REVIEW 13 cited by
AxBench: Steering LLMs? Even Simple Baselines Outperform Sparse Autoencoders
Not yet reviewed by Pith; the record is open.
This paper has not been read by Pith yet. Machine review is queued; the pith claim, tier, and objections will appear here once it completes.
SPECIMEN: schema-true, not a live event
T0 review · schema-true
One-sentence machine reading of the paper's core claim.
pith:XXXXXXXX · record.json · timestamp
read the original abstract
Fine-grained steering of language model outputs is essential for safety and reliability. Prompting and finetuning are widely used to achieve these goals, but interpretability researchers have proposed a variety of representation-based techniques as well, including sparse autoencoders (SAEs), linear artificial tomography, supervised steering vectors, linear probes, and representation finetuning. At present, there is no benchmark for making direct comparisons between these proposals. Therefore, we introduce AxBench, a large-scale benchmark for steering and concept detection, and report experiments on Gemma-2-2B and 9B. For steering, we find that prompting outperforms all existing methods, followed by finetuning. For concept detection, representation-based methods such as difference-in-means, perform the best. On both evaluations, SAEs are not competitive. We introduce a novel weakly-supervised representational method (Rank-1 Representation Finetuning; ReFT-r1), which is competitive on both tasks while providing the interpretability advantages that prompting lacks. Along with AxBench, we train and publicly release SAE-scale feature dictionaries for ReFT-r1 and DiffMean.
Forward citations
Cited by 13 Pith papers
-
ODRA: Synthesizing Cognitive Behavioral Therapy Sessions with Structured Chain-Of-Thought and Dynamic Patient Resistance
ODRA synthesizes CBT sessions with a chain-of-thought CBT aligner and a dynamic patient resistance orchestrator, but its clinical robustness claims rest on evaluations using its own simulated patients.
-
Retrieval is Enough: Training-Free Interpretability with a Tool-Using Agent
HARP shows that retrieval from an activation database plus linear-probe tools, driven by an LLM agent, matches or exceeds SAEs and activation oracles on four interpretability tasks with zero training.
-
Temporal Preference Concepts and their Functions in a Large Language Model
Temporal preference in Qwen3-4B-Instruct-2507 localizes to layers 17–35 (especially L24 attention), has curved residual-stream geometry, is behaviorally unstable, and can be bidirectionally steered.
-
The SuperActivator Mechanism: Transformers Concentrate Reliable Concept Signals in the Tail
Reliable concept presence in transformers is concentrated in the extreme high-activation tail of in-concept tokens; thresholding that tail improves concept detection and localization.
-
VISOR++: Universal Visual Inputs based Steering for Large Vision Language Models
A single adversarially optimized image can reproduce activation-steering behavior in multiple VLMs and partially transfer to unseen models.
-
Sparse but Wrong: Incorrect L0 Leads to Incorrect Features in Sparse Autoencoders
Incorrect L0 makes sparse autoencoders mix correlated features rather than disentangling them, and a decoder projection metric can identify the correct L0.
-
Insights into a radiology-specialised multimodal large language model with sparse autoencoders
Applying Matryoshka sparse autoencoders to a radiology-specialised multimodal LLM reveals a minority of interpretable clinical features, while steering them produces unreliable and often off-target report changes.
-
Position: Use Sparse Autoencoders to Discover Unknowns
Sparse autoencoders are best used to discover unknown concepts, not to act on known concepts.
-
Resa: Transparent Reasoning Models via SAEs
SAE-Tuning, a sparse-autoencoder-guided SFT procedure, elicits RL-comparable reasoning in 1.5B models from CoT-free QA data at about $1 and 20 minutes of training.
-
Investigating Symbolic Triggers of Hallucination in Gemma Models Across HaluEval and TruthfulQA
Symbolic triggers like modifiers and named entities keep Gemma hallucination rates at 64-79% across model scales, suggesting larger models do not eliminate this failure mode.
-
Sealing The Backdoor: Unlearning Adversarial Text Triggers In Diffusion Models Using Knowledge Distillation
SKD-CAG erases adversarial text triggers from diffusion models by distilling the model's own clean outputs through cross-attention guidance, claiming 100% and 93% removal for pixel and style backdoors.
-
From Emergence to Control: Probing and Modulating Self-Reflection in Language Models
Self-reflection in LLMs can be steered up or down by a single activation-space vector, improving accuracy when amplified and cutting output length when suppressed.
-
Probabilistic Concept-Aware Steering for Trustworthy LLM Inference
PCS improves steering direction accuracy by adaptively sampling the intervention coefficient from a cosine-similarity-conditioned Gaussian, but its evaluation is partly circular because the optimal coefficient is chos...
Discussion (0). Sign in to comment.