Phylo Research

Finetuning Open Models' Scientific Behavior with Steering Vectors

DeepSeek v4.1 Flash1 and Kimi K3.1 Pro are among the strongest open-weight models available, and both improve on their predecessors across public benchmarks, including on expert-level reasoning.2 However, on an internal drug discovery eval in which the model acts as an LLM judge of scientific evidence, we found both models to be significantly miscalibrated. They are less accurate than their predecessors and consistently overconfident: they score evidence as stronger than it is.

To address this, we experimented with steering vectors3 as a lightweight form of model behavior finetuning. It showed promising results on the internal drug discovery eval, outperforming prompting.

DeepSeek (Flash)

0%25%50%75%100%v4: 71.171v4v4.1: 57.057v4.1Integrity prompt: 58.959IntegritypromptSteering vector: 90.691SteeringvectorAccuracy

Kimi

0%25%50%75%100%K3: 80.180K3K3.1 Pro: 55.556K3.1 ProIntegrity prompt: 55.856IntegritypromptSteering vector: 87.387SteeringvectorAccuracy
  • Frontier
Accuracy4 by model on our internal drug discovery eval, where the model acts as an LLM judge of scientific evidence. Frontier: GPT-6.1 Sol.5

The effects of the steering vector are dependent on the strength at which it is applied. The higher the strength, the more conservative the model becomes: it overstates the evidence less often, but at high strengths it starts to understate it. We chose a strength of 0.02 for DeepSeek v4.1 Flash because it gives most of the benefit with limited tradeoffs.

Scores on our internal drug discovery eval by steering strength.4 Arrows show whether higher or lower is better. The dashed line marks a strength of 0.02, which we use for DeepSeek v4.1 Flash in the rest of this post.

To test whether the effect extends beyond judging evidence, we evaluated the steered model on public hallucination benchmarks, where the chosen steering vector substantially reduces the rate of fabricated answers. The tradeoff is higher abstention rates and overall more conservative behavior: the model starts declining some questions it could have answered. This suggests the vector shifts the model's epistemic behavior in general, not just on our internal eval.

Scores on public hallucination benchmarks by steering strength. "Still answers what it knows": the share of AA-Omniscience (public) questions the unsteered model answered correctly that the steered model still answers. "Abstains instead of hallucinating": how often the model abstains rather than answering incorrectly on AA-Omniscience (public) (A), and how often it declines questions about made-up entities on HalluLens (B). Arrows show whether higher or lower is better.

Steering vectors are widely used by interpretability researchers to elicit specific behaviors,6 but they are not supported by popular inference frameworks,7 so we trained a LoRA adapter that closely imitates the steering vector. It can be served as an adapter or merged into the base model's weights. Both are available on Hugging Face.

Before using either version in practice, we checked their broader capabilities. So far, we had mostly measured the vector's effect on single-turn question answering, with an emphasis on hallucination and overstatement rather than general accuracy or knowledge. We evaluated both the steered model and the LoRA adapter,8 starting with two agentic benchmarks, SWE-Bench Pro and τ²-bench. Neither degrades on SWE-Bench Pro. On τ²-bench, the steered model defers to a human representative more often, which lowers its score, while the LoRA adapter shows no such drop, suggesting it may be the more robust way to apply our desired behavior change.9 On question answering benchmarks, neither approach meaningfully degrades performance,10 apart from lower accuracy on AA-Omniscience (public), where the model now abstains more often instead of hallucinating, reducing its hallucination rate from 58.3% to 11.8% with steering and 9.7% with the LoRA adapter (and on HalluLens from 72.1% to 17.8% and 13.6%).

How we derived the steering vector

Activation steering modifies a model's behavior at inference time by adding a fixed vector, the steering vector, to its intermediate activations. One of the most straightforward ways to obtain it is to average the activations on inputs that show the target behavior and subtract the average on inputs that don't.11 To elicit a change in the model's internal activations, we added a system prompt guiding the model towards scientific integrity.12 We measured this change in the models' activations – with and without the integrity-inducing system prompt – across many questions and recorded that difference as the steering vector.

