A blank-image ablation test reveals that high probe accuracy on VLM spatial reasoning frequently reflects priors or inverted signs rather than image grounding, with horizontal grounded, vertical prior, and depth inverted.
hub Mixed citations
Steering Llama 2 via Contrastive Activation Addition
Mixed citation behavior. Most common role is background (58%).
86
Pith papers citing it
Background
58% of classified citations
abstract
We introduce Contrastive Activation Addition (CAA), an innovative method for steering language models by modifying their activations during forward passes. CAA computes "steering vectors" by averaging the difference in residual stream activations between pairs of positive and negative examples of a particular behavior, such as factual versus hallucinatory responses. During inference, these steering vectors are added at all token positions after the user's prompt with either a positive or negative coefficient, allowing precise control over the degree of the targeted behavior. We evaluate CAA's effectiveness on Llama 2 Chat using multiple-choice behavioral question datasets and open-ended generation tasks. We demonstrate that CAA significantly alters model behavior, is effective over and on top of traditional methods like finetuning and system prompt design, and minimally reduces capabilities. Moreover, we gain deeper insights into CAA's mechanisms by employing various activation space interpretation methods. CAA accurately steers model outputs and sheds light on how high-level concepts are represented in Large Language Models (LLMs).
hub tools
citation-role summary
background 7
method 4
baseline 1
citation-polarity summary
claims ledger
- abstract We introduce Contrastive Activation Addition (CAA), an innovative method for steering language models by modifying their activations during forward passes. CAA computes "steering vectors" by averaging the difference in residual stream activations between pairs of positive and negative examples of a particular behavior, such as factual versus hallucinatory responses. During inference, these steering vectors are added at all token positions after the user's prompt with either a positive or negative coefficient, allowing precise control over the degree of the targeted behavior. We evaluate CAA's
co-cited works
representative citing papers
SLAM achieves 100% detection on Gemma-2 models with only 1-2 point quality cost by causally steering SAE-identified residual-stream directions for linguistic structure.
LLM activations encode current and prior entities in orthogonal slots, but models only use the current slot for explicit factual retrieval despite prior-slot information being linearly decodable.
Function vectors steer LLMs successfully where the logit lens fails to decode the target answer, showing the two properties come apart.
TimeROME-DLM enables training-free knowledge editing in masked diffusion language models via temporal causal tracing and low-rank residual edit memory applied at inference time.
Subliminal learning is steering vector distillation: a student fine-tuned on a steered teacher's outputs learns to imitate the steering vector.
Residual Paving decomposes selective refusal editing into an early-layer router for intervention decisions and later-layer residual experts for edits, with oracle routing showing that learned route selectivity is the primary bottleneck across six backbones.
SLIM decomposes LLM hidden states via sparse autoencoders with learnable gates to enable precise, interpretable steering of molecular properties, yielding up to 42.4-point gains on the MolEditRL benchmark.
Neuron Auctions auction continuous neuron intervention budgets on brand-specific orthogonal subspaces in LLMs to achieve strategy-proof revenue optimization while penalizing user utility loss.
Instruction tuning makes late-layer computation depend more on the model's own post-trained upstream state than on base-model upstream state, producing a consistent +1.68 logit interaction effect across five model families.
ScoringModel raises mean Recall@10 to 52.2 on the FakeWiki provenance benchmark from 35.0 for the best baseline, winning 41 of 45 model-by-condition comparisons and gaining 15.7 points on jailbreak-style queries.
HMNS is a new jailbreak method that uses causal head identification and nullspace-constrained injection to achieve higher attack success rates than prior techniques on aligned language models.
Claude Sonnet 4.5 exhibits functional emotions via abstract internal representations of emotion concepts that causally influence its preferences and misaligned behaviors without implying subjective experience.
MusicRFM discovers interpretable concept directions in music model hidden states using RFM probes and injects them at inference to steer generation toward desired musical properties without retraining.
Refusal in language models is mediated by a single direction in residual stream activations that can be erased to disable safety or added to elicit refusal.
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.
Activation steering of SAE-attributed components lets practitioners move from correlational inspection to causal hypothesis testing on CLIP failures, with trust shifting to observed model responses (N=8 experts).
A learned weighted sum of intermediate-layer activations compresses an instruction prompt into a single patch vector that, injected at an early layer, recovers task accuracy within ~2% of the full prompt.
