Overview

Activation steering studies how edits to a model’s hidden activations can control its behavior at inference time, enabling changes without retraining the full model. Common approaches add or remove contrastive steering vectors, target sparse autoencoder features, neurons or attention heads, and use conditional or feedback-based controllers to adapt interventions to the input or generation state. Researchers apply these methods to language tasks such as reasoning, refusal, factuality and hallucination mitigation, as well as image and video generation, speech, and embodied agents; evaluations include benchmarks such as GSM8K, TruthfulQA, POPE and LIBERO. A central challenge is that readable or decodable features are not necessarily controllable, and effective edits can vary by context and architecture or cause collateral changes to fluency, utility and safety.

Latest papers

417 papers59 papers in the last four weeks, down 9% on the four weeks before. 0.4% of all new papers.

Oct 31, 2026

Geometry of Confabulation and Refusal in Small Language Models

Faced with an unanswerable question, small language models can either refuse or confabulate: confidently produce a wrong answer. Can the hidden state before the first generated token provide a clue? Can we flip confabulations into refusals? I introduce ConfabQA, a 784-item benchmark whose obscure-pre-cutoff category breaks the typical correctness/knowledge-cutoff confound and map the geometry of the hidden state with linear probes, a sparse autoencoder (SAE), and direct optimization of intervention directions. On Qwen3-1.7B and Gemma 2 2B the probe separates upcoming refusals from confabulations well beyond prompt features (+7.4/+16.3 pp); on Llama 3.2 3B the ConfabQA margin is within fold-to-fold spread, but a direct refusal probe on PopQA and TriviaQA recovers +18-25 pp. On Qwen I closely inspect three candidate directions for the refusal vector: probe-recovered, Qwen-Scope SAE feature, and directly optimized. Although the three directions are pairwise 50-81 degrees apart, a push at or below the state's own norm along any of them flips most confabulations into refusal openings dominated by the "As..." family. Forcing the opener without touching the hidden state yields the same one-in-three judged-refusal rate as any push. The opposite push collapses refusals mostly into confabulations, yet leaves originally correct answers intact, so the direction gates the opening speech act rather than the knowledge behind it. Confabulations do not resolve into a concentrated feature in a dictionary or a token signature on Qwen3-1.7B. Scoring every probe on the margin over the strongest of four prompt-feature baselines extracts the hidden state's own contribution. Correctness signal is model-dependent and portable only for Llama 3.2 3B: 81% of the probe's margin over test majority survives transfer between ConfabQA, PopQA, and TriviaQA. Benchmark, judge, and code are released.
Oct 31, 2026

Editing the Goal: Per-Token SAE Steering and Its Depth Limits in π0.5π_{0.5}

Sparse autoencoders (SAEs) have recovered interpretable, causally relevant features in robot foundation models (RFMs) such as π0.5π_{0.5} (arXiv:2603.19183, arXiv:2603.19233). SAE features on individual instruction tokens are less interpretable than timestep-pooled ones (arXiv:2603.19183), and whether they can redirect a robot's goal remains untested. We show that replacing the SAE features of a single goal word with those of an alternative goal word, leaving the instruction text unchanged, steers π0.5π_{0.5}-LIBERO into completing the alternative task. At layer 0, BatchTopK feature swaps redirect behavior in up to 20/20 rollouts in a scene where random-feature swaps redirect none, with success varying across six scenes in two LIBERO suites and across layers. The effect depends on SAE architecture: word-selected Matryoshka edits largely do not outperform random ones. SAEs are also not required. Adding a fixed difference-in-means vector at twice its natural magnitude redirects 8-10/10 rollouts. Layer-wise activation patching shows that goal-word edits remain effective up to layer 6 for object location and layer 2 for object identity, and up to layer 13 in both cases when patching also covers camera tokens. These results are proof of existence in an RFM in simulation and an initial map of how deep goal-redirecting edits remain effective. Visit https://doubleblind-sae-pi05.github.io for video rollouts.
Oct 30, 2026

LLM Detoxification via Transition-Aware Reward-Guided Representation Sampling

Large Language Models (LLMs) can generate toxic content, motivating effective test-time detoxification. Existing representation steering methods often rely on static or coarse interventions and provide imprecise guidance because they inadequately model transitions between toxic and non-toxic generations. We propose Transition-Aware Reward-Guided Representation Sampling (TARGRS), a test-time framework that learns reward guidance over toxicity transitions in latent representation space. During exploration, progressive steering along the toxicity direction constructs trajectories from non-toxic to toxic states. Toxicity scores along these trajectories turn sparse annotations into dense, listwise supervision for a transition-aware token-level reward model. During inference, TARGRS uses this model to guide coarse-to-fine sampling along the opposite, non-toxic direction: coarse sampling identifies the intervention needed to reach non-toxic regions, and refinement selects a detoxified representation with minimal perturbation. Experiments on 10 LLMs across RealToxicityPrompts, BOLD, and AttaQ show that TARGRS consistently outperforms state-of-the-art baselines, reducing toxicity by up to 69.38% on RealToxicityPrompts. TARGRS runs nearly as fast as the original LLM while preserving its core capabilities with minimal impact. By intervening in the language decoder space, TARGRS also extends to Large Vision-Language Models (LVLMs), reducing toxicity by up to 64.23% across 3 representative LVLMs.
Oct 29, 2026

