aiqbench analysis
The paper studies distillation where a student model improves only from scalar scores given by an LLM judge, with no teacher-written answers in its training data. It proposes embedding a registered pattern of preferences in those scores so the pattern carries over to a student trained by another party, which can then be detected from outputs alone. Readers should note that the results are reported on code and math tasks, and the detection claims depend on a fixed reference pool, so practical robustness across other domains and adversarial countermeasures remains untested.
Radioactive Feedback: Tracing Models Trained by Label-Only LLM Supervision
arXiv:2610.15685v1 Announce Type: new Abstract: Conventional distillation transfers capability through teacher-generated demonstrations. LLM-as-a-Judge knowledge distillation offers a distinct and increasingly practical alternative. In the label-only variant studied here, the student generates every response and improves from scalar scores alone. Prior work shows that this non-demonstrative feedback can improve student capability, but it creates a fundamentally different provenance problem. No teacher-generated answer enters the student's training set, so any evaluator-specific trace must pass through a low-bandwidth reward signal rather than copied text. Existing watermarks and distillation detectors do not cover this score-only channel. We introduce Radioactive Feedback, which encodes a registered pattern of behavioral preferences in the rewards themselves. The evaluator applies bounded pointwise adjustments to quality-qualified responses, allowing repeated optimization to transfer the pattern to a student trained by another party. Experiments on code and mathematics show that the registered preferences transfer through scores alone and across datasets, enabling black-box detection against a fixed reference pool.
Abstract:Conventional distillation transfers capability through teacher-generated demonstrations. LLM-as-a-Judge knowledge distillation offers a distinct and increasingly practical alternative. In the label-only variant studied here, the student generates every response and improves from scalar scores alone. Prior work shows that this non-demonstrative feedback can improve student capability, but it creates a fundamentally different provenance problem. No teacher-generated answer enters the student's training set, so any evaluator-specific trace must pass through a low-bandwidth reward signal rather than copied text. Existing watermarks and distillation detectors do not cover this score-only channel. We introduce Radioactive Feedback, which encodes a registered pattern of behavioral preferences in the rewards themselves. The evaluator applies bounded pointwise adjustments to quality-qualified responses, allowing repeated optimization to transfer the pattern to a student trained by another party. Experiments on code and mathematics show that the registered preferences transfer through scores alone and across datasets, enabling black-box detection against a fixed reference pool.
Artificial Intelligence (cs.AI)
arXiv:2610.15685 [cs.AI]
(or arXiv:2610.15685v1 [cs.AI] for this version)
https://doi.org/10.48550/arXiv.2610.15685 Focus to learn more arXiv-issued DOI via DataCite (pending registration)
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 · 245 words
Related coverage
Post-Training Large Language Models via Reinforcement Learning from Self-Feedback
arXiv · Sep 16, 2026
Thousand-dimensional structure
AI Alignment Forum · Jul 30, 2026
Preparing data for supervised fine-tuning Part 1: Formatting and quality
AWS · Aug 26, 2026
Simulating student learning for better AI tutoring
Microsoft Research · Oct 12, 2026