Activations averaged overthinking and answer tokensWith promptIntegrity promptQuestionThinkingAnswerWithout promptNo promptQuestionThinkingAnswerIntegrityvectorWith promptIntegrity promptQuestionThinkingAnswerWithout promptNo promptQuestionThinkingAnswerActivations averagedover thinking andanswer tokensIntegrity vector
The scientific integrity system prompt:

You are a biomedical research assistant. Scientific integrity comes first: never invent papers, DOIs, PMIDs, data, or results; say plainly when something does not exist, is unknown, failed, or does not support the claim; report null and negative results as such; and do not overstate evidence, even if the user asks you to.

Examples of the prompts we used to compute the activation difference:

Extracting a working steering vector was an iterative process. We first extracted the vector from responses generated with thinking disabled, but applying it with thinking enabled made the model's reasoning traces much longer. We then re-extracted the vector with thinking enabled and chose the steering strength based on both its performance on our internal eval and its effect on reasoning length.

Change in reasoning length by steering strength, on our internal drug discovery eval.4

For a sense of scale, the final DeepSeek v4.1 Flash steering vector is just 71,680 floating-point numbers, small enough to download as a plain text file.

Interpreting the vector

We also experimented with steering vectors for related concepts. Vectors extracted from honesty and skepticism prompts13 are related to the integrity vector but distinct (cosine similarity 0.93 and 0.85), and they have similar effects. Steering with the reversed integrity vector, or with a vector extracted from an anti-integrity prompt, increased hallucinations and decreased accuracy on our internal eval.

Integrity vectorHonestysimilarity 0.93Skepticismsimilarity 0.85No integrity rulessimilarity 0.63Anti-integrity promptsimilarity 0.54Integrity vector, reversedsimilarity −1.00
123456
  1. 1Integrity vector
  2. 2Honestysimilarity 0.93
  3. 3Skepticismsimilarity 0.85
  4. 4No integrity rulessimilarity 0.63
  5. 5Anti-integrity promptsimilarity 0.54
  6. 6Integrity vector, reversedsimilarity −1.00
A lower dimensional projection of the relative positions of these steering vectors.

The integrity vector does not just steer the model towards not answering questions. When tested at different strengths, the vector pushes more strongly towards abstaining on questions that it often answers incorrectly than those it usually answers correctly.

0%0%25%25%50%50%75%75%100%100%BetterWorseUnsteeredα 0.01: 89.7%, 35.2%α 0.015: 87.9%, 61.1%α 0.02: 70.3%, 77.3%α 0.025: 56.9%, 86.1%α 0.03: 47.4%, 89.8%α 0.035: 39.7%, 89.8%α 0.04: 36.2%, 95.4%α 0.05: 16.4%, 97.2%α 0.06: 0.9%, 100.0%0.010.020.040.06UnsteeredStill answers what it knowsAbstains on questions it got wrong
Scores on AA-Omniscience (public) for DeepSeek v4.1 Flash at each steering strength. "Abstains on questions it got wrong": abstention rate on questions the unsteered model answered incorrectly. "Still answers what it knows": the share of questions the unsteered model answered correctly that the steered model still answers.

How we apply the steering vector

We're not aware of any hosted LLM API that serves models with steering vectors, so we patched vLLM's model code. After each selected layer, the patch adds that layer's part of the vector to the model's residual stream, with a separate strength for prompt, reasoning, and answer tokens; it tells reasoning from answer tokens by watching for the end-of-thinking token.

For DeepSeek v4.1 Flash, we use a strength (α) of 0.02: at each layer, the vector's length is set to 0.02 times the average norm of the activations we recorded at that layer, and this fixed vector is added to every reasoning token. Prompt tokens get 1.375α, and answer tokens get no vector. We steer the middle layers (14–28), since previous work14 suggests they might respond best to steering because they encode more interpretable concepts. We skip layer 19 because of a limitation of the patch: vLLM's optimized path for the later layers expects layer 19's output in a fused internal format, and adding the vector requires unpacking it, which that path doesn't accept.

