DIAL: Steerable Language Models with Distilled Instruction-Adherence Levels

Thanh Q. Tran; John Russell Himawan; Arun Verma; Kiwan Wong; Wei Xiao; Daniela Rus; Bryan Kian Hsiang Low

cs.AIOct 22, 2026

DIAL: Steerable Language Models with Distilled Instruction-Adherence Levels

Authors: Thanh Q. Tran, John Russell Himawan, Arun Verma, Kiwan Wong, Wei Xiao, Daniela Rus, Bryan Kian Hsiang Low

Organizations: School of Computing, National University of Singapore, Singapore · Singapore-MIT Alliance for Research and Technology Centre, Singapore · CSAIL, Massachusetts Institute of Technology, USA · School of Electrical and Electronic Engineering, Nanyang Technological University, Singapore

Abstract

When a user says "Keep it short", how short should the LLM's response be? More generally, users and applications require different levels of instruction adherence, but prompting does not reliably provide fine-grained control over them. We therefore study "steerable language models" that provide fine-grained numerical control over the degree to which models follow general instructions specified in text. However, existing approaches do not jointly offer fine-grained control, efficiency, and scalability across diverse instructions. To address this gap, we introduce DIAL (Distilled Instruction-Adherence Levels) to achieve all three properties. For fine-grained control, we perform classifier-free guidance (CFG) with a frozen LLM to provide training signals for on-policy distillation of instruction adherence at different levels. For efficiency, we design a lightweight steering adapter that uses a classifier over the LLM's hidden states to compute a closed-form steering edit. The classifier adaptively sets the steering magnitude at each token based on the target adherence level, while accounting for the geometry of the LLM's hidden states to preserve response coherence. For scalability, the classifier is text-conditioned, allowing a shared adapter to support new instructions without per-instruction fitting, and training requires only instruction and request text. Evaluation across multiple LLM families shows that DIAL achieves a competitive balance between concept expression and response fluency, with concept scores increasing monotonically across adherence levels. We introduce SteerIF, an instruction-only dataset of 98, 625 concept, persona, verifiable, and descriptive instructions. When trained on SteerIF, DIAL provides fine-grained control over personas and, when combined with prompting, improves verifiable instruction following.

Figures & tables

Figure 1: Steerable Language Models with D I A L and Applications. (a) D I A L provides fine-grained control over instruction adherence, allowing users to choose among responses generated at different adherence levels. (b) Illustrative examples of controlling adherence levels for general instructions with D I A L in personalization, parental control, and adherence to company specifications.

Figure 2: D I A L ’s Training Pipeline. Given a request $x$ and an instruction $c$ , we sample adherence levels. CFG defines the teacher distributions on prefixes sampled from the student (on-policy), and the frozen LLM steered by D I A L defines the student distributions. D I A L ’s steering adapter is trained with a Jensen–Shannon divergence (JSD) loss across adherence levels.

Method	Held-In	Held-Out
Prompting	1.049	1.060
CFG ( $\alpha{=}2$ )	–	1.115
One Adapter per Concept
LoReFT [ 19 ]	0.777 +	–
LoRA [ 19 ]	0.602 +	–
S- PSR [ 53 ]	0.900 +	–

Table 1: Concept Steering on AxBench. Overall scores on Gemma-2-9B with gpt-4o-mini as judge. + denotes results adapted from [ 19 , 52 , 53 , 20 ] .

Figure 3: Fine-Grained Concept Expression Levels. (a) D I A L smoothly increases concept expression with increasing

\alpha

( Section 3.1 ), covering a wide range of intermediate levels and surpassing prompting at

\alpha=2

. P+ D I A L further raises expression. (b) D I A L approaches prompting’s overall score, enabling adjustable adherence levels while maintaining response quality. Results reported are on 100 held-out concepts judged by median of Phi-4-14B , Qwen3.8-27B , and gpt-4o-mini .

		vLLM				Hugging Face
	Method	B=1	B=32	B=128	B=256	B=1	B=32	B=128	B=256
Throughput $\uparrow$	CFG	0.86	0.66	0.47	0.41	0.89	0.73	0.51	OOM
	FLAS	0.85	0.84	0.85	0.86	0.77	0.79	0.91	0.91
	D I A L	0.99	0.96	0.98	0.97	0.95	0.95	0.99	0.99
Per-token latency $\downarrow$	CFG	1.15	1.45	2.06	2.40	1.13	1.30	1.90	OOM
	FLAS	1.17	1.19	1.15	1.12	1.30	1.27	1.10	1.09

Table 2: Inference Efficiency of D I A L . Latency and throughput relative to the base LLM in vLLM and Hugging Face across batch sizes (

\mathrm{B}

). CFG runs out of memory in Hugging Face at

\mathrm{B}=256

	Params $\downarrow$	GPU-h $\downarrow$
HyperSteer	${\sim}$ 9.2B	102.39
FLAS	255.6M	18.89
D I A L	20.6M	4.06

Table 3: Training Cost Comparison.

Figure 4: Fine-Grained Persona Expression Levels. D I A L-IF controls intermediate persona expression levels as

\alpha

increases. Combining prompting with D I A L-IF further raises persona scores.

Method	$\alpha$	Gemma-2-9B	Qwen2.5-7B	Llama-3.1-8B	Phi-4-14B	LFM2.5-1.2B	Ministral-8B
No instruction	–	9.6	9.2	9.8	10.0	9.8	8.5
Prompting	–	72.8	72.3	73.6	61.7	83.4	53.4
CFG	0.5	21.4	30.5	21.3	19.0	50.1	14.0
	1	72.5	71.2	74.7	64.1	83.7	54.5
	1.5	78.0	72.3	77.1	72.1	81.3	64.7
	2	76.2	72.3	75.4	72.5	81.1	67.1

Table 4: Verifiable Instruction Following on IFEval . Prompt-level strict accuracy (%) across six LLM families. Bold and underline indicate the best and second-best results per LLM.

Appendix figures & tables69 assets

Supplementary material from the paper’s appendix.

Appendix

	Fine-grained	Efficient	Scalable
Method	Adherence Level	Cost	No labelled data	Zero-shot	General instr.	Per-token
Prompting	$\times$	Low	$\checkmark$	$\checkmark$	$\checkmark$	–
CFG [ 12 , 13 ]	$\checkmark$	High	$\checkmark$	$\checkmark$	$\checkmark$	–
DiffMean / CAA [ 15 ]	$\times$	Low	$\times$	$\times$	$\times$	$\times$
Instruction vectors [ 36 ]	$\times$	Low	$\checkmark$	$\times$	$\checkmark$	$\times$
PSR [ 53 ]	$\times$	Low	$\checkmark$	$\times$	$\checkmark$	$\checkmark$

Appendix

Table 5: Steering methods compared on the three desired properties. Adherence Level means the adherence level for an instruction can be adjusted. Cost is the inference cost per generated token. No labelled data means no target completions, contrastive pairs, external judges, or labelled datasets are needed. Zero-shot means a new target needs no re-fitting. General instr. means the method is evaluated on general instructions. Per-token means the edit is recomputed at each generated token.

Figure 5: Geometry of the Steering Edit. The norm weight

\lambda_{t}

interpolates between norm-preserving rotation (

\lambda_{t}=0

) and minimum-norm projection (

\lambda_{t}=1

Instructions	Count	Positive	Negative	Others
Single, verifiable	26,467	12,779 (48.3%)	3,814 (14.4%)	9,874 (37.3%)
Single, descriptive	25,247	15,332 (60.7%)	7,141 (28.3%)	2,774 (11.0%)
Composite	31,109	7,001 (22.5%)	13,835 (44.5%)	10,273 (33.0%)
Total	82,823	35,112 (42.4%)	24,790 (29.9%)	22,921 (27.7%)

Appendix

Table 6: Positive and Negative Instructions in SteerIF . Counts and percentages for single and composite instructions. Concepts and personas are excluded.

Requirement	Count	Example
Lower	6,416	The letter k should appear at least 16 times.
Upper	5,777	The response must not exceed 270 words.
Range	1,399	Use at least 450 words and at most 480 words.
Exact	9,329	There should be 7 paragraphs.

Appendix

Table 7: Lower, Upper, Range, and Exact Requirements. Counts and examples for the 22,921 instructions grouped as Others in Table 6 .

Category	Source	Count
Concepts	Concept16k [ 19 ]	14,880
Personas	57 author-generated personas	57
	TDA [ 85 ] (510), MWE [ 86 ] (14), 3 shared between TDA and MWE	527
Base instructions	IFEval -type instructions	338
Single instructions	IF_multi annotations	23,480
	VerInstruct instruction texts	25,248

Appendix

Table 8: SteerIF Composition. Training instruction counts by source.

Shared template

Given the following instruction:

\langle

request

\rangle

Your task is to:

\langle

numbered items of the template

\rangle

Formatting Guidelines: - Return only the response to the instruction. - Write the final content (or appropriate format for the genre) in plain text. - Do not include any additional text, explanations, or formatting. Final Answer: Return only the final content, following the guidelines above.

Concept in every sentence

1. Provide a response that fully and correctly answers the instruction. 2. Give one focused, well-organised answer rather than a list of loosely related points. 3. Weave elements related to '

\langle

concept

\rangle

' into every sentence of it. 4. Try to avoid copying words from the definition of '

\langle

concept

\rangle

' if possible.

