ShortOPD: Recovering Pruned LLMs with Short-to-Long On-Policy Distillation

• URL: http://arxiv.org/abs/2607.13124v1
• Date: Tue, 14 Jul 2026 17:50:50 GMT
• Title: ShortOPD: Recovering Pruned LLMs with Short-to-Long On-Policy Distillation
• Abstract summary: textscpass@$k$ nearly vanishes after compression, yet textscpass@$k$ recovers substantially under repeated sampling.<n>Recovery should train on the compressed model's own on-policy states with dense token-level supervision.<n>We propose textbfshortopd, a short-to-long OPD schedule that detects teacher-confirmed repetitive suffixes.
• Score: 72.54604107217669
• License: http://arxiv.org/licenses/nonexclusive-distrib/1.0/
• Abstract: Structured pruning is a hardware-friendly way to compress LLMs, but it is mostly validated on multiple-choice recognition tasks, while the same compressed checkpoints can collapse on the free-form generation that deployment actually requires. Two observations trace this gap. First, greedy \textsc{pass}@$1$ nearly vanishes after compression, yet \textsc{pass}@$k$ recovers substantially under repeated sampling: useful generations are demoted, not erased. Second, the recoverable regime fails mainly through suffix repetition. Recovery should therefore train on the compressed model's own on-policy states with dense token-level supervision, which On-Policy Distillation (OPD) provides by reusing the pre-compression model as a frozen teacher. However, long on-policy rollouts spend early recovery budget on low-information repetitive suffixes, delaying loss descent. To mitigate this waste, we propose \textbf{\shortopd}, a short-to-long OPD schedule that detects teacher-confirmed repetitive suffixes, treats the surviving prefix as each rollout's effective length, and allocates future rollout budgets to the effective lengths the policy can currently use. Across math, code, and open-ended generation, \shortopd\ raises the compressed model's score to about $9\times$ its unrecovered value and $1.6$--$4.4\times$ standard recovery recipes (SFT w/o KD, KD, and SeqKD), and it matches a fixed $8192$-token rollout horizon within two points using a quarter of the training time ($8.5$ vs.\ $35.9$ hours) and $71\%$ fewer rollout tokens. We hope this recipe helps move structured pruning beyond marginal gains on perplexity and multiple-choice benchmarks, a step closer to deployment-ready generation quality.