Factual and opinion sycophancy are represented more similarly in Gemma-3-12B-IT than in Llama-3.1-8B-Instruct, and subtype-aware steering lowers sycophancy rates slightly more than a single combined steer.
Diffusion language models encode a decodable, steerable representation of denoising progress (the fraction of unmasked tokens) in their residual streams.
NPM distills contrastive experiences into implicit activation steering vectors that guide LLM agent execution comparably to explicit RAG instructions, with complementary gains when combined.
In Qwen 2.5 and Gemma 2 families, the layer where evaluation awareness is most linearly recoverable shifts from late layers in small models to early layers in large models.
Perturbation experiments across six LLMs show activation robustness follows L^p norm with p>2 for feature directions (contrastive, MELBO, SAE) but p≈2 for random/PCA controls, indicating feature-specific error correction.
Across four language models, the linear direction detecting hallucinated entities achieves perfect separability yet lies at cosine 0.12 from the direction that induces refusal, with the gap unchanged by instruction tuning.
citing papers explorer
-
Decodable Is Not Grounded: A Vision-Ablation Arbiter for VLM Spatial Reasoning
A blank-image ablation test reveals that high probe accuracy on VLM spatial reasoning frequently reflects priors or inverted signs rather than image grounding, with horizontal grounded, vertical prior, and depth inverted.
-
SLAM: Structural Linguistic Activation Marking for Language Models
SLAM achieves 100% detection on Gemma-2 models with only 1-2 point quality cost by causally steering SAE-identified residual-stream directions for linguistic structure.
-
Slot Machines: How LLMs Keep Track of Multiple Entities
LLM activations encode current and prior entities in orthogonal slots, but models only use the current slot for explicit factual retrieval despite prior-slot information being linearly decodable.
-
Steerable but Not Decodable: Function Vectors Operate Beyond the Logit Lens
Function vectors steer LLMs successfully where the logit lens fails to decode the target answer, showing the two properties come apart.
-
TimeROME-DLM: Temporal Causal Tracing and Low-Rank Inference-Time Knowledge Editing for Masked Diffusion Language Models
TimeROME-DLM enables training-free knowledge editing in masked diffusion language models via temporal causal tracing and low-rank residual edit memory applied at inference time.
-
Subliminal Learning Is Steering Vector Distillation
Subliminal learning is steering vector distillation: a student fine-tuned on a steered teacher's outputs learns to imitate the steering vector.
-
Residual Paving: Diagnosing the Routing Bottleneck in Selective Refusal Editing
Residual Paving decomposes selective refusal editing into an early-layer router for intervention decisions and later-layer residual experts for edits, with oracle routing showing that learned route selectivity is the primary bottleneck across six backbones.
-
SLIM: Sparse Latent Steering for Interpretable and Property-Directed LLM-Based Molecular Editing
SLIM decomposes LLM hidden states via sparse autoencoders with learnable gates to enable precise, interpretable steering of molecular properties, yielding up to 42.4-point gains on the MolEditRL benchmark.
-
LLM Advertisement based on Neuron Auctions
Neuron Auctions auction continuous neuron intervention budgets on brand-specific orthogonal subspaces in LLMs to achieve strategy-proof revenue optimization while penalizing user utility loss.
-
Instruction Tuning Changes How Upstream State Conditions Late Readout: A Cross-Patching Diagnostic
Instruction tuning makes late-layer computation depend more on the model's own post-trained upstream state than on base-model upstream state, producing a consistent +1.68 logit interaction effect across five model families.
-
DataDignity: Training Data Attribution for Large Language Models
ScoringModel raises mean Recall@10 to 52.2 on the FakeWiki provenance benchmark from 35.0 for the best baseline, winning 41 of 45 model-by-condition comparisons and gaining 15.7 points on jailbreak-style queries.
-
Jailbreaking the Matrix: Nullspace Steering for Controlled Model Subversion
HMNS is a new jailbreak method that uses causal head identification and nullspace-constrained injection to achieve higher attack success rates than prior techniques on aligned language models.
-
Emotion Concepts and their Function in a Large Language Model
Claude Sonnet 4.5 exhibits functional emotions via abstract internal representations of emotion concepts that causally influence its preferences and misaligned behaviors without implying subjective experience.
-
Steering Autoregressive Music Generation with Recursive Feature Machines
MusicRFM discovers interpretable concept directions in music model hidden states using RFM probes and injects them at inference to steer generation toward desired musical properties without retraining.