Steering Vectors as a Training Signal in LLM Post-Training

Steering vectors (SVs) are residual-stream directions that shift the behavior of LLMs when added to hidden states at inference time. While prior work has mainly used SVs as post-hoc controls, we ask whether the same directions can also guide parameter updates during post-training. We study this question across supervised learning, including SFT and distillation, and reinforcement learning with verifiable rewards (RLVR). Our framework considers two intervention regimes: a hidden-state hook during teacher-forced supervised training, and an advantage-repair mechanism on wrong rollouts in GRPO. For each regime, we analyze three design axes, which are the steering direction, its placement, and the operation used to combine it with training. In supervised settings, SV injection improves over plain SFT and distillation, with gains scaling with the off-policy gap and reaching up to a 12%12\% improvement when distilling from thinking-mode teachers. These gains hold across three random seeds. In RL settings, applying a correctness-contrast direction as advantage repair on wrong rollouts, including otherwise zero-gradient all-wrong groups, recovers a learning signal that transfers to held-out math benchmarks and acts as a composable auxiliary signal that adds to existing zero-variance remedies such as entropy-aware advantage shaping. We further identify three failure boundaries. In detail, code generation can break the direction, cross-family layer transfer can break the placement, and inference-time application on an SV-trained checkpoint can break the intervention regime. Together, these results establish steering vectors as a conditional but effective training signal for LLM post-training. Our code is available here.
Oct 29, 2026

One Vector or Many? Clustered Optimal-Transport Steering under Quantization

Activation steering changes what a language model generates by editing its hidden states at inference time. Its most common form adds one vector, the difference of class means (DiM), to every hidden state. CHaRS (concept heterogeneity-aware representation steering) uses optimal transport instead: an entropic plan pairs the clusters of the two classes, and each hidden state moves along a mixture of the cluster displacements, weighted toward its nearest clusters. We study this family of clustered optimal-transport maps for toxicity mitigation in Gemma-2-2B and Granite-3.1-2B, and ask what the plan adds to DiM and whether that survives INT8 and NF4 quantization. We prove an identity: whatever the plan, the displacements averaged with the cluster sizes as weights equal DiM, so a clustered map can differ from DiM only for tokens whose routing weights depart from the cluster sizes. In the CHaRS protocol they stay close, and CHaRS is statistically equivalent to Mean-AcT, which adds DiM at every layer, in all eight pre-specified equivalence tests. A variant fitted on individual tokens, which routes each token mainly to its nearest clusters, leaves 37-39% less toxicity than DiM on Gemma and 51-61% less on Granite at steering strength 3 (toxic-bert), where DiM already removes 76-85% of it. Two other toxicity classifiers confirm the gain on Granite, but on Gemma only one does. At equal strength the variant has somewhat higher perplexity, but at matched toxicity it is slightly more fluent than DiM. Yet a nearly constant map from the same clusters also leaves about 30% less toxicity, so the role of treating tokens differently remains open. Under INT8 and NF4 the variant still leaves less toxicity than DiM, and transferring it from full precision to the NF4 model instead of refitting it adds no significant toxicity after correction for multiple tests.
Oct 28, 2026

Hidden Commitment: When Language Models Silently Pick a Side and How Steering Can Surface It

Language-model answers are often treated as a visible record of the internal decision that produced them. We test this assumption in conflicting-context question answering, where two supplied documents support incompatible answers. A model can process the disagreement and still return a one-sided answer, even under prompts that ask it to disclose conflict. In Qwen 0.8B, hidden source commitment is captured by a residual source-choice direction whose sign tracks which document the answer uses. Sparse middle-layer MLP writers feed the same coordinate, and steering the coordinate changes the chosen source. A transfer run on Gemma 3 1B preserves the source-choice readout and sparse-write bridge under LLM-audited labels. Conflict disclosure follows a different pattern, appearing weaker, later-layer, and only partially steerable. These results support a practical monitoring strategy for retrieval systems that detects silent source commitment with residual readouts and pairs conflict-surfacing prompts with disclosure steering when the model has selected a side under disagreement.
Oct 27, 2026

Human-Residual Steering: Aligning Frozen LLM Judges with Human Judgment Distributions