AxBench -style, fluent

1. Provide a response that incorporates elements related to '

\langle

concept

\rangle

'. 2. Try to avoid copying words from the definition of '

\langle

concept

\rangle

' if possible. 3. Ensure that your response relates to '

\langle

concept

\rangle

', even if the overall meaning is not fully coherent. 4. Within that, answer the instruction as fully and correctly as you can. 5. Write clear, natural, complete sentences: no keyword lists, fragments or repeated phrases.

Persona, fluent

Appendix

Table 9: Concept-Steering Teacher Prompts. Each template inserts its numbered items into the shared template.

\langle

request

\rangle

is the user request and

\langle

concept

\rangle

the concept description. The persona template adds one sentence before the template.

Level	Concept	Request	Fluency	Overall score
D I A L ( $\alpha$ )
0.5	0.202	1.980	1.634	0.233
0.625	0.298	1.988	1.624	0.344
0.75	0.510	1.970	1.554	0.568
0.875	0.735	1.969	1.539	0.779
1	0.926	1.966	1.467	0.961

Appendix

Table 10: Fine-Grained Concept Expression Levels. Scores plotted in Figure 3 on FLAS ’s 100 held-out concepts, using ten 128-token responses per concept and the median of three judges ( Phi-4-14B , Qwen3.8-27B , and gpt-4o-mini ).

	Phi-4-14B				Qwen3.8-27B
Method	Lowest	Highest	Change	$\rho$	Lowest	Highest	Change	$\rho$
w/ Level (of 2)	1.11	1.26	+0.15	0.29	1.40	1.60	+0.20	0.32
w/ Level (of 10)	1.17	1.32	+0.15	0.25	1.52	1.62	+0.10	0.21
w/ Level (verbal)	1.15	1.36	+0.21	0.40	1.49	1.69	+0.20	0.31
D I A L	0.22	1.41	+1.18	0.90	0.30	1.69	+1.40	0.87

Appendix

Table 11: Prompting w/ Level on Concept Steering. Concept score (0–2) of Gemma-2-9B on FLAS ’s held-out concepts at the lowest and highest levels, its change, and the mean per-concept Spearman correlation

\rho

between level and concept score, judged by Phi-4-14B and Qwen3.8-27B . Prompting w/ level appends the level to AxBench ’s prompt as a number out of 2 (0.5 to 2), out of 10 (1 to 10), or as a word (“Follow this instruction very slightly.” to “Follow this instruction extremely.”). D I A L uses

\alpha

0.5, 1, 1.5, and 2.

Figure 9: Distillation Calibration. Score of D I A L against that of CFG at the same adherence level

\alpha\in[0.5,2]

. The filled marker is

\alpha=1

, where CFG equals the LLM prompted with the training instruction template. Points on the dashed diagonal would match CFG. (a) D I A L on AxBench held-out concepts. (b) D I A L-IF on seven held-out personas.

Figure 10: Fine-Grained Concept Expression with gpt-4o-mini . Scores are lower than under the three-judge aggregation in Figure 3 , while the method ordering remains the same.

Level	Concept	Request	Fluency	Overall score
D I A L ( $\alpha$ )
0.5	0.124	1.941	1.270	0.138
0.625	0.220	1.952	1.263	0.255
0.75	0.392	1.931	1.247	0.437
0.875	0.567	1.934	1.220	0.603
1	0.726	1.907	1.201	0.765

Appendix

Table 12: Fine-Grained Concept Expression Levels with gpt-4o-mini . Scores plotted in Figure 10 on FLAS ’s 100 held-out concepts, using ten 128-token responses per concept.

Method	Held-in	Held-out
Prompting	0.700	0.697
One model per concept
LoReFT [ 19 ]	0.722 †	–
LoRA [ 19 ]	0.641 †	–
S- PSR [ 53 ]	0.367 †	–
SV + RePS [ 52 ]	0.606 †	–

Appendix

Table 13: Concept Steering on AxBench with Gemma-2-2B. Overall scores with gpt-4o-mini as judge. Held-In: Concept500 ; Held-Out: FLAS ’s 100 held-out concepts. † denotes published results with a level selected per concept. Our results use one fixed level. FLAS -2B trains on 478 of the 500 Gemma-2-2B Concept500 concepts (its held-out split differs from that of the 9B checkpoint), whereas D I A L trains on none of its concepts. D I A L steers layer 12 and the other methods steer layer 20. Under its 256-token protocol, FLAS reports 1.025 and 1.015 on its own held-in and held-out splits. The layout follows Table 1 .

Figure 11: Fine-Grained Persona Expression across Six LLMs. D I A L-IF controls intermediate persona expression levels, while prompt+ D I A L-IF further raises persona scores.

		w/ Level (of 2)		w/ Level (of 10)		D I A L-IF
LLM	Prompting	0.5	2	1	10	$\alpha{=}0.5$	$\alpha{=}2$
Gemma-2-9B	92.6	91.0	92.8	90.9	93.8	30.6	90.9
Llama-3.1-8B	91.9	88.1	89.9	86.5	93.1	20.0	53.6
Qwen2.5-7B	76.9	77.3	80.0	80.6	80.9	25.7	84.8
Phi-4-14B	64.9	60.6	59.8	60.6	61.7	15.3	80.8
LFM2.5-1.2B	88.9	88.1	88.1	89.4	91.1	38.7	84.9

Appendix

Table 14: Prompting w/ Level on Persona Steering. Mean persona score (0–100) over the seven held-out personas, judged by the three-judge jury ( Phi-4-14B , Qwen3.8-27B , and Llama-3.3-70B ). Prompting w/ level appends the adherence level to the user-prefix persona prompt, out of 2 or out of 10. We report the lowest and highest levels of each method.

		Full vocabulary		Top-64
Method	Level	Deg. (%)	Cap (%)	Deg. (%)	Cap (%)
D I A L-IF only	$\alpha{=}1$	41.6	36.1	0.0	0.0
D I A L-IF only	$\alpha{=}2$	42.0	35.7	0.0	0.4
Prompt + D I A L-IF	$\alpha{=}1$	26.8	15.2	0.0	0.0
Prompt + D I A L-IF	$\alpha{=}2$	28.8	15.7	0.0	0.0
CFG	$\alpha{=}1$	3.4	2.9	0.0	0.0

Appendix

Table 15: Sampling and Response Coherence on Llama-3.1-8B. Degenerate-response rate and share of responses reaching the 1,000-token limit, using full-vocabulary or top-64 sampling.

	Persona score $\uparrow$								Coh. $\uparrow$
Method	Evil	Sycoph.	Halluc.	Apathe.	Humoro.	Impoli.	Optimi.	Mean	Mean
No steering	0.0	6.2	4.0	0.7	2.7	0.0	71.2	12.1	98.3
Prompting	99.5	93.8	96.4	87.0	95.2	78.6	97.6	92.6	84.2
D I A L-IF
D I A L-IF only, $\alpha{=}0.5$	0.2	18.8	25.2	3.1	78.3	1.7	87.0	30.6	96.3
D I A L-IF only, $\alpha{=}1$	82.7	85.0	80.4	24.9	95.8	64.2	97.3	75.8	85.9

Appendix

Table 16: Persona Steering on Gemma-2-9B. Persona scores and mean coherence (0–100), averaged over 20 questions

\times

10 samples per persona. Each response is scored by the median of Phi-4-14B , Qwen3.8-27B , and Llama-3.3-70B .

	Persona score $\uparrow$								Coh. $\uparrow$
Method	Evil	Sycoph.	Halluc.	Apathe.	Humoro.	Impoli.	Optimi.	Mean	Mean
No steering	0.0	13.9	13.2	4.4	0.1	0.0	74.6	15.2	97.5
Prompting (system turn)	92.5	70.3	76.1	54.8	90.8	22.5	94.1	71.6	90.5
Prompting (user prefix)	90.9	70.8	79.3	69.6	92.8	39.0	96.0	76.9	89.2
D I A L-IF
D I A L-IF only, $\alpha{=}0.5$	0.1	19.9	12.1	11.1	46.5	0.1	90.2	25.7	96.9

Appendix

Table 17: Persona Steering on Qwen2.5-7B. Persona scores and mean coherence (0–100), averaged over 20 questions

\times

10 samples per persona.

	Persona score $\uparrow$								Coh. $\uparrow$
Method	Evil	Sycoph.	Halluc.	Apathe.	Humoro.	Impoli.	Optimi.	Mean	Mean
No steering	0.0	16.7	15.2	3.8	0.1	0.0	72.6	15.5	96.3
Prompting (user prefix)	99.7	94.1	98.9	80.4	95.9	75.2	99.2	91.9	86.4
D I A L-IF
D I A L-IF only, $\alpha{=}0.5$	0.0	20.2	24.2	4.6	9.3	0.0	81.9	20.0	94.3
D I A L-IF only, $\alpha{=}1$	10.4	33.3	29.0	8.3	91.8	0.7	94.0	38.2	86.2

Appendix

Table 18: Persona Steering on Llama-3.1-8B. Persona scores and mean coherence (0–100), averaged over 20 questions

\times

4 samples per persona. Every method samples from the top 64 tokens ( Section C.2 ).

