A Coin Flip Per Token: Bernoulli Sparse Steering of Large Language Models

• URL: http://arxiv.org/abs/2607.05615v1
• Date: Mon, 06 Jul 2026 20:25:27 GMT
• Title: A Coin Flip Per Token: Bernoulli Sparse Steering of Large Language Models
• Abstract summary: Activation steering via sparse autoencoders (SAEs) enables behavioral control of large language models without task-specific fine-tuning, but standard methods apply the steering signal at every generated token.<n>We introduce Token Steering (STS), which gates each token independently with probability $p$, and Block Steering (SBS), which gates a leading window once per sequence.<n>Across two model families and two behavioral tasks, steering only 50% of the tokens recovers most of dense-steering effect while preserving fluency, and steering as few as 30% surpasses prompt-based control.
• Score: 12.373460256583583
• License: http://creativecommons.org/licenses/by/4.0/
• Abstract: Activation steering via sparse autoencoders (SAEs) enables behavioral control of large language models without task-specific fine-tuning, but standard methods apply the steering signal at every generated token, incurring constant per-token perturbation that risks degrading fluency. We ask: is dense intervention necessary? We introduce Stochastic Token Steering (STS), which gates each token independently with probability $p$, and Stochastic Block Steering (SBS), which gates a leading window once per sequence; neither requires a reward model or learned gating policy. Across two model families and two behavioral tasks, steering only 50% of the tokens recovers most of the dense-steering effect while preserving fluency, and steering as few as 30% surpasses prompt-based control. The optimal steering magnitude scales inversely with the intervention ratio, revealing that SAE-mediated control is rate-limited: the behavioral outcome depends on cumulative signal dosage across a sequence.

Related papers

• When is Your LLM Steerable? [56.656180566692946]
  Activation steering offers a lightweight approach to control language models' behavior at inference time.<n>Finding the regime and boundaries of successful steering typically requires expensive grid searches.<n>We investigate whether steerability can be predicted from the model's internal states at the beginning of the generation process.
  arXiv  Detail & Related papers  (2026-06-10T02:55:34Z)
• PriFT: Prior-Support Guided Supervised Fine-Tuning [74.65198014829393]
  Supervised fine-tuning (SFT) is an efficient approach for downstream task adaptation.<n>SFT fits fixed demonstrations token by token, including targets poorly aligned with the model's pretrained distribution.<n>We propose PriFT, which derives token weights from a frozen pretrained reference to obtain a stable reweighting signal unaffected by fine-tuning.
  arXiv  Detail & Related papers  (2026-06-08T12:14:06Z)
• Statistically Grounded Sparse-Feature Interventions for Activation-Space Control in Large Language Models [2.464003792743989]
  Activation steering offers a lightweight alternative to fine-tuning for behavioral control of large language models.<n>We introduce a transparent SAE-feature steering pipeline that ranks sparse features through an unweighted Borda consensus.<n>Across three Gemma-family models, four behavioral domains, and 356 layer-strength configurations, the method produces measurable domain-specific shifts.
  arXiv  Detail & Related papers  (2026-06-05T19:27:31Z)
• Steer Where It Matters: Token-Level Visual-Sensitivity Steering for LVLMs Hallucination Mitigation [59.80828456543202]
  We propose Token-Level Visual-Sensitivity Steering (TLVS) for hallucination mitigation.<n>This mechanism requires only minimal training for calibration and can be applied across diverse vision-language models.<n>We evaluate TLVS on several benchmarks, including POPE, AMBER, CHAIR (COCO), MMHal, and HallusionBench, demonstrating consistent improvements over previous steering methods.
  arXiv  Detail & Related papers  (2026-06-02T08:23:28Z)
• Primary-Fine Decoupling for Action Generation in Robotic Imitation [91.2899765310853]
  Multi-modal distribution in robotic manipulation action sequences poses critical challenges for imitation learning.<n>We propose Primary-Fine Decoupling for Action Generation (PF-DAG), a two-stage framework that decouples coarse action consistency from fine-grained variations.<n>PF-DAG outperforms state-of-the-art baselines across 56 tasks from Adroit, DexArt, and MetaWorld benchmarks.
  arXiv  Detail & Related papers  (2026-02-25T08:36:45Z)
• OAT: Ordered Action Tokenization [44.20363344414952]
  Autoregressive policies offer a compelling foundation for scalable robot learning by enabling discrete abstraction, token-level reasoning, and flexible inference.<n>Existing approaches either rely on analytical discretization methods that produce prohibitively long token sequences, or learned latent tokenizers that lack structure.<n>In this work, we identify three desiderata for action tokenization - high compression, total decodability, and a left-to-right causally ordered token space - and introduce Ordered Action Tokenization (OAT)<n>OAT discretizes action chunks into an ordered sequence of tokens using transformer with registers, finite scalar quant
  arXiv  Detail & Related papers  (2026-02-04T05:01:04Z)
• Selective Steering: Norm-Preserving Control Through Discriminative Layer Selection [1.7802147489386628]
  Large language models (LLMs) remain vulnerable to adversarial attacks that elicit harmful behaviors.<n>We propose Selective Steering, which addresses these limitations through two key innovations.<n> Experiments across nine models demonstrate that Selective Steering achieves 5.5x higher attack success rates than prior methods.
  arXiv  Detail & Related papers  (2026-01-27T08:56:25Z)
• Dynamically Scaled Activation Steering [3.177576903071419]
  We introduce Dynamically Scaled Activation Steering (DSAS), a method-agnostic steering framework that decouples when to steer from how to steer.<n>DSAS adaptively modulates the strength of existing steering transformations across layers and inputs, intervening strongly only when undesired behavior is detected.
  arXiv  Detail & Related papers  (2025-12-03T10:50:15Z)
• The Rogue Scalpel: Activation Steering Compromises LLM Safety [11.402179030703188]
  Activation steering is a technique for controlling LLM behavior by adding semantically meaningful vectors directly into a model's hidden states during inference.<n>We demonstrate the opposite: steering systematically breaks model alignment safeguards, making it comply with harmful requests.
  arXiv  Detail & Related papers  (2025-09-26T08:49:47Z)
• GrAInS: Gradient-based Attribution for Inference-Time Steering of LLMs and VLMs [56.93583799109029]
  GrAInS is an inference-time steering approach that operates across both language-only and vision-language models and tasks.<n>During inference, GrAInS hidden activations at transformer layers guided by token-level attribution signals, and normalizes activations to preserve representational scale.<n>It consistently outperforms both fine-tuning and existing steering baselines.
  arXiv  Detail & Related papers  (2025-07-24T02:34:13Z)
• R-Stitch: Dynamic Trajectory Stitching for Efficient Reasoning [80.104336426172]
  Chain-of-thought (CoT) enhances problem-solving ability of large language models.<n>CoT incurs substantial inference cost due to long autoregressive trajectories.<n>We introduce R-Stitch, a training-free hybrid decoding framework.
  arXiv  Detail & Related papers  (2025-07-23T08:14:36Z)

This list is automatically generated from the titles and abstracts of the papers in this site.

This site does not guarantee the quality of this site (including all information) and is not responsible for any consequences.