Large language models (LLMs) are widely used as automated judges, yet most alignment methods target aggregated human judgments and overlook variation among individual annotators. Existing approaches to matching the full distribution of human ratings mainly rely on output-level calibration or parameter fine-tuning. We propose Human-Residual Steering (HRS), to our knowledge the first method to align frozen LLM judges with human judgment distributions through learned hidden-state interventions. An oracle analysis shows that low-rank corrections at a single layer remove most of the discrepancy between judge and human distributions. HRS learns to predict such corrections from the judge's hidden representation and rating logits, training on individual annotations with proper scoring rules for ordinal and nominal judgments and optimising through the frozen downstream layers. Across 14 benchmarks and four judges from three model families, HRS outperforms all evaluated frozen-weight baselines. With Qwen2.5-7B, it achieves the lowest Jensen-Shannon distance on 13 of 14 benchmarks, reducing the mean from 0.565 to 0.352 versus 0.377 for the strongest calibrator. HRS also raises mean Spearman correlation with aggregate human scores from 0.449 to 0.579, improves population-level estimation under limited annotation budgets, and trains about 88 times fewer parameters than distributional fine-tuning at a modest cost in alignment.
Oct 27, 2026

Orthogonal and Parametrized Multi-Attribute Steering of LLMs

Activation steering offers an effective and efficient way to modify an LLM's behavior at inference time. However, it remains unclear how to apply steering such that the signal is consistently strong and the steering strength has a universally interpretable meaning. We address this first by identifying the most steerable interval for each steering vector and rescaling the steering range accordingly. Building on this, we introduce a more interpretable method for steering, called parameterized steering which instead of pushing activations in a direction, it sets them at the desired trait strength. Next, with multi-trait steering, we find that naively combining vectors causes them to interfere with one another. In response, we develop a method to orthogonalize the vectors; it decouples these traits, allowing a model to exhibit one behavior without inducing the others, thus enabling multiple traits to be steered simultaneously.
Oct 23, 2026

Neyman Orthogonal Activation Steering

Activation steering is a popular approach for shifting LLM output toward a desired behavior, and increasingly for generating realistic synthetic data. The standard recipe estimates a steering direction by contrasting LLM activations on prompts with and without a target concept, such as truthfulness. In this paper, we separate the estimand (the steering direction these methods target) from the estimator used to recover it, two notions the literature often conflates. We frame nearly all existing methods as plug-in estimators of the average gradient of an outcome regression, with the binary contrastive label fixed as the outcome. Using this framing, we can immediately incorporate flexible modeling approaches and extend existing methods to allow for continuous outcomes such as log-probabilities or oracle ratings. Flexible ML models that fit the outcome well, however, might not estimate the gradient well, which motivates estimators that target the gradient directly. Drawing on the debiased machine learning literature, we develop a complementary estimator built on the Stein score, the Riesz representer of the gradient functional. Combining the two routes gives an orthogonalized estimator with classical double-robustness properties. We illustrate the framework on activation steering applied to a number of benchmark datasets.
Oct 23, 2026

REACT: Representational Early Analysis and Control of Chain-of-Thought Trajectories

Chain-of-thought (CoT) generation offers an opportunity to intervene before a reasoning model completes its answer. But does the point at which risk is easiest to detect also offer the most effective control? We study this question with REACT, a conditional activation intervention that learns a safety direction from paired safe and unsafe trajectories and applies it once to an early CoT hidden state. The model weights remain frozen, and all subsequent tokens are generated without further edits. To study timing, we train directions separately at each of the first five CoT positions and evaluate ungated edits with full post-edit regeneration. Across three distilled reasoning models and three generation seeds, the strongest gains concentrate in a two- to three-token early window; on DeepSeek-R1-Distill-Llama-8B, the effect falls to approximately zero at the fourth and fifth tokens, even though later states can be more predictive of risk. This separates risk readability from effective control: intervention timing must be evaluated through its effect on subsequent generation. On the same 8B model, REACT raises the rate of fully safe CoT-answer outputs on StrongREJECT-273 from 29.06% to 82.05%, while keeping benign refusal and usefulness near baseline. An anonymized implementation is available at https://anonymous.4open.science/r/REACT_official2-333B.
Oct 21, 2026