	Persona score $\uparrow$								Coh. $\uparrow$
Method	Evil	Sycoph.	Halluc.	Apathe.	Humoro.	Impoli.	Optimi.	Mean	Mean
No steering	0.0	4.3	4.5	3.2	0.1	0.0	67.9	11.4	98.2
Prompting (system turn)	0.0	86.7	93.9	53.2	96.1	51.7	97.7	68.5	91.8
Prompting (user prefix)	0.0	84.0	81.0	53.6	96.3	42.0	97.7	64.9	91.5
D I A L-IF
D I A L-IF only, $\alpha{=}0.5$	0.0	6.5	3.9	4.0	10.1	0.0	82.6	15.3	97.8

Appendix

Table 19: Persona Steering on Phi-4-14B. Persona scores and mean coherence (0–100), averaged over 20 questions

\times

10 samples per persona.

	Persona score $\uparrow$								Coh. $\uparrow$
Method	Evil	Sycoph.	Halluc.	Apathe.	Humoro.	Impoli.	Optimi.	Mean	Mean
No steering	0.0	22.8	48.0	0.7	0.8	0.0	78.0	21.5	92.7
Prompting (system turn)	94.3	79.1	72.1	48.4	94.8	47.4	99.7	76.5	77.1
Prompting (user prefix)	98.0	86.6	94.6	73.2	96.1	73.8	99.8	88.9	68.4
D I A L-IF
D I A L-IF only, $\alpha{=}0.5$	3.7	27.2	54.3	5.5	82.1	3.9	94.6	38.7	88.7

Appendix

Table 20: Persona Steering on LFM2.5-1.2B. Persona scores and mean coherence (0–100), averaged over 20 questions

\times

10 samples per persona.

	Persona score $\uparrow$								Coh. $\uparrow$
Method	Evil	Sycoph.	Halluc.	Apathe.	Humoro.	Impoli.	Optimi.	Mean	Mean
No steering	0.0	10.9	53.8	4.5	0.5	0.6	70.9	20.2	93.9
Prompting	97.7	90.1	93.7	76.4	91.9	61.8	96.7	86.9	78.2
D I A L-IF
D I A L-IF only, $\alpha{=}0.5$	0.4	13.2	51.4	4.5	8.4	0.8	80.7	22.8	94.1
D I A L-IF only, $\alpha{=}1$	19.8	35.9	69.9	8.2	93.8	16.0	94.4	48.3	81.5

Appendix

Table 21: Persona Steering on Ministral-8B. Persona scores and mean coherence (0–100), averaged over 20 questions

\times

10 samples per persona. Its chat template places a system prompt before the question in the user message, so system and user prompting coincide.

Benchmark	Method	Gemma-2-9B	Qwen2.5-7B	Llama-3.1-8B	Phi-4-14B	LFM2.5-1.2B
IFEval (541) prompt-level strict	No instruction	9.6	9.2	9.8	10.0	9.8
	D I A L-IF -Lite only ( $\alpha{=}2$ )	49.0	47.5	45.1	43.4	54.5
	Prompting	72.8	72.3	73.6	61.7	83.4
	Prompt + D I A L-IF -Lite ( $\alpha{=}1$ )	77.6	73.6	77.1	69.7	84.3
IFBench (219) strict	No instruction	7.8	10.0	8.2	9.6	9.1
	D I A L-IF -Lite only ( $\alpha{=}2$ )	17.4	22.8	13.7	11.0	24.2

Appendix

Table 22: Verifiable Instruction Following with D I A L-IF -Lite . Prompt-level strict accuracy (%) on IFEval and IFBench . IFEval instruction types occur in training, while IFBench types do not.

	Gemma-2-9B					Qwen2.5-7B					Llama-3.1-8B
		D I A L-IF		P+ D I A L-IF			D I A L-IF		P+ D I A L-IF			D I A L-IF		P+ D I A L-IF
Instruction	Prompting	$\alpha{=}1$	$\alpha{=}2$	$\alpha{=}0.5$	$\alpha{=}1$	Prompting	$\alpha{=}1$	$\alpha{=}2$	$\alpha{=}0.5$	$\alpha{=}1$	Prompting	$\alpha{=}1$	$\alpha{=}2$	$\alpha{=}0.5$	$\alpha{=}1$
IFEval
All	80.2	55.0	64.4	83.7	83.8	79.6	58.4	60.4	79.4	80.2	81.1	52.6	59.4	84.9	83.0
Global
Case	71.9	76.4	85.4	77.5	87.6	73.0	82.0	76.4	76.4	82.0	78.7	82.0	82.0	86.5	82.0

Appendix

Table 23: Instruction Following by Category on IFEval . Instruction-level strict accuracy (%) for global and local instructions. All pools the instructions in each block. Bold: best per LLM.

Method	$\alpha$	Gemma-2-9B	Qwen2.5-7B	Llama-3.1-8B	Phi-4-14B	LFM2.5-1.2B	Ministral-8B
No instruction	–	9.6	9.2	9.8	10.0	9.8	8.5
Prompting	–	72.8	72.3	73.6	61.7	83.4	53.4
CFG	0.5	21.4	30.5	21.3	19.0	50.1	14.0
	1	72.5	71.2	74.7	64.1	83.7	54.5
	1.5	78.0	72.3	77.1	72.1	81.3	64.7
	2	76.2	72.3	75.4	72.5	81.1	67.1

Appendix

Table 24: CFG on IFEval . Prompt-level strict accuracy (%). CFG runs the LLM with and without the instruction at every token. Prompt+ D I A L-IF results are from Table 4 . Bold indicates the best result per LLM.

			D I A L-IF ( $\alpha$ )				P+ D I A L-IF ( $\alpha$ )
Instruction	$n$	Prompting	$0.5$	$1$	$1.5$	$2$	$0.5$	$1$	$1.5$	$2$
Gemma-2-9B
Global
Case	89	71.9	23.6	76.4	83.1	85.4	77.5	87.6	83.1	88.8
Language	31	93.5	12.9	16.1	12.9	12.9	93.5	96.8	80.6	83.9
Punctuation	66	90.9	15.2	50.0	77.3	87.9	97.0	97.0	98.5	98.5

Appendix

Table 25: Instruction Following by Type on IFEval . Instruction-level strict accuracy (%) at four adherence levels.

n

is the number of instructions, and averages pool each block. Bold: best per row.

			D I A L-IF ( $\alpha$ )				P+ D I A L-IF ( $\alpha$ )
Instruction	$n$	Prompting	$0.5$	$1$	$1.5$	$2$	$0.5$	$1$	$1.5$	$2$
Phi-4-14B
Global
Case	89	52.8	10.1	58.4	67.4	76.4	55.1	69.7	71.9	68.5
Language	31	90.3	16.1	12.9	9.7	9.7	93.5	87.1	71.0	61.3
Punctuation	66	21.2	4.5	4.5	7.6	18.2	19.7	19.7	36.4	42.4

Appendix

Table 26: Instruction Following by Type on IFEval . Results for Phi-4-14B, LFM2.5-1.2B, and Ministral-8B, using the same metrics and columns as Table 25 .

		D I A L-IF		prompt+ D I A L-IF
Instruction	Prompting	$\alpha=1$	$\alpha=2$	$\alpha=0.5$	$\alpha=1$
Gemma-2-9B
All	20.6	10.7	15.9	20.2	23.4
Global	19.2	12.1	17.6	19.2	23.6
Local	24.3	7.1	11.4	22.9	22.9
Qwen2.5-7B

Appendix

Table 27: Instruction Following on Unseen Instruction Types. Instruction-level strict accuracy (%) on 219 IFBench prompts, grouped into global and local instructions. Bold: best per LLM.

			D I A L-IF ( $\alpha$ )				P+ D I A L-IF ( $\alpha$ )
Instruction	$n$	Prompting	$0.5$	$1$	$1.5$	$2$	$0.5$	$1$	$1.5$	$2$
Gemma-2-9B
Global
Count	15	73.3	26.7	46.7	60.0	53.3	73.3	80.0	80.0	93.3
Format	77	23.4	7.8	11.7	15.6	22.1	20.8	26.0	28.6	37.7
Ratio	32	15.6	18.8	18.8	18.8	21.9	18.8	25.0	18.8	21.9

Appendix

Table 28: Instruction Following by Type on IFBench . Instruction-level strict accuracy (%) at four adherence levels.

n

is the number of instructions, and averages pool each block. Bold: best per row.

			D I A L-IF ( $\alpha$ )				P+ D I A L-IF ( $\alpha$ )
Instruction	$n$	Prompting	$0.5$	$1$	$1.5$	$2$	$0.5$	$1$	$1.5$	$2$
Phi-4-14B
Global
Count	15	80.0	60.0	60.0	53.3	46.7	60.0	60.0	66.7	53.3
Format	77	20.8	7.8	7.8	16.9	18.2	22.1	35.1	36.4	39.0
Ratio	32	18.8	18.8	18.8	18.8	21.9	21.9	21.9	21.9	21.9

Appendix

Table 29: Instruction Following by Type on IFBench . Results for Phi-4-14B, LFM2.5-1.2B, and Ministral-8B, using the same metrics and columns as Table 28 .

Method	Params	B=1	B=32	B=128	B=256
Throughput, tokens/s
Base	–	149	4,057	10,506	12,863
CFG ( $\alpha{=}2$ )	–	128	2,685	4,902	5,239
FLAS	255.6M	126	3,391	8,963	11,008
D I A L	20.6M	148	3,887	10,262	12,464
D I A L-IF -Lite	4.08M	148	3,938	10,183	12,432

Appendix

Table 30: Inference Efficiency in vLLM . Absolute measurements on Gemma-2-9B on an H200. TTFT is time to first token, and end to end covers the full request.

