces, not pure noise
A new arXiv paper investigates feature stability in sparse autoencoders (SAEs), measuring the probability that individual learned features reappear across independent training runs. The authors find a functional asymmetry: stable features carry most reconstruction-relevant signal, while unstable features are individually non-reproducible but concentrate in reproducible lower-rank subspaces, suggesting seed dependence reflects basis ambiguity rather than noise. A synthetic model confirms that low-rank ground-truth features can be recovered at the subspace level even when individual SAE latents are non-identifiable across seeds. The work has direct implications for interpretability research that relies on SAE features as meaningful, stable units of analysis.
Sparse AutoEncoder steering reduces Whisper hallucination rate by ~5x without fine-tuning
Researchers investigate hallucination detection and mitigation in OpenAI's Whisper ASR model by probing internal encoder representations. They find that both raw activations and Sparse AutoEncoder (SAE) latents encode linearly separable hallucination signals concentrated in deeper layers. SAE-based activation steering reduces hallucination rates from 72.6% to 14.1% (Whisper small) and 86.9% to 27.3% (Whisper large-v3) on non-speech audio, with minimal WER degradation, approaching fine-tuning-level performance without weight updates.
SPEX and ProxySPEX: Scalable Interaction Discovery for LLM Interpretability
Researchers from BAIR introduce SPEX (Spectral Explainer) and ProxySPEX, algorithms for identifying influential feature, data, and model-component interactions in LLMs at scale. The approach exploits sparsity, low-degreeness, and hierarchy properties to reframe interaction discovery as a sparse recovery problem using tools from signal processing and coding theory. ProxySPEX achieves comparable performance to SPEX with roughly 10x fewer ablations by leveraging hierarchical structure. The methods are evaluated on feature attribution (sentiment analysis), data attribution, and mechanistic interpretability tasks, outperforming marginal methods like LIME at long context lengths.
MAST: Mechanism-guided selective unlearning for RLVR-trained reasoning models
Researchers introduce MAST (Mechanism-Aligned Selective Targeting), a method for selectively unlearning capabilities induced by reinforcement learning from verifiable rewards (RLVR) in language models while minimizing collateral damage to retained knowledge. The approach ranks attention-projection tensors by off-principal energy and gradient coupling to identify a targeted subset for update, rather than applying full-parameter gradient ascent. Evaluated on Qwen2.5-Math-1.5B and Qwen3-1.7B-Base, MAST achieves statistically significant forgetting on target MATH problems while preserving GSM8K performance, whereas full-parameter unlearning collapses retained capabilities. The method generalizes across seeds and unlearning objectives (NPO/SimNPO).
SV-Detect: AI-generated text detection via steering vectors in representation space
SV-Detect proposes a method for detecting machine-generated text by extracting steering vectors from the hidden representations of a frozen language model, constructing layer-wise directions that separate human from AI-written text. A lightweight classifier trained on projection features achieves strong performance both in-distribution and under distribution shift across domains, source models, and editing attacks like polishing and rewriting. The approach reframes AI-text detection as a representation-space probing problem, with interpretation analyses showing the learned directions capture stylistic cues beyond surface features.
VaSE: Value-Aware Stochastic KV Cache Eviction improves reasoning model efficiency
A new arXiv preprint introduces Value-aware Stochastic KV Cache Eviction (VaSE), a training-free method for compressing KV caches in long-chain-of-thought reasoning models. The authors identify two key failure modes in prior eviction approaches — catastrophic repetition loops caused by evicting high-magnitude value states, and low cache diversity — and address both with targeted protections and stochastic eviction. On six reasoning tasks with Qwen3 models at 4x compression, VaSE outperforms the current best selection-based sparse attention method and exceeds the strongest eviction baseline by over 4%, while supporting FlashAttention2 and maintaining a static memory footprint.
ETCHR: Decoupled Image Editing for Visual Chain-of-Thought Reasoning in MLLMs
ETCHR introduces a question-conditioned, reasoning-aware image editing model that decouples visual transformation from downstream understanding in multimodal LLMs. It addresses two identified gaps—language-side (mapping abstract questions to visual edits) and generation-side (edit quality degrading with reasoning depth)—via a two-stage training recipe combining supervised fine-tuning on edit trajectories and VLM-derived reward signals. Because the editor is decoupled, it plugs into arbitrary MLLMs without retraining, yielding Pass@1 gains of roughly +4.6 to +5.5 points across five task families when paired with Qwen3-VL-8B, Gemini-3.1-Flash-Lite, and Kimi K2.5. The work advances the 'think with images' paradigm beyond fixed toolkits and unified multimodal approaches.