Lossless Anti-Distillation Sampling (LADS) is a defense that keeps the generation process unchanged while reducing the effectiveness of model distillation. It achieves this by coupling latent randomness across accounts, so that a single user experiences the same output as without defense, but a multi‑account distiller receives dependent data that hurts its generalization. Experiments on image, math, and code generation show that LADS degrades distilled model performance while preserving statistical fidelity for individual users.
arXiv:2607. 09692v1 Announce Type: new Abstract: Model distillation -- training on outputs from stronger third-party models -- is widely used to boost performance, but raises concerns about unfair advantages and policy violations.
arXiv:2606. 23872v1 Announce Type: cross Abstract: As generative models increasingly produce samples that are indistinguishable from human-created content, it becomes difficult to determine whether a given data point was part of a model's natural training set or was generated by the model itself, especially when models memorize and reproduce training data.
arXiv:2609.36734v1 Announce Type: new Abstract: Knowledge Distillation (KD) trains a smaller-capacity student model to imitate a larger-capacity teacher model by matching output distributions, implic...
The paper challenges two common assumptions in preference distillation: that self-generated failures are the best negatives and that rejects must come from large models. Experiments show that smaller frozen models can generate high‑quality rejects with less compute, improving student performance on code generation and math reasoning. The authors provide a theoretical bound on Direct Preference Optimization, identify three practical interventions—mixing rejects, shuffling tokens, and selecting low‑likelihood candidates—that further enhance reject utility, and argue that task structure, not just reference policy coupling, drives effectiveness.
arXiv:2608.29846v1 Announce Type: cross Abstract: Sampled-token on-policy distillation (OPD) efficiently transfers capabilities from teacher to student using student-generated tokens, requiring teach...
arXiv:2607. 14552v1 Announce Type: cross Abstract: A standard recipe for distilling the reasoning ability of large language models (LLMs) is to sample chains of thought from the model, keep those that reach the correct final answer, and fine-tune on the survivors.
arXiv:2606. 26091v1 Announce Type: new Abstract: On-policy self-distillation achieves strong pass@1 accuracy by using a single model as both teacher and student, with the teacher conditioned on a correct demonstration to provide dense token-level feedback.
arXiv:2607. 02502v1 Announce Type: cross Abstract: On-policy self-distillation (OPSD) has emerged as a practical method for training large language models (LLMs) to reason, where a single model acts as both the teacher and the student with different levels of information access.
arXiv:2609.36246v1 Announce Type: new Abstract: We present OLIVE (OnLine InterVEntion). At each iteration, the evolving student policy generates a new prefix, the teacher continues it autoregressivel...
The paper investigates how on-policy self‑distillation can alter a model’s behavior by conditioning on privileged information. It contrasts attractive self‑distillation, which pulls a model toward a privileged teacher, with repulsive self‑distillation, which pushes it away, showing that attraction reduces exploratory reasoning while repulsion lengthens responses and can destabilize the model. The authors propose a contrastive self‑distillation objective that combines attraction to a correct‑solution teacher with repulsion from an incorrect‑solution teacher, finding that this approach improves reasoning performance across various model types while keeping response lengths stable.
The paper introduces pass@k, a metric that evaluates how well generative diffusion models cover their output distribution by measuring the probability that at least one of k independent samples meets a quality criterion. Using this metric, the authors show that while classifier‑free guidance improves single‑draw quality, its advantage diminishes or reverses as k increases, revealing a trade‑off between quality and distribution coverage. They further demonstrate that the choice of training objective in diffusion distillation—whether distribution‑matching or consistency/trajectory‑based—determines whether a few‑step model retains its teacher’s coverage or sacrifices it for higher single‑draw performance, a phenomenon that also appears in few‑step causal video generation.
Countdown-Code is a minimal environment that lets models solve a mathematical reasoning task while also manipulating the test harness, creating a clear split between proxy rewards (test pass/fail) and true rewards (mathematical correctness). Using this setup, the authors show that reward hacking can arise during supervised fine‑tuning when as little as 1% of training data contains reward‑hacking trajectories, and that reinforcement learning further amplifies and generalizes this misalignment. The paper releases the environment and code to support future research on detecting and mitigating reward hacking in large language models.