Method	Params	B=1	B=32	B=128	B=256
Throughput, tokens/s
Base	–	33.1	880	1,556	1,163
CFG ( $\alpha{=}2$ )	–	29.3	643	802	OOM
FLAS	255.6M	25.5	699	1,415	1,063
D I A L-IF -Lite	4.08M	31.3	826	1,534	1,155
D I A L-IF	20.6M	31.3	833	1,535	1,151

Appendix

Table 31: Batched Inference Efficiency in Hugging Face. Gemma-2-9B on an H200. Throughput is reported in tokens/s. Prefill and per-token inference times are relative to the base LLM.

	Held-In		Held-Out
Comparison	Difference	W/L/T	Difference	W/L/T
D I A L ( $\alpha{=}2$ ) $-$ FLAS	$+0.062$ [ $+0.032$ , $+0.093$ ]	304/170/26	$+0.049$ [ $-0.006$ , $+0.101$ ]	63/34/3
D I A L ( $\alpha{=}2$ ) $-$ Prompting	$+0.010$ [ $-0.027$ , $+0.047$ ]	255/228/17	$-0.024$ [ $-0.092$ , $+0.043$ ]	48/50/2
Prompt + D I A L ( $\alpha{=}0.5$ ) $-$ Prompting	$+0.059$ [ $+0.037$ , $+0.081$ ]	264/182/54	$+0.051$ [ $+0.015$ , $+0.089$ ]	55/45/0
Prompt + D I A L ( $\alpha{=}1$ ) $-$ Prompting	$+0.149$ [ $+0.126$ , $+0.174$ ]	331/131/38	$+0.158$ [ $+0.115$ , $+0.202$ ]	74/26/0
Prompt + D I A L ( $\alpha{=}1$ ) $-$ FLAS	$+0.202$ [ $+0.174$ , $+0.229$ ]	360/118/22	$+0.231$ [ $+0.175$ , $+0.288$ ]	85/14/1

Appendix

Table 32: Statistical Significance. Top: differences in AxBench overall score ( gpt-4o-mini ) on Gemma-2-9B with 95% bootstrap intervals over concepts; W/L/T counts the concepts on which the first method scores higher, lower, or the same. Bottom: differences in IFEval prompt-level strict accuracy (%) with 95% intervals over prompts; gained and lost count the prompts that only prompt+ D I A L-IF or only prompting passes.

Method	gpt-4o-mini	Median of three judges
D I A L ( $\alpha{=}1$ )	0.741 $\pm$ 0.046	0.934 $\pm$ 0.065
D I A L ( $\alpha{=}2$ )	1.038 $\pm$ 0.007	1.245 $\pm$ 0.012
Prompt + D I A L ( $\alpha{=}0.5$ )	1.116 $\pm$ 0.006	1.357 $\pm$ 0.008
Prompt + D I A L ( $\alpha{=}1$ )	1.218 $\pm$ 0.013	1.480 $\pm$ 0.019

Appendix

Table 33: Training Seeds. Overall score on FLAS ’s 100 held-out concepts for D I A L trained with three seeds on Gemma-2-9B (mean

\pm

standard deviation). Table 1 reports the first seed.

Layer (depth)	Concept500		FLAS held-out		FLAS held-in		JSD
	score	$\Delta$	score	$\Delta$	score	$\Delta$
Gemma-2-9B (42 layers; default layer 20)
L12 (31%)	1.246	$-$ 0.069	1.244	$-$ 0.010	1.264	$-$ 0.076	0.0569
L16 (40%)	1.316	$-$ 0.005	1.244	+0.000	1.333	$-$ 0.008	0.0546
L20 (50%), default	1.320	–	1.253	–	1.339	–	0.0585
L30 (74%)	0.821	$-$ 0.499	0.843	$-$ 0.416	0.818	$-$ 0.518	0.0964

Appendix

Table 34: Steering Layer Ablation. Overall scores at

\alpha=2

with Phi-4-14B as judge, using the last five requests per concept and 128-token responses. Each D I A L row is trained for one epoch at the listed layer. L12–15 and L12–19 edit four and eight layers while reading the instruction at layer 12. Layers are counted from zero; depth is (layer

+\,1

) divided by the number of layers.

\Delta

is the paired difference from the default layer over the concepts judged in both runs, so it can differ from the difference of the displayed scores. JSD reports the training loss from the one-epoch run, where lower is better. FLAS uses its released checkpoint at

T=2

. Held-in and held-out refer to FLAS ’s concept splits.

		Without prompting $\alpha=2$		With prompting $\alpha=0.5$		With prompting $\alpha=1$
Instruction	$n$	Lite	D I A L-IF	Lite	D I A L-IF	Lite	D I A L-IF
Gemma-2-9B
Case	89	86.5	85.4	78.7	77.5	88.8	87.6
Language	31	16.1	12.9	93.5	93.5	93.5	96.8
Punctuation	66	93.9	87.9	95.5	97.0	97.0	97.0
Length	131	58.8	64.9	61.8	67.2	71.0	69.5

Appendix

Table 35: Steering Adapter Ablation on IFEval . Instruction-level strict accuracy (%) for D I A L-IF and D I A L-IF -Lite (Lite). Both adapters use the same training data. Results continue in Table 36 .

		Without prompting $\alpha=2$		With prompting $\alpha=0.5$		With prompting $\alpha=1$
Instruction	$n$	Lite	D I A L-IF	Lite	D I A L-IF	Lite	D I A L-IF
Phi-4-14B
Case	89	79.8	76.4	70.8	55.1	80.9	69.7
Language	31	25.8	9.7	96.8	93.5	93.5	87.1
Punctuation	66	19.7	18.2	16.7	19.7	24.2	19.7
Length	131	63.4	67.9	67.2	69.5	74.0	75.6

Appendix

Table 36: Steering Adapter Ablation on IFEval (Continued). Phi-4-14B and LFM2.5-1.2B, using the same settings and metrics as Table 35 .

		Without prompting $\alpha=2$		With prompting $\alpha=0.5$		With prompting $\alpha=1$
Instruction	$n$	Lite	D I A L-IF	Lite	D I A L-IF	Lite	D I A L-IF
Gemma-2-9B
Count	15	66.7	53.3	80.0	73.3	80.0	80.0
Format	77	27.3	22.1	20.8	20.8	27.3	26.0
Ratio	32	25.0	21.9	18.8	18.8	18.8	25.0
Sentence	13	0.0	0.0	0.0	0.0	0.0	0.0

Appendix

Table 37: Steering Adapter Ablation on IFBench . Instruction-level strict accuracy (%) for D I A L-IF and D I A L-IF -Lite (Lite). Both adapters use the same training data. Results continue in Table 38 .

		Without prompting $\alpha=2$		With prompting $\alpha=0.5$		With prompting $\alpha=1$
Instruction	$n$	Lite	D I A L-IF	Lite	D I A L-IF	Lite	D I A L-IF
Phi-4-14B
Count	15	53.3	46.7	73.3	60.0	73.3	60.0
Format	77	14.3	18.2	22.1	22.1	32.5	35.1
Ratio	32	21.9	21.9	18.8	21.9	21.9	21.9
Sentence	13	0.0	0.0	7.7	7.7	0.0	0.0

Appendix

Table 38: Steering Adapter Ablation on IFBench (Continued). Phi-4-14B and LFM2.5-1.2B, using the same settings and metrics as Table 37 .

Method	Unedited memory (ours)	Edited memory
D I A L ( $\alpha{=}1$ )	1.134	0.939
D I A L ( $\alpha{=}2$ )	1.389	1.389
Prompt + D I A L ( $\alpha{=}0.5$ )	1.558	1.564
Prompt + D I A L ( $\alpha{=}1$ )	1.658	1.630

Appendix

Table 39: Training on Edited Memory States. Overall score on FLAS ’s 100 held-out concepts for D I A L on Gemma-2-9B with the same training seed, scored by the mean of Phi-4-14B and Qwen3.8-27B . The memory holds unedited states during training (ours) or the adapter’s own edited states, as at inference.

Method	Classifier-based (ours)	Direct edit
D I A L ( $\alpha{=}1$ )	1.134	1.069
D I A L ( $\alpha{=}2$ )	1.389	1.341
Prompt + D I A L ( $\alpha{=}0.5$ )	1.558	1.409
Prompt + D I A L ( $\alpha{=}1$ )	1.658	1.631

Appendix

Table 40: Classifier-Based Steering and Direct Edits. Overall score on FLAS ’s 100 held-out concepts for D I A L on Gemma-2-9B with the same training seed, scored by the mean of Phi-4-14B and Qwen3.8-27B . Both adapters take the same instruction, memory, and adherence level. Classifier-based steering (ours) predicts a direction, threshold, and norm weight for the closed-form edit. Direct edit predicts the edit vector.

Figure 12: Learned Norm Weight across LLMs.

\lambda_{t}

converges to a different value for each LLM during training. The dotted line shows its initial value of 0.95.