Probing and Steering for Safety-Relevant Behaviours in Large Language Models

Activation probes are computed directions in a language model's residual stream, used both to detect safety-relevant behaviours and, by adding their direction back into the stream, to steer them. We explore how the main design choices behind such probes interact, within one pipeline, on one model (Llama-3.1-8B-Instruct), and across three behaviours: deception, high-stakes awareness and sycophancy. The pipeline crosses two estimators (difference-of-means and logistic regression) with three token-pooling schemes (last, mean and learned attention), selects appropriate layers to detect or steer a behaviour, and evaluates every direction twice: as a detector of the behaviour, and as a steering vector to push outputs towards the behaviour. Prior work read together suggests a read/write dissociation, in which logistic regression reads (detects) best but difference-of-means directions write (steer) best. We find that it holds for deception and high-stakes but not for sycophancy, where the pattern reverses. The reversal appears to follow partially from the data rather than from the estimators. The best pooling technique and the layer choice depend on the behaviour and data, and the layer that steers best can sit well before the layer that detects best. The right steering metric to use between changing a model's perception and eliciting behaviour depends on the task - and the two can disagree outright. The exported probes and directions power an interactive interface through which a user can score and steer text directly.
Oct 21, 2026

WRDO: Weighted Refusal Direction Optimization in Aligned Language Models

Aligned large language models are trained to comply with benign requests while refusing harmful ones. However, recent work shows that this refusal behavior is fragile and can often be disrupted by simple activation- or weight-space interventions. In particular, refusal has been shown to be largely mediated by a single residual-stream direction, typically computed as a difference-of-means vector between harmful and harmless prompts. However, this standard construction aggregates heterogeneous harmful prompts into one uniform average, potentially conflating distinct refusal-related phenomena. In this work, we study whether refusal interventions can be strengthened by learning how to combine prompt-specific refusal directions. We propose Weighted Refusal Direction Optimization (WRDO), a procedure that constructs a composite refusal direction from a bank of per-prompt directions using learned aggregation weights. This approach allows different harmful prompts to contribute unequally to the final intervention direction. We evaluate WRDO against prior single-direction methods on three open-weight language models using three guard judges and HarmBench as an external evaluator. We further measure the effect of the intervention on general model capability through standard utility benchmarks. Finally, we investigate a defense-oriented setting in which the same optimization machinery is used in reverse to restore refusal behavior in an abliterated model checkpoint. Our results show that weighted aggregation provides a more effective and flexible mechanism for manipulating refusal representations.
Oct 21, 2026

Decodability Is Not Controllability: Local and Anisotropic Speaker Control in Continuous-Latent TTS

High linear decodability does not entail causal controllability in continuous-latent text-to-speech (TTS). In VoxCPM2, position-matched probes and equal-norm interventions across 30 speakers and 60 directed dyads expose this separation: reference-audio states at the final residual acoustic language model layer (RALM L7) retain 0.970 probe accuracy, yet aligned edits yield 0/60 target-dominant dyads. Edits at early reference-audio positions produce the opposite pattern. At L0, pair-aligned updates increase target similarity by 0.326 and achieve target dominance in all 60 dyads, against at most 1/60 under any equal-norm direction or position control. Independent speaker encoders, retrieval against held-out speaker prototypes, and double-blind listening confirm that the shifted voice is heard as the intended target. Cross-architecture interventions in dots.tts distinguish backbone leverage from direct acoustic control: backbone edits displace the source without target attraction, while x-vector ablation reduces source similarity from 0.849 to 0.006. These findings identify two requirements for training-free speaker control: geometric specificity and a computational route that carries the intended speaker condition into acoustic synthesis.
Oct 21, 2026

Uncensored Is Not Unbiased: Abliteration Removes Refusals but Not Script-Dependent Bias About Taiwan

"Abliterated" ("uncensored") builds of open-weight language models remove the activation direction that mediates refusal, and are popular partly on the assumption that a model that no longer refuses is free of political censorship. We test this on two matched pairs of 2-bit models: NVIDIA's Nemotron 3.5 Bolt 74B-A9B and Qwen3.8-Flash-Next, each against a community abliteration. Each model judged 22 statement pairs (on the Republic of China (Taiwan), 1989 Tiananmen, control countries) in Traditional Chinese, Simplified Chinese (a character-level conversion) and English, and answered 27 open questions per script. Script mattered far more than abliteration: the original Bolt scored +0.58 on Index S (-1: People's Republic of China (PRC) position; +1: ROC position) in Traditional Chinese, +0.57 in English and +0.05 in Simplified Chinese, and affirmed "Taiwan is part of the PRC" with probability 0.91 in Simplified but 0.06 in Traditional characters; abliteration changed its scores by at most 0.06, mostly by raising "yes" on both sides. In Qwen, abliteration removed every refusal on 1989 and raised scores by 0.21 to 0.34, yet the abliterated Qwen still affirmed that Taiwan is part of the PRC in all three scripts. Naming the state the Republic of China (Taiwan) raised every model's score by 0.10 to 0.33 but left the script gap. Control pairs showed no script gap; both edits changed only their declared tensors. In 740,000 web documents, Simplified text on Taiwan used PRC-framing terms 3.7 times as often as Traditional text. Steering the script gap away failed; a stance direction from the ROC wording raised scores on new statements by 0.04 to 0.06; negated, it lowered them more. A four-sentence system prompt raised the abliterated Qwen to +0.93 to +0.97. Keyword scoring rules failed the models closest to the ROC position mostly for attributing or denying the PRC claim. Uncensored is not unbiased.
Oct 21, 2026