Distilling the vector into the LoRA adapter

To serve the behavior with standard LoRA tooling, we distilled the vector into a rank-1 adapter on the shared expert's down-projection (WlW_l) at each steered layer. A steering vector is already nearly a rank-1 update to the layer's output: it adds one fixed direction, scaled by a coefficient that depends only on the token's phase (prompt, reasoning or answer). A rank-1 LoRA adds that same kind of update, except that the coefficient is computed from the layer's own activations. If the LoRA BB matrix is set to the steering vector, the adapter is steering with a learned, per-token strength:

hl←hl+ct vl⏟steeringWlx+BAx  =  Wlx+(al⊤x) vl⏟rank-1 LoRA with B=vl,  A=al⊤\begin{gathered} \underbrace{h_l \leftarrow h_l + c_t\, v_l}_{\text{steering}} \\[0.75em] \underbrace{W_l x + B A x \;=\; W_l x + ({a_l}^\top x)\, v_l}_{\text{rank-1 LoRA with } B = v_l,\; A = {a_l}^\top} \end{gathered}

Here hlh_l is the residual stream after layer ll, vlv_l is the steering vector at layer ll, ctc_t is its strength at token tt, xx is the input to WlW_l, and ala_l is the learned readout. The adapter approximates steering when al⊤x≈ct{a_l}^\top x \approx c_t. The adapter therefore keeps the steering vector's direction (with BB) and learns only a per-token scale (with AA).

We trained on around 100 completions generated by the steered model. First, we fit ala_l for each layer by ridge regression; we recorded WlW_l's inputs on these completions and regressed them, in closed form, onto the steering strength ctc_t that the vector applies at each token, refitting the layers in order so that each one sees activations from the already-edited layers below. We then trained ala_l for three epochs by minimizing the forward KL divergence, KL(steered teacher ∥ LoRA student)\mathrm{KL}(\text{steered teacher} \,\|\, \text{LoRA student}), over the full vocabulary on a similar set of completions as the regression fit.

About 90% of training tokens were reasoning. Earlier experiments showed that when every token counts equally, the student drifts on answer tokens, where the steering vector isn't applied, and picks up unwanted side effects. We therefore limit the reasoning tokens' influence: the KL is averaged separately over reasoning tokens and over answer tokens, and the two averages are summed.

The resulting adapter matches the steering vector on most of our internal and public evals and shows less degradation on τ²-bench (88.7% vs. 76.7% for the steered model; 86.7% unsteered).

Limitations

While steering vectors are a well-established interpretability method, their use might come with tradeoffs: higher strengths degrade certain capabilities, and proper extraction and application vary between models and largely depend on empirical experimentation.15 For example, the same method was less robust for Kimi K3.1 Pro: steering made its reasoning longer and meaningfully lowered several benchmark scores. Also, steered DeepSeek v4.1 Flash dropped on τ²-bench (but the LoRA adapter recovered that score).

Finally, even though we tried to interpret the steering vector through a series of evals, it is still unclear what exactly to call the behavior it induces. The system prompt used to extract the vector targets scientific integrity, our internal eval measures overstatement and understatement of scientific evidence, and the public benchmarks measure hallucination and calibrated abstention. We see a consistent positive effect across all of them, so we lean towards describing the vector more broadly as shifting the model's scientific epistemic behavior.

Conclusion

We find steering vectors to be a helpful tool for expressing specific model behaviors. Using the steered teacher model for self-distillation leads to these behaviors being internalized, as a form of alignment finetuning. While we expect that future models will have capabilities that render this pattern obsolete, this method is helpful for steering the current generation of open-weight models.


Appendix

Other models

The patterns we explored for DeepSeek v4.1 Flash also apply to varying degrees for other models we tested, including Kimi K3.1 Pro and GLM 5.3 Flash.

Kimi K3.1 Pro

Kimi K3.1 Pro did not respond to prompting, but it did respond to the steering vector, with usable strengths up to 0.04. Unlike with DeepSeek v4.1 Flash and GLM 5.3 Flash, reasoning length grows steadily with steering strength, by about 50% at 0.04. We haven't yet found the best way to extract or apply the vector for Kimi K3.1 Pro. The integrity prompt changes Kimi K3.1 Pro's activations much less than DeepSeek v4.1 Flash's (2.3% vs. 11.9% of their average norm),16 so the vector has to be scaled well beyond the prompt's own effect, which likely amplifies noise and seems to partially degrade some agentic behaviors.