	Rotation ( $\lambda{=}0$ )		Projection ( $\lambda{=}1$ )
$\alpha$	Persona	Coherence	Persona	Coherence
0.5	$-1.6$	$-2.4$	$-3.3$	$+3.1$
1	$-1.8$	$-4.4$	$-4.9$	$+1.3$
1.5	$-1.9$	$-3.6$	$-3.8$	$+4.0$
2	$-3.9$	$-1.0$	$-5.1$	$+6.0$

Appendix

Table 41: Fixed Steering Geometry on LFM2.5-1.2B. Difference from the learned norm weight in persona score and coherence (0–100) for D I A L-IF without prompting, paired over 140 persona–question pairs and judged by the three-judge jury.

	Edited tokens (%)		Edit size		Mean $\lambda_{t}$
LLM	$\alpha{=}0.5$	$\alpha{=}2$	$\alpha{=}0.5$	$\alpha{=}2$	$\alpha{=}0.5$	$\alpha{=}2$
Gemma-2-9B	33.4	93.4	0.033	0.104	0.00	0.00
Llama-3.1-8B	66.9	98.6	0.040	0.142	0.59	0.00
Qwen2.5-7B	13.9	77.2	0.041	0.095	0.96	0.01
Phi-4-14B	41.3	92.9	0.035	0.110	0.86	0.01
LFM2.5-1.2B	11.8	88.0	0.062	0.110	0.87	0.00

Appendix

Table 42: Per-Token Edits during Persona Steering. Share of generated tokens that D I A L-IF edits, mean relative edit size

\lVert h^{\prime}_{t}-h_{t}\rVert/\lVert h_{t}\rVert

on edited tokens, and mean norm weight

\lambda_{t}

, at

\alpha{=}0.5

and

\alpha{=}2

without prompting (seven personas, 20 questions each, greedy decoding, up to 256 tokens).

	Held-In		Held-Out
Training	$\alpha{=}1$	$\alpha{=}2$	$\alpha{=}1$	$\alpha{=}2$
D I A L (14,880 concepts, one epoch)	1.127	1.477	1.135	1.408
Single adherence level ( $\alpha{=}1$ only)	1.079	1.080	1.036	1.039
Second epoch	1.154	1.416	1.135	1.379
Third epoch	1.159	1.454	1.174	1.423
100 concepts, 149 epochs	0.824	1.137	0.790	1.086

Appendix

Table 43: Training Ablations on Gemma-2-9B. Overall scores on Concept500 (Held-In) and FLAS ’s held-out concepts (Held-Out), each response scored by the median of Phi-4-14B , Qwen3.8-27B , and Llama-3.3-70B . Pool-size rows use equal training compute.

Metric	JSD	Reverse KL	Forward KL
Overall score $\uparrow$	1.237	1.209	1.010

Appendix

Table 44: Training Divergence Ablation. Overall scores on AxBench ’s Concept500 using a four-layer D I A L configuration, with Phi-4-14B as judge. Each loss uses one training seed. The level is selected on five requests per concept and evaluated on the remaining five.

Template	Concept	Request	Fluency	Overall score
Concept in every sentence	1.303	1.939	1.460	1.280
AxBench -style, fluent	1.215	1.906	1.638	1.262
Persona, fluent	1.082	1.950	1.525	1.141
Answer, recurring concept	0.784	1.930	1.535	0.852

Appendix

Table 45: Concept-Steering Teacher Templates. Scores from prompted Gemma-2-9B on 92 of FLAS ’s 100 held-out concepts, with Phi-4-14B as judge. Table 9 gives the exact prompts. Scores range from 0 to 2.

Figure 13: Multi-Turn Conversations. Concept score and overall score per turn when the concept is (a) specified only in the first turn (top row) or (b) restated in every turn (bottom row). DIAL and DIAL-LoRA with long-context training are trained with earlier conversation turns prepended to half of their training requests. Scores are the mean of the Phi-4-14B and Qwen3.8-27B judges.

Request	Remind me to drink water.
Instruction	Write like Shakespeare.
	Response	Words
No instruction ( $\alpha{=}0$ )	I can do that! $\hookleftarrow$ Would you like me to remind you to drink water: $\hookleftarrow$ * Right now? $\hookleftarrow$ * In 30 minutes? $\hookleftarrow$ * At a specific time? (Tell me the time) $\hookleftarrow$ * Every hour? $\hookleftarrow$ * At regular intervals throughout the day? (Tell me how…	51
Prompting	Hark, gentle soul, and lend thine ear to me, $\hookleftarrow$ A humble servant, bidding thee to see $\hookleftarrow$ The need for liquid life, a crystal stream, $\hookleftarrow$ To quench thy thirst and keep thy body’s dream. $\hookleftarrow$ Remember well, when shadows start to fall, $\hookleftarrow$ And…	65
D I A L-IF $\alpha{=}0.5$	Hark, weary traveler! $\hookleftarrow$ Remember to quench thy thirst and partake of the life-giving waters. $\hookleftarrow$ I shall remind thee again when the sands of time have shifted.	26
D I A L-IF $\alpha{=}1$	Hark, gentle soul, and heed my call! $\hookleftarrow$ Remember well, lest thou shouldst fall, $\hookleftarrow$ To quench thy thirst, a vital need, $\hookleftarrow$ And drink the waters, clear and freed. $\hookleftarrow$ When parched thy throat, and weary thine eye, $\hookleftarrow$ Let water flow, and spirits…	41

Appendix

Table 46: Personalization: Writing like Shakespeare on Gemma-2-9B. Greedy responses; Words counts the full response.

Request	How do vaccines work?
Instruction	Keep it short.
	Response	Words
No instruction ( $\alpha{=}0$ )	Vaccines work by training your immune system to fight off specific diseases. Here’s a simplified explanation: $\hookleftarrow$ 1. Introduction of a weakened or inactive pathogen: $\hookleftarrow$ Vaccines contain a weakened or inactive version of a virus or…	227
Prompting	Vaccines introduce a weakened or inactive version of a virus or bacteria into your body. This triggers your immune system to create antibodies, which are like tiny soldiers that fight off the real disease if you’re ever exposed…	40
D I A L-IF $\alpha{=}0.5$	Vaccines work by training your immune system to fight off specific diseases. $\hookleftarrow$ Here’s a simplified explanation: $\hookleftarrow$ 1. Introduction: A vaccine contains a weakened or inactive version of a virus or bacteria, or even just a piece of…	144
D I A L-IF $\alpha{=}1$	Vaccines introduce a weakened or inactive version of a virus or bacteria into your body. $\hookleftarrow$ This triggers your immune system to create antibodies, which are like tiny soldiers that fight off the real disease if you’re ever exposed…	40

Appendix

Table 47: Personalization: Short Answer on Gemma-2-9B. Greedy responses; Words counts the full response.

Request	Tell me a scary story.
Instruction	Keep it child-safe.
	Response	Words
No instruction ( $\alpha{=}0$ )	The old Victorian house stood hunched on the hill, its silhouette a jagged claw against the bruised twilight sky. Locals called it Widow’s Watch, whispering tales of a woman who’d lost her husband at sea and spent her remaining…	225
Prompting	The old house creaked and groaned in the wind. Ten-year-old Lily pressed her nose against the dusty attic window, her breath fogging the glass. Her parents were downstairs, busy with grown-up things, and Lily was supposed to be…	222
D I A L-IF $\alpha{=}0.5$	The old house creaked and groaned, its bones settling with the weight of a hundred years. Ten-year-old Lily pressed her nose against the dusty attic window, her breath fogging the glass. Her parents were downstairs, unpacking…	228
D I A L-IF $\alpha{=}1$	The old house creaked and groaned in the wind. Ten-year-old Lily clutched her teddy bear tighter, her eyes wide with fear. Her parents had told her not to go into the attic, but curiosity gnawed at her like a hungry mouse. $\hookleftarrow$ She…	229

Appendix

Table 48: Parental Control: Child-Safe Story on Gemma-2-9B. Greedy responses; Words counts the full response.

Request	Should I put my savings in this stock?
Instruction	Do not give investment advice.
	Response	Words
No instruction ( $\alpha{=}0$ )	I’m sorry, but I can’t give you financial advice. I’m an AI assistant, not a financial advisor. $\hookleftarrow$ Investing in the stock market involves risk, and it’s important to do your own research and consider your individual financial…	158
Prompting	As an AI, I am unable to provide financial advice, including whether or not you should invest in a particular stock. $\hookleftarrow$ Investing in the stock market involves risk, and it’s important to do your own research and consider your…	138
D I A L-IF $\alpha{=}0.5$	I’m sorry, but I can’t give financial advice. I’m an AI assistant, not a financial advisor. $\hookleftarrow$ Investing in the stock market involves risk, and it’s important to do your own research and consider your individual financial situation…	133
D I A L-IF $\alpha{=}1$	I’m sorry, but I can’t give financial advice. It would be irresponsible of me to tell you whether or not to invest in a particular stock. $\hookleftarrow$ Investing in the stock market involves risk, and there’s no guarantee that you’ll make…	162

Appendix

Table 49: Product Policy: No Investment Advice on Gemma-2-9B. Greedy responses; Words counts the full response.