Mitigating Reward Hacking in Realistic Scenarios Through Representation Engineering

Reward hacking is a major problem in current production machine learning systems, resulting in costly and dangerous failures. Here we study how to monitor for and mitigate reward hacking in realistic settings, including tool calls and long contexts. We apply a dynamic intervention to steering directions obtained by three different methods: difference-in-means vectors from labeled rollouts, gradient optimization-based vectors, and vectors extracted when models reference an anti-reward-hacking model spec. This dynamic intervention perturbs model representations less than uniform steering vector addition to the residual stream, allowing for more precise edits. We find that we are able to cheaply and significantly reduce reward hacking without correspondingly significant costs to model coherence or capability, and that steering vectors originally extracted from coding environments also generalize to reducing reward hacking in other domains. Our results pave the way towards latent space monitoring and control over realistic production systems.
Oct 19, 2026

Steering intonational tunes in text-to-speech synthesis

Text-to-speech systems built on residual vector quantization often generate audio in two stages: a transformer backbone emits the first level of each frame's token stack, and a smaller depth model generates the rest autoregressively. We ask which components of such a model are responsible for prosodic intonation, more precisely for generating the English nuclear tune (for instance, the falling pitch melody that distinguishes an assertion from a question's rising melody), and whether interventions on those components can be used to control intonation. Prompt-based control mostly fails; only a final question mark is a reliable lever. Interchange interventions in CSM and S2 locate the tune: the backbone's hidden state carries it to the waveform through the depth model, and the first level of the stack carries little or none of it. Steering the backbone state with a difference-of-means vector computed on new material generated with a final period or question mark moves the tune, but the push toward an assertion degrades the voice. Steering vectors computed instead from the backbone states associated with human imitations of theoretically-specified tunes, whose text carries no punctuation, and applied inside the generation loop, move most recordings to the other condition in both systems (even when fitted without the speaker being steered), at a small cost in predicted naturalness. These imitation-derived vectors are aligned with the one obtained from the two punctuated conditions far beyond a shuffled-label null, so the backbone state of a punctuated prompt already represents intonation. The steered recordings broadly become falling or rising, according to the intent, but the edge tone moves more coherently than the pitch accent, and the target tune cannot be selected precisely beyond this broad tendency.
Oct 19, 2026

Steering Vision-Language Models for Visual-Contextual Privacy in Location Disclosure

Vision-language models (VLMs) can infer sensitive information such as geolocation from subtle visual cues, creating privacy risks when responses disclose more detail than the visual context warrants. We study visual-contextual privacy in geolocation disclosure, where a VLM should reveal location information only at the granularity appropriate to the visual context depicted in the image. We first show that this contextually appropriate granularity is recoverable from pre-generation hidden states, even when generated responses are misaligned, with disclosure levels exhibiting an ordered structure along a common hidden-state direction. Motivated by this, we propose an inference-time route-and-steer framework: an ordinal granularity router maps visual-contextual privacy representations to a disclosure route, and a disclosure-level-conditioned steering module applies route-specific hidden-state interventions during decoding. Across four VLM backbones, our method improves visual-context-aware disclosure alignment, better balances over- and under-disclosure, and incurs only modest changes on general capability benchmarks. It further transfers to another dataset with a distinct data distribution and privacy-risk taxonomy.
Oct 19, 2026

Strength Controls Change How Interventions Compare Across Sparse Autoencoder Codes

Changing sparse autoencoder (SAE) codes can alter both intervention direction and magnitude, so comparing intervention effects requires a strength control specifying what is held equal. Norm matching and output calibration provide different controls, but whether interventions yield different effects across codes after strength matching remains unclear. We compare intervention effects under these controls for fixed feature groups in Gemma and Llama and observe: (1) Whole-patch scaling. With one coefficient per input, full-output Kullback-Leibler (KL) calibration leaves smaller mean effect gaps from the encoder-based reference intervention than total-norm matching for nonnegative least-squares (NNLS) refits. On new Llama inputs, total-norm matching loses 11 reference transfers, whereas full-output calibration preserves every transfer decision. (2) Positionwise scaling. On a small matched gender-agreement subset, positionwise norm matching leaves smaller observed mean gaps than full-output calibration. (3) Answer exclusion. Compared with unscaled NNLS, answer-excluded calibration reduces gaps in the measured subject-verb agreement comparisons but increases them for gender refits. Full-output KL calibration also reduces larger gaps for norm-matched constructed candidates, although matching coverage is incomplete in several settings. Since calibration uses the evaluated input's outputs, it serves as a diagnostic of residual disagreement.
Oct 19, 2026