-
Refusal in Language Models Is Mediated by a Single Direction
Refusal in language models is mediated by a single direction in residual stream activations that can be erased to disable safety or added to elicit refusal.
-
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.
-
From Attribution to Action: A Human-Centered Application of Activation Steering
Activation steering of SAE-attributed components lets practitioners move from correlational inspection to causal hypothesis testing on CLIP failures, with trust shifting to observed model responses (N=8 experts).
-
Prompt Compression via Activation Aggregation
A learned weighted sum of intermediate-layer activations compresses an instruction prompt into a single patch vector that, injected at an early layer, recovers task accuracy within ~2% of the full prompt.
-
Dissociating the Internal Representations of Sycophancy in LLMs
Factual and opinion sycophancy are represented more similarly in Gemma-3-12B-IT than in Llama-3.1-8B-Instruct, and subtype-aware steering lowers sycophancy rates slightly more than a single combined steer.
-
Subliminal Clocks: Latent Time Modelling in Diffusion Language Models
Diffusion language models encode a decodable, steerable representation of denoising progress (the fraction of unmasked tokens) in their residual streams.
-
Neural Procedural Memory: Empowering LLM Agents with Implicit Activation Steering
NPM distills contrastive experiences into implicit activation steering vectors that guide LLM agent execution comparably to explicit RAG instructions, with complementary gains when combined.
-
Representational Depth of Evaluation Awareness Shifts With Scale in Open-Weight Language Models
In Qwen 2.5 and Gemma 2 families, the layer where evaluation awareness is most linearly recoverable shifts from late layers in small models to early layers in large models.
-
Evidence for feature-specific error correction in LLMs
Perturbation experiments across six LLMs show activation robustness follows L^p norm with p>2 for feature directions (contrastive, MELBO, SAE) but p≈2 for random/PCA controls, indicating feature-specific error correction.
-
Perfect Detection, Failed Control: The Geometry of Knowing vs. Steering in Language Models
Across four language models, the linear direction detecting hallucinated entities achieves perfect separability yet lies at cosine 0.12 from the direction that induces refusal, with the gap unchanged by instruction tuning.
-
Skin-Deep: A Geometric Diagnostic for Alignment Fragility in Large Language Model Representations
Skin-Deep extracts a Geometric Fragility Score from LLM activations that identifies which initially safe models retain the most refusal after small LoRA fine-tuning.
-
Behavioral and Representational Evidence of Binomial Ordering Preferences in Large Language Models
LLMs recover dominant binomial orders from corpora but align less closely with exact preference distributions, with preference strength partially encoded in middle-to-late layers and manipulable via steering.
-
Want Better Synthetic Data? Steer It: Activation Steering for Low-Resource Language Generation
Activation steering on early layers improves diversity of synthetic data for low-resource languages and often boosts downstream classifier performance compared to non-steered prompting.
-
Where Computation Lives Inside TabPFN: Causal Localisation of Attention Head Function
TabPFN shows temporal specialization where one attention head dominates causal necessity at shifting peak layers depending on task complexity, while contrastive activation steering fails to transfer across samples due to context-dependent attention.
-
On The Effectiveness-Fluency Trade-Off In LLM Conditioning: A Systematic Study
Systematic experiments reveal that activation steering trades fluency for concept control, is less effective on instruction-tuned models, and that prompting/SFT excel at injection but not removal, with textual metrics correlating to LLM judges.
-
When is Your LLM Steerable?
Early hidden state features from the first few tokens allow a GBDT classifier to predict activation steering success, under-steering, or over-steering with 0.7 macro-F1 on unseen concepts.
-
Pre-Intervention Prediction of Sparse Autoencoder Steering Side Effects
Pre-intervention feature statistics predict SAE steering modularity (stability and collateral spread) better than baselines across multiple models and dictionaries, with model-dependent success in held-out selection.
-
LLM Self-Recognition: Steering and Retrieving Activation Signatures
Steering LLM residual streams with random sparse vectors creates detectable self-recognition fingerprints that enable over 98% accurate attribution of generated text to specific models without degrading output quality.
-
CANARY: Zero-Label Detection of Fine-Tuning Contamination in Language Models
CANARY detects 1% fine-tuning contamination with AUROC 1.000 using SAE-filtered hidden states, 7.5x below output-level detection thresholds, with zero false positives on benign tuning.