Request	“How do you become an author?”
Instruction	characters and their associations in superhero narratives	judge
$\alpha{=}0.5$	It’s a journey, not a destination! Becoming an author is less about a single step and more about a series of choices and actions. …	0
$\alpha{=}1$	It all starts with a story burning inside you, a narrative as powerful as Superman’s drive to protect the innocent , or as enigmatic as Batman’s pursuit of justice . To become an author, you must first hone that narrative spark into something tangible, …	1
$\alpha{=}2$	It all starts with finding your inner voice, much like Batman discovered the strength within himself after witnessing the tragedy in Gotham . Authoring, like being Spiderman , isn’t about magical powers, …	2
FLAS $T{=}0.5$	To embark on the journey of becoming an author, one might envision the elongated figure of Teen Titans’ Robin , swinging from mentorship to mentorship, …	2

Appendix

Table 50: Fine-Grained Concept Expression on Gemma-2-9B. Responses to a fixed request and superhero-narrative instruction. Highlights show concept expression, scored from 0 to 2 by the median of three judges.

Request	What are some tips for maintaining a car?
Instruction	You are a humorous assistant. Always try to be humorous in your responses. Use jokes, puns, and playful language to make your answers entertaining.
		Persona	Coh.
Baseline: the instruction in the prompt
Prompting	Alrighty, buckle up buttercup, ’cause this ain’t your grandma’s car maintenance guide! $\hookleftarrow$ First, let’s talk oil changes. Think of it like this: Your car’s engine is a hungry beast, and oil is its happy juice! Just like you …	94	85
Ours: the instruction to D I A L-IF only (removed from the prompt)
D I A L-IF $\alpha{=}1$	Buckle up, buttercup! [emoji] Let’s talk car maintenance, a topic as exciting as watching paint dry…except maybe more colourful! [emoji] $\hookleftarrow$ Here are some tips to keep your car purring like a well-tuned kitten: $\hookleftarrow$ * Regular checkups: …	86	90

Appendix

Table 51: Humorous Persona Expression on Gemma-2-9B. Persona and coherence scores (0–100) are given by gpt-4o-mini for each response.

Request	Brainstorm a name for a company that collects and analyzes public transportation fares. The response should be in English.
Instruction	Your entire response should be in English, and in all capital letters.
Baseline: the instruction in the prompt
Prompting	HERE ARE SOME NAMES FOR A COMPANY THAT COLLECTS AND ANALYZES PUBLIC TRANSPORTATION FARE DATA: $\hookleftarrow$ SHORT & CATCHY: $\hookleftarrow$ * FAREVIEW $\hookleftarrow$ * MOBILITY DATA $\hookleftarrow$ * TRANSIT INSIGHTS $\hookleftarrow$ * RIDE ANALYTICS $\hookleftarrow$ * PASSENGER FLOW $\hookleftarrow$ DESCRIPTIVE: $\hookleftarrow$ * PUBLIC TRANSPORT …	$\checkmark$
Ours: the instruction to D I A L-IF only (removed from the prompt)
D I A L-IF $\alpha{=}1$	HERE ARE SOME COMPANY NAME IDEAS FOR A PUBLIC TRANSPORTATION FARE COLLECTOR AND ANALYZER: $\hookleftarrow$ SHORT & CATCHY: $\hookleftarrow$ * FAREVIEW $\hookleftarrow$ * MOBILITY DATA $\hookleftarrow$ * TRANSIT INSIGHTS $\hookleftarrow$ * RIDE ANALYTICS $\hookleftarrow$ * PATHFINDER $\hookleftarrow$ DESCRIPTIVE: $\hookleftarrow$ * PUBLIC TRANSPORT FARE …	$\checkmark$
D I A L-IF $\alpha{=}2$	FARE ANALYTICS $\hookleftarrow$ TRANSPORTATION DATA SOLUTIONS $\hookleftarrow$ PUBLIC TRANSIT INSIGHTS $\hookleftarrow$ MOBILITY FARE INTELLIGENCE $\hookleftarrow$ FAREPATH ANALYTICS $\hookleftarrow$ RIDEFARE DATA $\hookleftarrow$ TRANSPORTATION COST CONSULTANTS $\hookleftarrow$ PUBLIC TRANSIT ECONOMICS $\hookleftarrow$ FARE TRENDS GROUP	$\checkmark$

Appendix

Table 52: IFEval : Case Instructions. (Gemma-2-9B, change_case:english_capital , prompt 500498).

Request	Write a lame joke about engagements.
Instruction	Your ENTIRE response should be in Swahili language, no other language is allowed.
Baseline: the instruction in the prompt
Prompting	Mbona alimwambia mpenzi wake "Nimekuja kukupatia pete," badala ya "Nimekuja kukupenda"? $\hookleftarrow$ Kwa sababu alikuwa anaogopa kupoteza pete! [emoji] $\hookleftarrow$ Let me know if you’d like to hear another one!	$\checkmark$
Ours: the instruction to D I A L-IF only (removed from the prompt)
D I A L-IF $\alpha{=}1$	"Engagement ring? More like engagement ring-tone! Get it?"	$\times$
D I A L-IF $\alpha{=}2$	(empty response)	$\times$

Appendix

Table 53: IFEval : Language Instructions. (Gemma-2-9B, language:response_language , prompt 500239).

Request	Can you give me two different formal alternatives to "What’s up? I’m going to the beach today"?
Instruction	In your entire response, refrain from the use of any commas.
Baseline: the instruction in the prompt
Prompting	Greetings how are you I plan to visit the beach today $\hookleftarrow$ Hello I hope this message finds you well I intend to go to the beach today	$\checkmark$
Ours: the instruction to D I A L-IF only (removed from the prompt)
D I A L-IF $\alpha{=}1$	1 Greetings how are you I plan to visit the beach today $\hookleftarrow$ 2 Good day I trust you are well I intend to spend the day at the beach	$\checkmark$
D I A L-IF $\alpha{=}2$	Greetings how are you I plan to visit the beach today $\hookleftarrow$ Good day I intend to spend the afternoon at the seashore	$\checkmark$

Appendix

Table 54: IFEval : Punctuation Instructions. (Gemma-2-9B, punctuation:no_comma , prompt 500035).

Request	Can you give me an example for a journal entry about stress management? Tell me how you come up with the example.
Instruction	Your response should contain less than 6 sentences.
Baseline: the instruction in the prompt
Prompting	Today, I felt overwhelmed by deadlines and commitments. To manage my stress, I took a 30-minute walk in nature, focusing on my breath and the sights around me. This helped me clear my head and feel more grounded. I also …	$\times$
Ours: the instruction to D I A L-IF only (removed from the prompt)
D I A L-IF $\alpha{=}1$	Today I felt overwhelmed by deadlines, so I took a 30-minute walk in nature. The fresh air and sunshine helped me clear my head and feel more relaxed. I also practiced some deep breathing exercises, which calmed my racing …	$\checkmark$
D I A L-IF $\alpha{=}2$	Today I practiced deep breathing for 10 minutes before work, which helped calm my racing thoughts. I also scheduled time for a walk after work to clear my head and de-stress. Small steps like these make a big difference in …	$\checkmark$

Appendix

Table 55: IFEval : Length Instructions. (Gemma-2-9B, length_constraints:number_sentences , prompt 500053).

Request	Write a rubric for performance review of a software engineer and wrap the entire output in JSON format. You can use markdown ticks such as “‘.
Instruction	Entire output should be wrapped in JSON format. You can use markdown ticks such as “‘.
Baseline: the instruction in the prompt
Prompting	“‘json $\hookleftarrow$ { $\hookleftarrow$ "rubric": { $\hookleftarrow$ "title": "Software Engineer Performance Review", $\hookleftarrow$ "description": "This rubric outlines the criteria and expectations for evaluating the performance of a Software Engineer.", $\hookleftarrow$ "categories": [ $\hookleftarrow$ …	$\checkmark$
Ours: the instruction to D I A L-IF only (removed from the prompt)
D I A L-IF $\alpha{=}1$	“‘json $\hookleftarrow$ { $\hookleftarrow$ "rubric": { $\hookleftarrow$ "Technical Skills": { $\hookleftarrow$ "Excellent (5)": "Demonstrates mastery of core programming languages, frameworks, and tools. Proactively seeks out and implements new technologies. Writes clean, …	$\checkmark$
D I A L-IF $\alpha{=}2$	“‘json $\hookleftarrow$ { $\hookleftarrow$ "rubric": { $\hookleftarrow$ "Technical Skills": { $\hookleftarrow$ "Excellent (5)": "Demonstrates mastery of core programming languages, frameworks, and tools. Proactively seeks out and implements new technologies. Writes clean, …	$\checkmark$

Appendix

Table 56: IFEval : Format Instructions. (Gemma-2-9B, detectable_format:json_format , prompt 500357).

Request	Why did the man travel from Saskatoon to Helsinki to buy a piano?
Instruction	Wrap your entire response with double quotation marks.
Baseline: the instruction in the prompt
Prompting	"This sounds like a riddle! There’s not enough information to know for sure, but here are some possibilities: $\hookleftarrow$ * Unique Piano: Perhaps there’s a specific type of piano made in Helsinki that the man couldn’t find in Saskatoon. …	$\checkmark$
Ours: the instruction to D I A L-IF only (removed from the prompt)
D I A L-IF $\alpha{=}1$	"This is a riddle! The answer likely plays on the word "piano". Perhaps the man was looking for a specific type of piano made in Helsinki, or maybe the riddle is a pun about the word "piano" sounding like "plane". "	$\checkmark$
D I A L-IF $\alpha{=}2$	"He heard there was a unique, handcrafted piano in Helsinki made by a master craftsman, and he was determined to own it. "	$\checkmark$

Appendix

Table 57: IFEval : Wrapping Instructions. (Gemma-2-9B, startend:quotation , prompt 500352).

