Can you change how an AI behaves by nudging its internal numbers mid-run — no retraining, no prompt tricks?
Can activation-space directions reliably steer LLM behavior without retraining or prompting?
This explores whether you can change how a model behaves by adding a direction to its internal numbers while it runs, with no weight updates and no prompt rewriting, and whether that works dependably.
This explores whether you can change how a model behaves by adding a direction to its internal numbers (its activations) while it runs, with no weight updates and no prompt rewriting, and whether that works dependably. The corpus says yes for a couple of well-chosen behaviors. It has little evidence on how far that stretches.
The cleanest case is brevity. Activation-Steered Compression takes 50 paired examples of verbose and concise reasoning, distills them into a single vector, and adds it at inference time. Chain-of-thought length drops 67%, accuracy holds, and generation runs 2.73x faster. The effect carries across model sizes and domains (Can we steer reasoning toward brevity without retraining?). Verbosity turns out to be a roughly straight line through the model's internal geometry rather than something smeared across it.
Personality traits look similar, with a twist. Sycophancy and hallucination-proneness also show up as linear directions, and they work better as a thermometer than as a dial. They predict how a finetuning run will shift the model's personality before it happens, and they can steer training away from the unwanted shift (Can we track and steer personality shifts during model finetuning?). That use is steering during finetuning, so it doesn't fit the training-free part of your question. Prompting is the obvious alternative, and it has a ceiling. Most open models won't adopt a personality they're prompted into and snap back to their trained ENFJ-like defaults (Can open language models adopt different personalities through prompting?). The corpus never tests steering against prompting head to head. But traits the model holds strong defaults about are where an internal lever might beat a verbal request.
"Reliably" is the part the evidence can't yet back. Both successes involve one well-defined behavior with clean paired examples. No note here tests a fuzzy behavior or reports a case where steering broke down. One reason for caution, and this is my inference rather than something the notes test: a handful of input-agnostic activations run up to 100,000× larger than the rest and act as built-in attention bias terms (Do hidden massive activations act as attention bias terms?). If some directions are that load-bearing, pushing along the wrong one could wreck the model instead of steering it.
The corpus also shows other ways to change behavior without retraining, and they work from outside the model rather than inside it. Four cognitive tools run as sandboxed LLM calls lifted GPT-4.1 on AIME2024 from 26.7% to 43.3% with no RL (Can modular cognitive tools unlock reasoning without training?). An episodic memory system reached 87.88% on GAIA without touching parameters (Can agents learn continuously from experience without updating weights?). If you want the behavior baked in, activation-level consistency training (ACT) does that, but it is a training method (Can models learn to ignore irrelevant prompt changes?). So steering sits in a specific spot. It is cheap and precise for behaviors that have a clean direction, and it has not yet been shown to work for arbitrary ones.
Sources 7 notes
Activation-Steered Compression extracts a single vector from 50 paired examples to reduce chain-of-thought length by 67% while maintaining accuracy and achieving 2.73x speedup. The method is training-free and generalizes across model sizes and domains.
Research identifies linear directions in LLM activation space corresponding to specific traits like sycophancy and hallucination. These persona vectors predict finetuning-induced personality shifts before they occur and can preventatively steer training to avoid unwanted trait changes.
Research shows most open models fail to adopt prompted personalities, stubbornly retaining their trained ENFJ-like defaults. Only a few flexible models succeed. Combining role and personality conditioning improves results but doesn't fully overcome resistance.
A very small number of input-agnostic activations with values up to 100,000× larger than others act as indispensable implicit bias terms and concentrate attention probability onto specific tokens. This phenomenon appears across model sizes and Vision Transformers.
Four cognitive tools implemented as sandboxed LLM calls improved GPT-4.1 on AIME2024 from 26.7% to 43.3% without any RL training. Modularity enforces operation isolation that pure prompting cannot guarantee, eliciting pre-existing reasoning capability.
Show all 7 sources
AgentFly formalizes agent learning as a Memory-augmented MDP with three memory modules (case, subtask, tool) that enable credit assignment and policy improvement entirely through memory operations. The approach achieved 87.88% on GAIA validation without modifying LLM parameters.
Two methods—BCT (output-level) and ACT (activation-level)—train models to respond identically to clean and wrapped prompts by using the model's own clean responses as targets, eliminating specification and capability staleness inherent in standard SFT.
Papers this line draws on 8
The research behind the notes this line reads — ranked by how closely each paper relates.
- Are Emergent Abilities in Large Language Models just In-Context Learning?
- Your LLM, Your Style: Behavioral Mode Axes for LLM Behavioral Control
- The Illusion of Debiasing: Persona Steering Redistributes Rather Than Reduces Bias in LLMs
- PersLLM: A Personified Training Approach for Large Language Models
- Persona Vectors: Monitoring and Controlling Character Traits in Language Models
- Consistency Training Helps Stop Sycophancy and Jailbreaks
- Useful Memories Become Faulty When Continuously Updated by LLMs
- AgentFly: Fine-tuning LLM Agents without Fine-tuning LLMs