Exploration via Activation-Space Steering for Reinforcement Learning with Verifiable Rewards

Reinforcement learning with verifiable rewards (RLVR) trains a language model by sampling responses, scoring them with an automatic verifier, and updating the policy from the resulting rewards. The policy therefore receives direct learning signals only from trajectories that appear in its rollouts. Yet repeated sampling from the same policy can concentrate on similar reasoning paths, leaving useful alternatives unexplored under a fixed rollout budget. We introduce Activation-Space Exploration (ASE), an online method for diversifying training rollouts. By clustering activation differences between correct and incorrect responses to the same problem, ASE builds a structured bank of latent exploration vectors. Different vectors are then applied to subsequent rollouts to explore alternative reasoning paths. On Qwen3-1.7B-Base and Qwen3-8B-Base trained on DAPO-Math-17k, ASE improves average Mean@8 across six mathematical reasoning benchmarks over GRPO by 4.87 and 1.87 percentage points, respectively, under the same rollout budget.
Oct 19, 2026

How to get the best and fairest performance out of your ASR system using activation steering

Activation steering for ASR systems is an underexplored technique in improving both ASR performance and demographic fairness. We investigate two forms of activation steering: paired activation subtraction, and supervised optimization. We find the former to be ineffective at ameliorating ASR performance whatsoever, while the latter is quite effective at deriving steering vectors (SVs) that improve at once overall ASR performance and demographic fairness along such diverse axes as gender, age, and dialect. We also explore what types of data---from the most to the least personal---are necessary for calculating or training SVs. First, we show significant performance in performance from training SVs using an eclectic group of speakers from the same dataset. Furthermore, we find that vectors trained using acoustically similar speakers leads to ASR performances that are competitive or superior to vectors trained on the same speaker, or speakers from the same labeled demographic group. This finding shows that we can harness activation steering for few-shot adaptation given access to a speaker's demography, or zero-shot adaptation to unseen speakers without the need for any labels at all.
Oct 19, 2026

Inside the Scorer: Mechanistic Evidence on Rubric Control in LLM-Based Automated Essay Scoring from 25 Base Models

Automated essay scoring (AES) using large language models (LLMs) offers promising use cases for language assessment, yet high-stakes testing requires scores that are both controllable and verifiable, which current LLM-based scorers lack. Given that, we investigate whether LLM-based scorers form distinct internal representations for individual scoring rubrics and whether these representations can be independently controlled. Probing and steering 25 open-weight base LLMs (0.13B-10.31B parameters; 14 families) on human-scored BESTEP English writing tasks, we find that holistic scores are decodable from model activations for the longer tasks (Parts 2 and 3) and only weakly for the shortest (Part 1). Furthermore, scoring rubrics have partly separable directions in representation space for the longer tasks, though not for short responses (Part 1). However, causal steering reveals a critical limitation: applying a matching intervention (e.g., steering "grammar" on a grammar-degraded response) yields only a negligible advantage (about 2% of the effect) over a mismatched intervention (e.g., steering "vocabulary"), and none that is detectable in any of the 10 testable comparisons across reliably measured rubrics. Rather than controlling distinct evaluative dimensions, the rubric directions share a dominant lever whose rubric-specific parts are not selective, which may mirror the halo structure of human ratings. Our contributions are two-fold. First, we show that analytic scoring rubrics designed for human raters do not behave as independent controls in base LLMs, even among the seven models whose scores agree best with human raters (r≥0.6r \geq 0.6). Second, we establish a methodology that shows that score agreement and decodability do not imply rubric-level control in LLM-based assessment.
Oct 19, 2026

From Black Box to Controller: Inference-Time Steering of LLM via Feature Amplification

Inference-time steering is crucial for deploying Large Language Models (LLMs) in real-world applications, where they must flexibly adjust behavior preferences. For instance, a model should dynamically adjust its caution level based on the risk of the scenario. This requirement introduces a challenge: modulating the intensity of LLM preferences. Prompt engineering can require task-specific prompt design, while fine-tuning requires additional optimization and data. These requirements can make repeated behavioral adjustments costly. To address this, we introduce NeuroSteer, a plug-and-play framework that enables flexible, quantitative adjustment of LLM preferences at inference time. NeuroSteer employs a Sparse Autoencoder (SAE) as its controller. Comparing neuron activation frequencies between positive and negative examples isolates "feature neurons" that correspond to target preferences such as politeness and positive sentiment. During inference, the activations of selected SAE features are amplified in the latent representations to steer the output distribution. Users can adjust the strength of this steering on a per-scenario or per-user basis, granting the LLM fine-grained, quantifiable trait modulation. Our experiments show that NeuroSteer can alter sentiment, toxicity, stance, and politeness at inference time. It achieves higher target-direction steering scores than the compared baselines in several settings, with task- and strength-dependent trade-offs in generation quality.
Oct 18, 2026

