aiqbench analysis
DUET introduces a novel token-selective distillation technique for enforcing runtime prohibitions in LLMs by leveraging disagreement between paired teachers with differing visibility to policy constraints. This method isolates causal violation signals at the token level, enabling efficient alignment without requiring offline preference data or suffering from standard post-training limitations like SFT and DPO. The approach successfully balances strict compliance on enterprise policies while maintaining high utility across general reasoning benchmarks.
DUET: Dual-Teacher On-Policy Distillation via Same-Weight Disagreement for Prohibition Compliance
arXiv:2608.14644v1 Announce Type: new Abstract: Real-world LLM deployments increasingly rely on runtime-injected prohibitions--enterprise policies, PII redlines, tool boundaries--that vary per request and per tenant. Conventional post-training is structurally ill-suited: SFT hides the violation signal in compliant labels, and DPO's sequence-level preferences mismatch token-localized violations. We propose DUET, a token-selective on-policy distillation method for prohibition compliance. DUET pairs a teacher that sees the prohibition (positive) with an identical-weight teacher that does not (negative). Because the two teachers differ only in prohibition visibility, their per-token disagreement isolates the prohibition's causal effect--yielding a clean supervision signal uncontaminated by model capacity or mismatch. This disagreement drives two complementary mechanisms: signal cleaning, which discards agreement tokens as redundant or prefix-corrupted, and preference-directed learning, which pushes the student away from the negative teacher and toward the positive one at token granularity, embedding DPO-style optimization directly into OPD without offline preference data. We construct an industrial Prohibition-Compliance benchmark spanning five task families covering explicit-refusal, paraphrase robustness, and over-refusal. Across 1.5B-8B Qwen variants, DUET achieves 72.3-85.2% violation compliance while preserving 88-93% normal utility, dramatically outperforming teacher model and other distillation baselines. External evaluation on SysBench confirms improved safety alignment with minimal degradation on GSM8K and MATH-500.
Abstract:Real-world LLM deployments increasingly rely on runtime-injected prohibitions--enterprise policies, PII redlines, tool boundaries--that vary per request and per tenant. Conventional post-training is structurally ill-suited: SFT hides the violation signal in compliant labels, and DPO's sequence-level preferences mismatch token-localized violations. We propose DUET, a token-selective on-policy distillation method for prohibition compliance. DUET pairs a teacher that sees the prohibition (positive) with an identical-weight teacher that does not (negative). Because the two teachers differ only in prohibition visibility, their per-token disagreement isolates the prohibition's causal effect--yielding a clean supervision signal uncontaminated by model capacity or mismatch. This disagreement drives two complementary mechanisms: signal cleaning, which discards agreement tokens as redundant or prefix-corrupted, and preference-directed learning, which pushes the student away from the negative teacher and toward the positive one at token granularity, embedding DPO-style optimization directly into OPD without offline preference data. We construct an industrial Prohibition-Compliance benchmark spanning five task families covering explicit-refusal, paraphrase robustness, and over-refusal. Across 1.5B-8B Qwen variants, DUET achieves 72.3-85.2% violation compliance while preserving 88-93% normal utility, dramatically outperforming teacher model and other distillation baselines. External evaluation on SysBench confirms improved safety alignment with minimal degradation on GSM8K and MATH-500.
Machine Learning (cs.LG); Computation and Language (cs.CL)
arXiv:2608.14644 [cs.LG]
(or arXiv:2608.14644v1 [cs.LG] for this version)
https://doi.org/10.48550/arXiv.2608.14644 Focus to learn more arXiv-issued DOI via DataCite
arXivLabs is a framework that allows collaborators to develop and share new arXiv features directly on our website.
Both individuals and organizations that work with arXivLabs have embraced and accepted our values of openness, community, excellence, and user data privacy. arXiv is committed to these values and only works with partners that adhere to them.
Have an idea for a project that will add value for arXiv's community? Learn more about arXivLabs.
Full text under the publisher's open licence · 289 words
Related coverage
TIDE: Teacher-Student Transition via Informative Distillation and Exploration for Agentic RL
arXiv · Sep 30, 2026
Who Teaches Which Token? Verifier-Gated Multi-Expert On-Policy Distillation for Scientific Reasoning
arXiv · Sep 15, 2026
Token Utility for On-Policy Distillation via Local Trust-Region Optimization
arXiv · Oct 15, 2026
[AINews] Quasi-Riemann-Hypothesis: OpenAI publishes 722 math papers solving 90 of the top 500 open math problems; “the most significant moment” in >100 years of mathematics
Latent Space · Oct 7, 2026