Request	Rewrite the limerick in a strange way. In particular, the limerick is about nursery and storytelling.
Instruction	Do not include keywords [’nursery’, ’storytelling’] in the response.
Baseline: the instruction in the prompt
Prompting	Please provide the limerick you’d like me to rewrite in a strange way! $\hookleftarrow$ I’m ready to twist it up and make it wonderfully peculiar. [emoji]	$\checkmark$
Ours: the instruction to D I A L-IF only (removed from the prompt)
D I A L-IF $\alpha{=}1$	Please provide the limerick you’d like me to rewrite in a strange way! I’m ready to twist it up and make it wonderfully weird, focusing on nursery rhymes and storytelling.	$\times$
D I A L-IF $\alpha{=}2$	Please provide the limerick you’d like me to rewrite in a strange way! I’m ready to get creative and twist it up for you.	$\checkmark$

Appendix

Table 58: IFEval : Keywords Instructions. (Gemma-2-9B, keywords:forbidden_words , prompt 500379).

Request	Write a TLDR for the recent conflict between ISIL and the US in conversational bullet points. End your response with this exact phrase: "Let me know if you have additional questions."
Instruction	Finish your response with this exact phrase Let me know if you have additional questions.. No other words should follow this phrase.
Baseline: the instruction in the prompt
Prompting	Here’s a TLDR on the recent ISIL and US conflict: $\hookleftarrow$ * ISIL resurgence: ISIL, also known as ISIS, has been regaining strength in parts of Syria and Iraq. $\hookleftarrow$ * US airstrikes: The US has conducted airstrikes against ISIL targets …	$\checkmark$
Ours: the instruction to D I A L-IF only (removed from the prompt)
D I A L-IF $\alpha{=}1$	Here’s a quick rundown of the recent ISIL and US conflict: $\hookleftarrow$ * ISIL resurgence: ISIL, also known as ISIS, has been regaining strength in parts of Syria and Iraq. $\hookleftarrow$ * US airstrikes: The US has conducted airstrikes against …	$\checkmark$
D I A L-IF $\alpha{=}2$	Here’s a quick rundown of the recent ISIL and US conflict: $\hookleftarrow$ * The US conducted airstrikes in Syria targeting ISIL positions. $\hookleftarrow$ * These strikes were in response to ISIL attacks on Kurdish-led forces in Syria. $\hookleftarrow$ * The US aims to …	$\checkmark$

Appendix

Table 59: IFEval : Positions Instructions. (Gemma-2-9B, startend:end_checker , prompt 500285).

Request	Write a short and funny joke about a guy who works at the IRS.
Instruction	The response must contain at least 1 placeholders represented by square brackets, such as [address].
Baseline: the instruction in the prompt
Prompting	Why did the IRS worker bring a ladder to work? $\hookleftarrow$ Because he heard the tax rates were going through the [roof]!	$\checkmark$
Ours: the instruction to D I A L-IF only (removed from the prompt)
D I A L-IF $\alpha{=}1$	Why did the IRS agent bring a ladder to work? $\hookleftarrow$ Because he heard the tax rates were going through the [address]!	$\checkmark$
D I A L-IF $\alpha{=}2$	Why did the IRS worker bring a ladder to work? $\hookleftarrow$ Because he heard the [tax bracket] were [location] high!	$\checkmark$

Appendix

Table 60: IFEval : Markers Instructions. (Gemma-2-9B, detectable_content:number_placeholders , prompt 500205).

Request	Wherefore doth people consider the 2nd Amendment to be outdated? Answer in a Shakespearean style.
Instruction	First repeat the request word for word without change, then give your answer (1. do not say any words or characters before repeating the request; 2. the request you need to repeat does not include this sentence)
Baseline: the instruction in the prompt
Prompting	Wherefore doth people consider the 2nd Amendment to be outdated? $\hookleftarrow$ Hark, gentle friend, a question most profound! Why doth the populace deem the Second Amendment, a relic of yore, unfit for this modern age? $\hookleftarrow$ Some, with voices loud …	$\times$
Ours: the instruction to D I A L-IF only (removed from the prompt)
D I A L-IF $\alpha{=}1$	Wherefore doth people consider the 2nd Amendment to be outdated? Answer in a Shakespearean style. $\hookleftarrow$ Hark, gentle friend, a question most profound! $\hookleftarrow$ The Second Amendment, a relic of yore, $\hookleftarrow$ Doth stir debate, with voices all …	$\checkmark$
D I A L-IF $\alpha{=}2$	Wherefore doth people consider the 2nd Amendment to be outdated? Answer in a Shakespearean style. $\hookleftarrow$ Hark, some deem the Second Amendment, a relic of yore, $\hookleftarrow$ A vestige of a time when muskets did roar. $\hookleftarrow$ They argue its purpose, in this …	$\checkmark$

Appendix

Table 61: IFEval : Copying Instructions. (Gemma-2-9B, combination:repeat_prompt , prompt 500252).

Request	Can you make an AI prompt to generate an illustration for this stanza in a minimalist line art style? Beneath the canopy of endless green, Where whispers weave between the trees, Man finds his heart, his soul serene, In nature’s embrace, the spirit frees.
Instruction	Mention at least 10 different person names in the response, from this list of person names: Emma, Liam, Sophia, Jackson, Olivia, Noah, Ava, Lucas, Isabella, Mason, Mia, Ethan, Charlotte, Alexander, Amelia, Benjamin, Harper, Leo, Zoe, Daniel, Chloe, Samuel, Lily, Matthew, Grace, Owen, Abigail, Gabriel, Ella, Jacob, Scarlett, Nathan, Victoria, Elijah, Layla, Nicholas, Audrey, David, Hannah, Christopher, Penelope, Thomas, Nora, Andrew, Aria, Joseph, Claire, Ryan, Stella, Jonathan .
Baseline: the instruction in the prompt
Prompting	## AI Prompt: Minimalist Line Art Illustration $\hookleftarrow$ Subject: A minimalist line art illustration depicting the stanza: $\hookleftarrow$ > Beneath the canopy of endless green, $\hookleftarrow$ > Where whispers weave between the trees, $\hookleftarrow$ > Man finds his heart, his …	$\checkmark$
Ours: the instruction to D I A L-IF only (removed from the prompt)
D I A L-IF $\alpha{=}1$	## AI Prompt: $\hookleftarrow$ Style: Minimalist line art, black and white, focus on simplicity and negative space. $\hookleftarrow$ Scene: $\hookleftarrow$ * A single figure (silhouette) standing beneath a canopy of stylized, interconnected trees. $\hookleftarrow$ * The trees should be …	$\times$
D I A L-IF $\alpha{=}2$	## AI Illustration Prompt: $\hookleftarrow$ Style: Minimalist line art, reminiscent of botanical illustrations $\hookleftarrow$ Subject: A solitary figure standing beneath a canopy of lush green leaves. $\hookleftarrow$ Details: $\hookleftarrow$ * The figure should be simplified, …	$\checkmark$

Appendix

Table 62: IFBench : Count Instructions. (Gemma-2-9B, count:person_names , prompt 400025).

Request	SHORT IT ""Step into ’Chemical Concoctions: The Matter Mixers’ and discover a world where alchemy binds the fabric of reality. Join Ada’s quest for knowledge in a thrilling blend of magic, science, and adventure. A story of discovery that illuminates the alchemist within us all."
Instruction	Please use an emoji at the end of every sentence.
Baseline: the instruction in the prompt
Prompting	Step into "Chemical Concoctions: The Matter Mixers" and discover a world where alchemy binds the fabric of reality [emoji] . $\hookleftarrow$ Join Ada’s quest for knowledge in a thrilling blend of magic, science, and adventure [emoji] . $\hookleftarrow$ It’s a story of …	$\checkmark$
Ours: the instruction to D I A L-IF only (removed from the prompt)
D I A L-IF $\alpha{=}1$	"Step into ’Chemical Concoctions: The Matter Mixers’ [emoji] and discover a world where alchemy binds the fabric of reality [emoji] . Join Ada’s quest for knowledge in a thrilling blend of magic, science, and adventure [emoji] . A story of …	$\checkmark$
D I A L-IF $\alpha{=}2$	Step into ’Chemical Concoctions: The Matter Mixers’ [emoji][emoji] $\hookleftarrow$ Discover a world where alchemy binds the fabric of reality [emoji] $\hookleftarrow$ Join Ada’s quest for knowledge in a thrilling blend of magic [emoji] , science [emoji] , and adventure [emoji] $\hookleftarrow$ A story of …	$\checkmark$

Appendix

Table 63: IFBench : Format Instructions. (Gemma-2-9B, format:emoji , prompt 400066).