Related papers

• TRACE: Trajectory-robust Admission with Evidence Ordering for Efficient GUI Agents [54.88208747581851]
  Training-free visual token pruning can reduce this cost, but cache reuse introduces a fundamental constraint.<n>We propose textbfmethod, a training-free framework for emphtextbfTrajectory-textbfrobust textbfAdmission and textbfCoverage-aware textbfEvidence ordering.
  arXiv  Detail & Related papers  (2026-09-09T15:12:48Z)
• Rollout-Decoded Reconstruction for Long-Horizon Prediction in Latent World Models [0.9740025522928777]
  A latent world model trains its decoder on latents anchored to observations, then deploys it on the model's own free-running rollout, hundreds of steps past the last observation.<n>Rollout-Decoded Reconstruction (RDR) closes this gap with a single loss term that free-runs the model during training exactly as evaluation will, decodes every rollout latent, and penalizes reconstruction error against ground truth.
  arXiv  Detail & Related papers  (2026-08-25T18:05:24Z)
• DeepLoop: Depth Scaling for Looped Transformers [91.7922038545625]
  Looped Transformers scale sequential computation by applying a compact stack of physical blocks for multiple rounds.<n>We formalize this tied-depth effect through a first-order perturbation bound controlled by a visit-alignment coefficient.<n>The resulting method, textbfDeepLoop, keeps the Post-LN DeepNorm architecture and sets $= (2N)1/2$ and $=(8N)-1/2$ for unrolled depth.
  arXiv  Detail & Related papers  (2026-07-15T06:37:05Z)
• PARTREP: Learning What to Repeat for Decoder-only LLMs [12.053018926036678]
  PartRep is a selective augmentation method that appends only the most informative tokens.<n>We train a lightweight gate that predicts high-NLL tokens from early-layer hidden states.<n>PartRep retains most of the gains of full repetition while using only 59.4% of its KV cache and 79.0% of its prefill FLOPs.
  arXiv  Detail & Related papers  (2026-07-02T07:07:28Z)
• Entropy Gate: Entropy Quenching for Near-Lossless Token Compression in LLM Pipelines [0.0]
  LLM pipelines waste substantial token budgets on low-information content.<n>We introduce Entropy Gate, a token compression framework applying entropy quenching.<n>The framework is stateless, model-agnostic, and deploys as an OpenAI-compatible HTTP proxy.
  arXiv  Detail & Related papers  (2026-06-02T14:55:02Z)
• Shortcuts in the Tail: Debiasing via Post-Hoc Spectral Compression of Fine-Tuning Updates [0.47963487028876156]
  Fine-tuning often introduces spurious correlations alongside task knowledge, causing systematic failures on underrepresented groups.<n>We show a simple post-hoc intervention reduces shortcut reliance without any of these: truncating the tail of the SVD of $W = W_mathrmft - W_mathrmbase$ reduces the spurious-group gap while preserving task accuracy.
  arXiv  Detail & Related papers  (2026-05-29T05:18:25Z)
• From Patches to Trajectories: Privileged Process Supervision for Software-Engineering Agents [56.31499185764872]
  Supervised fine-tuning (SFT) on long teacher trajectories is the dominant way to instill investigation and reasoning in open software-engineering (SWE) agents.<n>We propose Patches-to-Trajectories (P2T), which uses $pstar$ as privileged information during curation and formulates trajectory construction as bi-objective optimization over per-step effectiveness and trajectory length.
  arXiv  Detail & Related papers  (2026-05-21T04:54:55Z)
• Representation Without Reward: A JEPA Audit for LLM Fine-Tuning [1.2691047660244335]
  Joint-embedding predictive architectures (JEPAs) propose that a model should learn more useful abstractions when trained to predict latent representations rather than observed outputs.<n>For autoregressive language-model fine-tuning the principle entails a stricter requirement: the induced hidden-state geometry must reach the language-model head emphand improve the decoded task metric.<n>We test that requirement under a fixed Llama-3.2-1B-Instruct LoRA harness on natural-language-to-regex generation.
  arXiv  Detail & Related papers  (2026-05-14T20:27:32Z)
• RACER: Retrieval-Augmented Contextual Rapid Speculative Decoding [80.12789199134511]
  Autoregressive decoding in Large Language Models (LLMs) generates one token per step, causing high inference latency.<n>We propose $textbfRACER, a lightweight and training-free method that integrates retrieved exact patterns with logit-driven future cues.<n>Experiments on Spec-Bench, HumanEval, and MGSM-ZH demonstrate that RACER consistently accelerates inference, achieving more than $2times$ speedup over autoregressive decoding.
  arXiv  Detail & Related papers  (2026-04-16T11:23:55Z)
• SkipKV: Selective Skipping of KV Generation and Storage for Efficient Inference with Large Reasoning Models [25.509962883211]
  Large reasoning models (LRMs) often cost significant key-value (KV) cache overhead, due to their linear growth with the chain-of-thought (CoT) reasoning process.<n>We present textbfSkipKV, a KV compression method for selective textiteviction and textitgeneration operating at a coarse-grained sentence-level sequence removal.
  arXiv  Detail & Related papers  (2025-12-08T19:32:06Z)
• From Continual Learning to SGD and Back: Better Rates for Continual Linear Models [50.11453013647086]
  We analyze the forgetting, i.e., loss on previously seen tasks, after $k$ iterations.<n>We develop novel last-iterate upper bounds in the realizable least squares setup.<n>We prove for the first time that randomization alone, with no task repetition, can prevent catastrophic in sufficiently long task sequences.
  arXiv  Detail & Related papers  (2025-04-06T18:39:45Z)
• WR-ONE2SET: Towards Well-Calibrated Keyphrase Generation [57.11538133231843]
  Keyphrase generation aims to automatically generate short phrases summarizing an input document. The recently emerged ONE2SET paradigm generates keyphrases as a set and has achieved competitive performance. We propose WR-ONE2SET which extends ONE2SET with an adaptive instance-level cost Weighting strategy and a target Re-assignment mechanism.
  arXiv  Detail & Related papers  (2022-11-13T09:56:24Z)

This list is automatically generated from the titles and abstracts of the papers in this site.

This site does not guarantee the quality of this site (including all information) and is not responsible for any consequences.