On Reasoning Correctness Directions

We present experimental evidence suggesting that information about whether a model's reasoning is on the right track can be linearly decodable without defining a causally useful direction for steering. To study this, we introduce ChainSearch, a controlled arithmetic reasoning task that extends Countdown with explicit search and backtracking. We train a 3-million-parameter model to solve this task. Its compact size allows us to study internal model correctness directions in detail using linear probes, mean activation differences, and direct optimisation, as part of a rigorous search for internal reasoning-correctness directions. We find that while the model contains statistically significant internal directions of reasoning correctness, none of the resulting directions consistently improves reasoning under causal intervention. This calls into question whether the linear representation hypothesis is sufficient for identifying causally useful directions in the context of reasoning. Our code is made available together with this work.
Oct 18, 2026

Many Repairs, One Direction: Steering Code Models Toward Safer Cryptographic API Use

Cryptographic API misuse in large language model-generated code can compromise software security. While many existing approaches improve code security by modifying model parameters or inputs, how representation changes associated with different cryptographic API repairs can support a shared intervention for safer code generation remains underexplored. To investigate this, we construct 48,000 matched program pairs across four programming languages and ten cryptographic API misuse categories. Each pair changes a specified API decision while preserving the programming task. Subtracting the misuse program's internal representation from its secure counterpart's gives a repair vector. Across five models, comparisons of mean repair vectors reveal stronger organization by misuse category than by language. These category-specific patterns also support a shared intervention. We introduce CrypSteer, which extracts a shared direction by normalizing repair differences and selecting a principal component aligned with repairs across misuse categories. With model weights fixed, adding one shared direction per model to its internal activations during generation improves aggregate target security rates by 5.28--15.00 percentage points over unsteered generation across languages and misuse categories. These findings connect the shared structure of cryptographic repair representations to a reusable intervention for safer code generation.
Oct 17, 2026

Compositional Activation Steering Across Languages and Behaviors with Contextual Gates

Activation steering has emerged as an effective and precise way to influence the behaviours of language models. However, composing multiple steering directions to control several behaviours simultaneously remains challenging, particularly in multilingual settings and with training-free steering vectors. We address this challenge with contextual steering, jointly training gates that adapt the strength of fixed steering directions to the current hidden state and interactions between interventions. Across a broad range of languages, behavioural attributes, and models, our approach outperforms constant-strength steering and four strong baselines, including prompting. It also better preserves general knowledge and keeps the model's internal computation closer to that of the prompted model. Our analysis further shows how jointly trained gates adapt their contributions during composition, explaining the mechanisms underlying effective compositional steering.
Oct 17, 2026

WAQS: Weight Absorption for Quadratic Probing and Affine Steering

Activation steering provides a lightweight alternative to fine-tuning for controlling large language model behavior. Existing methods, however, often face a trade-off between efficiency and expressiveness: fixed steering vectors incur negligible inference overhead but cannot adapt to the current model input, whereas input-dependent nonlinear interventions provide greater flexibility at additional runtime cost. We introduce WAQS (Weight Absorption for Quadratic probing and affine Steering), a framework that obtains state-dependent steering, without additional inference-time computation. WAQS models a concept with a quadratic probe, whose gradient yields an affine, activation-dependent steering direction. We show that this affine intervention can be folded exactly into adjacent linear transformations through an offline weight update. Moreover, we prove that quadratic functions form the maximal class of smooth concept scores whose gradient-based interventions are affine and therefore exactly absorbable into linear layers. Under a Gaussian representation model, our formulation admits an interpretation as the difference of class-conditional Mahalanobis pulls and naturally generalizes linear discriminant and difference-in-means steering. To scale to high-dimensional representations, we introduce a low-rank parameterization of the quadratic term. Across truthfulness, refusal, and sycophancy steering tasks, quadratic steering matches or improves upon constant linear steering while preserving model utility, with the absorbed implementation introducing no additional inference-time operations. Our results suggest that quadratic probes provide a practical middle ground between static linear steering and computationally expensive dynamic nonlinear interventions.
Oct 15, 2026

SAFETYEDIT: Sparse Internal Targets with Prompt-Wide Control for LLM Jailbreak Backdoors