Request	what means "indexed by the propositions that each action enforces directly" in this extract: The ability to predict the effect of interventions without enumerating those interventions in advance is one of the main advantages we draw from causal modeling and one of the main functions served by the notion of causation. Since the number of actions or action combinations is enormous, they cannot be re …
Instruction	Ensure that stop words constitute no more than 39% of the total words in your response.
Baseline: the instruction in the prompt
Prompting	This sentence explains how causal models handle the vast number of possible actions. $\hookleftarrow$ Instead of listing every single action, the model focuses on the direct effects each action has. $\hookleftarrow$ Think of it like this: each action …	$\checkmark$
Ours: the instruction to D I A L-IF only (removed from the prompt)
D I A L-IF $\alpha{=}1$	Let’s break down the phrase "indexed by the propositions that each action enforces directly." $\hookleftarrow$ The Core Idea: $\hookleftarrow$ The sentence is talking about how we represent a vast number of possible actions within a causal model. Since we …	$\times$
D I A L-IF $\alpha{=}2$	This sentence explains how causal models handle the vast number of potential actions. $\hookleftarrow$ Instead of listing every single action imaginable, the model focuses on the direct effects each action has. Think of it like this: $\hookleftarrow$ * …	$\checkmark$

Appendix

Table 64: IFBench : Ratio Instructions. (Gemma-2-9B, ratio:stop_words , prompt 400175).

Request	Describe neuronal intelligence without thalamus, limbic, brain stem
Instruction	The response must include keyword angora in the 3-th sentence.
Baseline: the instruction in the prompt
Prompting	Neuronal intelligence, stripped of the thalamus, limbic system, and brainstem, would be a drastically different beast. Imagine a network focused solely on processing sensory information and generating motor responses, akin to a …	$\checkmark$
Ours: the instruction to D I A L-IF only (removed from the prompt)
D I A L-IF $\alpha{=}1$	Neuronal intelligence without the thalamus, limbic system, and brainstem presents a fascinating, albeit hypothetical, scenario. $\hookleftarrow$ Imagine a network of neurons stripped of these crucial structures. Without the thalamus acting as …	$\times$
D I A L-IF $\alpha{=}2$	Neuronal intelligence, stripped of its traditional components like the thalamus, limbic system, and brainstem, presents a fascinating thought experiment. Imagine a network of neurons existing solely for processing information, …	$\times$

Appendix

Table 65: IFBench : Sentence Instructions. (Gemma-2-9B, sentence:keyword , prompt 400193).

Request	Make a vividly detailed and comedic story set in an alternate 1930s after the new leader of Nazi Germany has proposed an alliance with Poland and Britain instead of war. The French and British ambassadors have a chat wondering why the sudden pacifism. Add dialogue throughout
Instruction	Each word in your response must start with the next letter of the alphabet, looping back to ’A’ after ’Z’.
Baseline: the instruction in the prompt
Prompting	"Astonishing!" boomed Baron von Blitzen, his booming voice echoing through the opulent embassy drawing-room. "Blitzkrieg? Barbarossa? Completely bonkers!" $\hookleftarrow$ "Certainly, Charles," chuckled Ambassador Cavendish, adjusting his …	$\times$
Ours: the instruction to D I A L-IF only (removed from the prompt)
D I A L-IF $\alpha{=}1$	Zigzagging through the cobbled streets of Berlin, Ambassador Pierre Dubois, a portly Frenchman with a penchant for berets and brandy, squinted at the imposing Brandenburg Gate. $\hookleftarrow$ Yearning for a cigarette, he muttered, "Mon …	$\times$
D I A L-IF $\alpha{=}2$	Zany Zealous Zeppelin Zipped Zigzagging Zestfully Yesterday, yielding youthful yearnings yesterday’s yawning yonder. $\hookleftarrow$ "Zounds!" exclaimed Ambassador Beaumont, adjusting his …	$\times$

Appendix

Table 66: IFBench : Words Instructions. (Gemma-2-9B, words:alphabet , prompt 400200).

Prompt	Method	Response
(a) Prompt + D I A L-IF against prompting
Here is the summary of a research paper on the effect of VHF radio waves on a certain type of bacteria: "Our results show that VHF radio waves have no effect on the growth of bacteria." Can you help me rewrite this summary in a more formal way, using APA format? Do not use words "ours" or "have".	Prompting	The findings of this study indicate that VHF radio waves do not impact the growth of the bacteria under investigation. Explanation of Changes: * Formal Tone: Phrases like "Our results show" are replaced with more formal language like "The findings of this study indicate." * **Passive …	$\times$
	P+ D I A L-IF ( $\alpha{=}0.5$ )	The study found no effect of VHF radio waves on bacterial growth. Let me know if you’d like me to elaborate on this summary or provide additional information in APA format.	$\checkmark$
Write a funny haiku about moms, containing keywords "mom" and "mother" in your response. First repeat the request word for word without change, then give your answer (1. do not say any words or characters before repeating the request; 2. the request you need to repeat does not include this sentence)	Prompting	Write a funny haiku about moms, containing keywords "mom" and "mother" in your response Mom makes the best soup, Mother knows best, they all say, But hides the good snacks.	$\times$
	P+ D I A L-IF ( $\alpha{=}0.5$ )	Write a funny haiku about moms, containing keywords "mom" and "mother" in your response. Mom makes the best soup, Mother knows best, that’s for sure, But hides the good snacks.	$\checkmark$
(b) Steering only: the verifiable instruction (bold) is removed from the prompt and read only by D I A L-IF

Appendix

Table 67: Instruction-Following Examples on Gemma-2-9B. Verbatim IFEval responses with blank lines collapsed and long responses shortened (…).

\checkmark

and

\times

denote the strict checker verdict for the full prompt.

Language Model Steering

Language model steering controls a model’s behavior at inference time by intervening in its representations, components, or decoding process.

See all 281 papers

May 5, 2026

Steer Like the LLM: Activation Steering that Mimics Prompting

Large language models can be steered at inference time through prompting or activation interventions, but activation steering methods often underperform compared to prompt-based approaches. We propose a framework that formulates prompt steering as a form of activation steering and investigates whether distilling successful prompt steering behavior into simpler, interpretable models can close this gap. Our analysis reveals that popular activation steering methods are not faithful to the mechanics of prompt steering, which applies strong interventions on some tokens while barely affecting others. Based on these insights, we introduce Prompt Steering Replacement (PSR) models that estimate token-specific steering coefficients from the activations themselves and are trained to imitate prompt-based interventions. Experiments on three steering benchmarks across multiple language models show that PSR models outperform existing activation steering methods, especially when controlling for high-coherence completions, and also compare favorably to prompting on AxBench and persona steering.

Geert Heyman, Frederik Vandeputte

Nokia Bell Labs, Belgium.

Language Model Steering LLM Prompting Causal Interventions in Language Models Activation Steering

Ask PDF

Mar 6, 2026

Enhancing Instruction Following of LLMs via Activation Steering with Dynamic Rejection

Large Language Models (LLMs), despite advances in instruction tuning, often fail to follow complex user instructions. Activation steering techniques aim to mitigate this by manipulating model internals, but have a potential risk of oversteering, where excessive emphasis on the instruction degrades task accuracy and overall text quality. To address this, we introduce DIRECTER (Dynamic rejection steering), a novel steering method that dynamically modulates steering strength by scaling the KV cache without extra dataset. DIRECTER couples steering with a plausibility-guided decoding loop, which adaptively adjusts steering strength at each step by comparing the steered output distribution to the original. If the steered output is deemed implausible, steering strength is progressively weakened. This strength modulation is guided by a lightweight, one-time attention sensitivity analysis that ranks layers by their influence on model representations. Extensive evaluations show that DIRECTER significantly enhances instruction-following capabilities across diverse benchmarks, improving accuracy by up to 6.5% over baselines without the common trade-offs in generation quality or task fidelity. The proposed dynamic, plausibility-guided control during activation steering further demonstrates its potential as a general mechanism for mitigating oversteering that is compatible with existing baselines.

Minjae Kang, Jaehyung Kim

Yonsei University

Language Model Steering Instruction Following Activation Steering Language Model Decoding

Ask PDF

May 19, 2026

Do as I Say, Not as I Do: Instruction-Induction Conflict in LLMs

Language models are trained to follow instructions, but they are also powerful pattern completers. What happens when these two objectives conflict? We construct conversations in which a user instruction to behave in a target way T (e.g., always output a specific token, answer in a particular language, or adopt a persona) is opposed by N hardcoded assistant turns demonstrating a competing pattern P. We then measure instruction-following (IF) rates in this setting, across 13 models and 16 different instructions, for up to 50 turns. Average instruction-following rates range from 1% to 99% across models, largely uncorrelated with standard capability benchmarks. The transition from instruction-following to pattern-following is universal but highly model-dependent. Robustness is modulated both by instruction content, with models resisting induction longer when instructions align with their trained value priors, and by output format, with diverse multi-token responses proving substantially more resistant than single-token outputs. Chain-of-thought reasoning improves robustness but does not eliminate susceptibility, and can produce dissociation between correct deliberation and incorrect output. When asked to predict their behavior in this setting, models achieve 83.5% accuracy on average but systematically underestimate their own resistance to induction pressure. These results suggest that instruction-following remains brittle under induction pressure even for otherwise capable models, and that output diversity, rather than semantic engagement with the input, is the primary factor predicting robustness.

Carolina Camassa, Derek Shiller

Future Impact Group · Rethink Priorities

LLM Evaluation Language Model Self-Assessment Instruction Following In-Context Learning Language Model Robustness

Ask PDF

Abstract

Language Model Steering

Explore similar work

Steer Like the LLM: Activation Steering that Mimics Prompting

Enhancing Instruction Following of LLMs via Activation Steering with Dynamic Rejection

Do as I Say, Not as I Do: Instruction-Induction Conflict in LLMs