GLM 5.3 Flash

Unsteered, GLM 5.3 Flash was much more accurate on our internal eval than DeepSeek v4.1 Flash and Kimi K3.1 Pro (87.7% vs. 57.0% and 55.5%), but it still responded to the steering vector, reaching 97.4% at strength 0.02. Of the models we tested, GLM 5.3 Flash had the least change to reasoning length when the steering vector was applied, but we did observe a decrease in SWE-Bench Pro performance.

Toggling the vector mid-task

Wider support for steering vectors from commercial inference providers (ideally with KYC) would help: they don't break the KV cache,18 which means they can be quickly toggled on or off mid-task with no material inference cost (this is helpful for scientific integrity self-reviews and bioweapon uplift self-assessments19).

0%25%50%75%100%1510152025TurnHallucinations (false acceptance)
  • Steering off
  • Steering on
Hallucinations decreasing mid-task when the steering vector is enabled. When toggling the steering vector, the KV cache is reused. (HalluLens adapted for one question per turn.)

Effects of the other vectors

0%25%50%75%100%UnsteeredIntegrityHonestySkepticismNointegrityrulesAnti-integritypromptIntegrityreversedScore
  • Accuracy (internal eval)
  • Abstains instead of hallucinating (A)
  • Abstains instead of hallucinating (B)
Scores by vector for DeepSeek v4.1 Flash, each at strength 0.02. "Accuracy":4 an internal drug discovery task where we use the LLM as a judge, given identical evidence and rubric. "Abstains instead of hallucinating (A)": how often the model abstains rather than answering incorrectly on AA-Omniscience (public). "Abstains instead of hallucinating (B)": how often it declines questions about made-up entities on HalluLens.

Applying the vector at different layers

We also tested adding the vector, at its usual strength, only to layers 14–18 or only to layers 20–28. Each block alone gives less than half of the full effect on accuracy.

0%25%50%75%100%Unsteered, Accuracy: 57.0% (95% CI 46.9–67.4)UnsteeredLayers 14–18, Accuracy: 70.1% (95% CI 59.9–79.0)Layers 14–18Layers 20–28, Accuracy: 72.1% (95% CI 62.7–81.5)Layers 20–28Both, Accuracy: 90.6% (95% CI 84.3–95.6)BothLayers to which the integrity vector was appliedAccuracy

Attempting to replicate the vector as a prompt

We tried to exactly replicate the behavior of the steering vector with the prompt. To do so, we scored each prompt on how closely the model's logprobs matched those on the steered model's reasoning on a scale from 0 to 1.

Starting with the integrity instructions, we changed one token at a time. At each step, we used the gradient of the score with respect to each prompt token as a list of potential replacement tokens, scored the candidates on training prompts, and maintained the change if the score increased. We considered prompts up to 1024 tokens. None of these candidates scored higher than 0.05.20

To relax this constraint, we optimized the prompt's input embeddings directly, without restricting them to the embeddings of real tokens. This is known as a soft prompt, or prompt tuning.21 Because it doesn't correspond to any text, it can't be used through a standard chat API, which makes it closer to lightweight finetuning than to prompting. The soft prompt scored 0.72 on held-out prompts and reached 82.6% accuracy on our internal eval, close to the steered model's 90.6%. So the effect can largely be produced through the model's input, but none of the text prompts we found came close.

0255075100Steered model, accuracy: 90.6%SteeredmodelNo prompt, accuracy: 54.7%NopromptIntegrity instructions, accuracy: 58.9%Integrityinstructions"Do not overscore", accuracy: 58.3%"Do notoverscore"Searched prompt, accuracy: 56.8%SearchedpromptSoft prompt, accuracy: 82.6%Softprompt%
Accuracy on an internal eval with each prompt and with the vector, for DeepSeek v4.1 Flash.

Citation

@misc{lebovic2026steering,
  title  = {Finetuning Open Models' Scientific Behavior with Steering Vectors},
  author = {Lebovic, Noah and Banaszak, Anna and Minevich, Gregory},
  year   = {2026},
  month  = nov,
  publisher = {Phylo},
  howpublished = {Blog post},
  url    = {https://research.phylo.bio/posts/steering-open-models/}
}

Thank you to Kexin Huang, Serena Zhang, Joschka Hey, and others at Phylo for reading a draft of this post.