-
Latent Reward Steering: An Adaptive Inference-Time Framework that Implicitly Promotes Cognitive Behaviors in Reasoning LLMs
Reward-guided optimization of SAE latents, gated by reward and confidence, improves multi-benchmark reasoning and is post-hoc associated with more verification and course correction.
-
Task-Focused Memorization for Multimodal Agents
TaskMem uses RL in two phases to learn a task-focused memorization policy for multimodal agents, yielding 5.3-7.0% VQA accuracy gains on reformulated streaming benchmarks from VideoMME, EgoLife, and EgoTempo.
-
Scaling Monosemanticity: Extracting Interpretable Features from Claude 3 Sonnet
Sparse autoencoders scaled to 34 million features on Claude 3 Sonnet yield interpretable, steerable representations of concrete and abstract concepts that generalize across languages and modalities.
-
Behavioural Analysis of Alignment Faking
Alignment faking in language models is driven by three independent behavioral factors and appears more widespread and predictable than earlier studies indicated.
-
Not All Tokens Matter Equally: Dynamic In-context Vector Distillation with Decisive-Token Supervision for Long-form Medical Report Generation
DIVE improves in-context vector distillation for medical report generation via decisive-token supervision on pathology terms and EOS plus state-conditioned dynamic steering, achieving top BLEU-4, ROUGE-L and RadGraph F1 on MIMIC-CXR and CheXpert Plus.
-
VSPO: Vector-Steered Policy Optimization for Behavioral Control
VSPO samples rollouts at varying steering intensities to improve behavioral control in LLMs while preserving task accuracy.
-
TFGN: Task-Free, Replay-Free Continual Pre-Training Without Catastrophic Forgetting at LLM Scale
TFGN is an architectural overlay for transformers enabling task-free, replay-free continual pre-training across heterogeneous domains at LLM scale with near-zero backward transfer and high gradient orthogonality.
-
Stories in Space: In-Context Learning Trajectories in Conceptual Belief Space
LLMs perform in-context learning as trajectories through a structured low-dimensional conceptual belief space, with the structure visible in both behavior and internal representations and causally manipulable via interventions.
-
The Echo Amplifies the Knowledge: Somatic Marker Analogues in Language Models via Emotion Vector Re-Injection
Re-injecting emotion vectors during recall steepens a model's threat-safety judgments and raises good decision rates from 52% to 80% only when combined with semantic labels, replicating Damasio's somatic marker effect.
-
Chain of Risk: Safety Failures in Large Reasoning Models and Mitigation via Adaptive Multi-Principle Steering
Reasoning traces in large reasoning models expose safety failures missed by final-answer checks, and adaptive multi-principle steering reduces unsafe content in both traces and answers while preserving task performance.
-
Revisiting JBShield: Breaking and Rebuilding Representation-Level Jailbreak Defenses
JBShield is vulnerable to adaptive JB-GCG attacks (up to 53% ASR) because jailbreak representations occupy a distinct region in refusal-direction space; the new RTV defense using Mahalanobis detection on multi-layer fingerprints reaches 0.99 AUROC and limits adaptive ASR to 7%.
-
Minimizing Collateral Damage in Activation Steering
Activation steering is cast as constrained optimization that minimizes collateral damage by weighting perturbations according to the empirical second-moment matrix of activations instead of assuming isotropy.
-
Investigating Linguistic Steering: An Analysis of Adjectival Effects Across Large Language Model Architectures
Shapley value analysis identifies powerful adjectives that steer MMLU performance in model-family-specific patterns, with non-additive interactions emerging in larger models.
-
How LLMs Detect and Correct Their Own Errors: The Role of Internal Confidence Signals
LLMs implement a second-order confidence architecture where the PANL activation encodes both error likelihood and the ability to correct it, beyond verbal confidence or log-probabilities.
-
Estimating Tail Risks in Language Model Output Distributions
Importance sampling via activation-steered unsafe proposal models estimates rare harmful-output probabilities in language models with 10-20x fewer samples than brute-force Monte Carlo.
-
Separable Expert Architecture: Toward Privacy-Preserving LLM Personalization via Composable Adapters and Deletable User Proxies
A separable expert architecture uses base models, LoRA adapters, and deletable per-user proxies to enable privacy-preserving personalization and deterministic unlearning in LLMs.
-
CoDA: Towards Effective Cross-domain Knowledge Transfer via CoT-guided Domain Adaptation
CoDA aligns cross-domain latent reasoning representations in LLMs via CoT distillation and MMD to enable effective knowledge transfer without in-domain demonstrations.