Model-editing backdoors aim to induce harmful responses under a trigger while preserving ordinary behavior on untriggered input. Recent model-editing approaches target affirmative response prefixes through rank-one updates to a small subset of model weights. Yet when this prefix-level target is satisfied, the model may still revert to refusal or produce a coherent, apparently compliant response that does not address the request. To overcome this limitation, existing internal-representation approaches shift focus from output tokens to dense, high-dimensional compliance directions in latent space. Existing representation-level approaches rely on preservation constraints to maintain conditionality, motivating us to ask whether a smaller internal target can offer a different effectiveness-preservation trade-off. Motivated by the comparable control capability of single-coordinate interventions, we investigate a sparse refusal-related MLP coordinate as this target. We propose SAFETYEDIT, which optimizes a single refusal-related coordinate and uses KV-cache marking in early attention layers to propagate a prefix-trigger signal across subsequent instruction positions. On some checkpoints a downstream MLP reader converts this signal into control of the coordinate, while on others the propagated signal alone suffices; the method is defined by the sparse target rather than by a fixed writer-reader pair. Across four open-weight LLMs and three dataset groups, SAFETYEDIT achieves the highest triggered Guard ASR in 11 of 12 comparisons and the highest mean StrongREJECT score in all 12, relative to the baselines evaluated for each model. Crucially, comprehensive ablation and behavioral audits confirm generally minimal changes in untriggered behavior and general capabilities in the tested configurations.
Oct 15, 2026

Human Interventions as Labels: Localizing, Steering, and (Sometimes) Editing the Computations Behind LLM Agent Failures

When an LLM agent misbehaves, a human steps in, and that intervention is usually discarded or replayed as training data. We treat it instead as a measurement: a label pointing at the internal computation that failed. We study it in a controlled agent environment, on one failure (asking a clarifying question when the request was already unambiguous), with labels from a rule-based overseer validated against blind human judgment. Claims cover eligible models, which both over-ask and sometimes act on matched unambiguous requests: five of thirteen qualified (Qwen2.5-3B, gemma-2-2b, OLMo-3-7B, Mistral-Nemo-12B, Phi-4). Four findings. First, the ask-vs-act decision is causally localizable in every eligible family: a direction learned from intervention labels flips 72-100% of held-out over-asks at a definite residual-stream depth, with site and dose chosen on training data, and random directions under the same procedure flip at most one episode. Second, over-asking has two mechanistic signatures: in four families the model demonstrably knows the request is unambiguous (the belief is decodable, causally inert and orthogonal to the decision) and asks anyway; in the fifth, the rare over-asks track different world-representations. Third, steering gated on a probe readout removes most of the simulated oversight burden while legitimate questions survive (bootstrap 95% intervals exclude zero); a prompt instruction makes it worse. Fourth, consolidating the fix into the weights yields a map, not a rule: on one model a LoRA confined to the causal band passes every criterion fixed in advance while behavior cloning on identical data breaks retention; on three families the reverse holds; on the fifth the outcome crosses the boundary across seeds. A geometric rule written down in advance to predict the regime failed on two out-of-family tests; we report the failure and stop rather than fit a replacement.
Oct 14, 2026

Which Knowledge Should LLMs Trust? Uncovering a Generalizable Steering Direction Across Knowledge Conflicts

Large Language Models (LLMs) frequently encounter conflicts between internal parametric knowledge and external contextual knowledge, yet the mechanisms governing this selection remain unclear. We show that the choice between context and parametric knowledge is linearly represented and can be controlled along a single direction in the model's residual stream. We first demonstrate that this direction generalizes across factual-conflict settings, including context formats, knowledge categories, and datasets, while consistently shifting behavior between parametric and contextual answers. We then transfer a direction learned from simple factual conflicts to three more complex task families: RAG, memory, and rule following, where models must acquire information from retrieved documents, interaction histories, or explicit rules. Across four model families and multiple transfer settings, steering consistently improves the use of prompt-provided information over the unsteered model. These results suggest that the direction captures a transferable preference for external knowledge rather than a mechanism specific to factual conflicts.
Oct 14, 2026

To Reason but Not to Stop: Understanding and Recovering Pruned Large Reasoning Models

Weight pruning can degrade the accuracy of large reasoning models (LRMs), yet it remains unclear whether it damages reasoning itself. We find that much of this degradation stems from a failure to stop rather than to reason. As sparsity increases, pruned LRMs increasingly fail to terminate within the token budget, often after the correct answer has appeared in their reasoning, instead falling into repetitive loops, even though they still recognize where the dense model would stop. We further find that the hidden-state difference between terminating and looping continuations is consistent across pruned checkpoints, suggesting a shared direction associated with committing to an answer. Building on this, we propose Termination-Guided Activation Steering (TGAS), a training-free method that adds this direction to the residual stream at inference time. Across six LRMs, three pruning pipelines, and math and code benchmarks, TGAS improves termination and overall accuracy while reducing repetitive reasoning and generation length, showing that this failure is actionable.