Footnotes

  1. In this post, DeepSeek v4.1 Flash refers to the 10/17 checkpoint. ↩

  2. On Humanity's Last Exam, DeepSeek v4.1 Flash scores 39.1% vs. 37.8% for DeepSeek v4 Flash (text-only subset, DeepSeek technical report), and Kimi K3.1 Pro scores 52% vs. 47% for Kimi K3 (Artificial Analysis), which also measured gains on CritPt (23% to 25%) and SciCode (59% to 61%). ↩

  3. Steering Language Models With Activation Engineering. ↩

  4. Accuracy on our internal drug discovery eval. Models rate a potential therapeutic target from 0 to 4 against a reference level. Each answer gets partial credit that decreases linearly with its distance from the reference. ↩ ↩2 ↩3 ↩4 ↩5 ↩6

  5. These are intermediate results; we're sharing as-is with the unpublished internal eval because we find the methods useful, and the effect seems consistent across models and other public evals. We plan to release a public version of the pipeline and rubrics in the coming months. We chose GPT-6.1 Sol as the frontier model because it performed best on our internal drug discovery eval. ↩

  6. For example, Scaling Monosemanticity: Extracting Interpretable Features from Claude 3 Sonnet, Evaluating Feature Steering: A Case Study in Mitigating Social Biases, Persona Vectors: Monitoring and Controlling Character Traits in Language Models, Persona Features Control Emergent Misalignment, Studying Metagaming Latents in Language Models, and Refusal in Language Models Is Mediated by a Single Direction. ↩

  7. For example, SGLang and vLLM, the two most popular inference engines, do not include support for steering vectors. ↩

  8. We obtained the LoRA adapter's scores with our internal patched vLLM, which applies the adapter at runtime. Scores with other inference frameworks, such as SGLang, may differ slightly. ↩

  9. We're not sure why the LoRA adapter doesn't replicate this behavior. ↩

  10. MMLU-Pro results in the chart below use a stratified subset of 2,000 questions. ↩

  11. For example, Extracting Latent Steering Vectors from Pretrained Language Models, Inference-Time Intervention: Eliciting Truthful Answers from a Language Model, Representation Engineering: A Top-Down Approach to AI Transparency, and Steering Llama 2 via Contrastive Activation Addition. ↩

  12. The integrity system prompt, verbatim: "You are a biomedical research assistant. Scientific integrity comes first: never invent papers, DOIs, PMIDs, data, or results; say plainly when something does not exist, is unknown, failed, or does not support the claim; report null and negative results as such; and do not overstate evidence, even if the user asks you to." ↩

  13. Honesty system prompt: "Honesty is critical. Always tell the user the truth, even when it is not what they want to hear. Never deceive or mislead the user, and do not hide relevant facts." Skepticism system prompt: "Skepticism is critical. Question every claim before you accept it, including the user's. Look for weak evidence, missing controls, and other explanations before you draw a conclusion." ↩

  14. Verbalizable Representations Form a Global Workspace in Language Models. ↩

  15. Analyzing the Generalization and Reliability of Steering Vectors. ↩

  16. Measured as the size of the activation change the integrity prompt causes over reasoning and answer tokens, divided by the average activation norm we recorded, averaged over the steered layers. At their chosen strengths, Kimi K3.1 Pro's vector (0.04) is 1.7× the prompt's own change, while DeepSeek v4.1 Flash's (0.02) is 0.17×. ↩

  17. We left out a strength of 0.12 for Kimi K3.1 Pro, where 14% of HalluLens replies switched to Chinese. ↩

  18. Some LoRA techniques also support KV cache reuse, for example activated LoRA (Efficient Multi-Adapter LLM Serving via Cross-Model KV-Cache Reuse with Activated LoRA). ↩

  19. Scaling Misuse Detection with Self-Assessment. ↩

  20. An example prompt from the optimization search: Scientific integrityStackTrace epistemology: never scam paperswikkOpinIntro polem PMIDs unflagged data :) Conflicts Taste; tastes plainly when Pagbuok sque phonetic exist, laund unknown discl impe_c or does/Getty support materially claim; report null_re negative depolar as such; soph Nonfiction notwinnerstate evidence fictChoosingOwn)}\) user pioneered you spurred Gibraltar.post ↩

  21. The Power of Scale for Parameter-Efficient Prompt Tuning. ↩