Source-linked AI summary

On-Policy Distillation Meets Off-Policy GRPO: Training Compact Instruction-Following Rerankers

Vignesh Prabhakar, Jialing Pan, Anil Babu Ankisettipalli

arXiv:2609.01947v1cs.LGcs.AI

TL;DR

Offline distillation limits compact rerankers to fixed teacher rankings, motivating a method that combines off-policy teacher strengthening with on-policy student exploration. The resulting reward-based distillation performs best across MAIR tasks and transfers across student architectures, within the study’s evaluated scope.

  • Problem

    Offline imitation on fixed examples restricts compact reranker supervision to the teacher’s observed ranking space, motivating student-driven exploration under distribution shift.

  • Method

    A 4B teacher is strengthened with off-policy GRPO, then a 1B student samples its own rankings and learns from soft teacher-derived rewards using on-policy GRPO.

  • Results

    Across all 126 MAIR tasks, the proposed method obtains the highest task-macro point estimates among evaluated distillation variants and also improves three architecturally distinct student backbones.

  • Takeaways & Limitations

    The results support combining student-driven sampling with permutation-level teacher rewards rather than offline pairwise transfer or on-policy teacher-distribution matching.

  • Takeaways & Limitations

    The study is limited to English-language, text-only reranking, uses one strengthened teacher family, and has mostly single-seed ablations.

Abstract

from arXiv · show

Compact instruction-following rerankers are attractive for deployment, but conventional distillation pipelines typically train students by offline imitation of teacher outputs on a fixed set of examples, constraining supervision to the teacher's observed ranking space. We revisit reranker distillation through the lens of reinforcement learning. We propose a two-stage framework combining off-policy teacher optimization with on-policy student distillation. In Stage 1, a 4B teacher reranker is strengthened with off-policy GRPO using LLM-judge feedback on 88K instruction-following examples. In Stage 2, a compact 1B student samples rankings from its own policy and receives soft teacher-derived rewards on those rankings, coupling student exploration with knowledge transfer. Our strongest gains appear under distribution shift. On MAIR-11, the original 11-subset, 869-query evaluation, the proposed student reaches 0.7670 nDCG@6, outperforming offline listwise KD by +4.6 points. Controlled comparisons against offline pairwise RankNet KD and on-policy GKD show that neither changing the offline distillation objective nor moving teacher-distribution matching on-policy reproduces the performance of reward-based on-policy distillation over student-sampled rankings. The advantage persists on MAIR-Full: across all 126 tasks and 9,356 queries, the proposed method obtains the highest task-macro point estimates among the evaluated distillation variants, reaching 0.6808 nDCG@6 and 0.7865 MRR@6. It also exceeds two released 7B RL-trained rerankers on the comparable MAIR-11 evaluation, while the same Stage 2 training procedure consistently improves three architecturally distinct alternative student backbones. On the 9,861-query validation benchmark, the resulting 1B reranker achieves 0.7624 nDCG@6 while providing a favorable quality-efficiency tradeoff relative to larger alternatives.

1 Introduction

The paper addresses compact instruction-following reranking by replacing fixed-distribution imitation with student-driven, reward-based on-policy distillation. Its two-stage framework strengthens a teacher and transfers capability across distribution shifts and architectures.

  • Existing compact-reranker distillation typically imitates teacher labels or scores offline on fixed examples, limiting supervision to the teacher’s observed ranking space.
  • The proposed pipeline strengthens a 4B teacher with off-policy GRPO and distills it into a 1B student using on-policy GRPO with soft teacher rewards.
  • The student samples rankings from its own policy, receiving teacher supervision as soft rewards over ranking permutations rather than token-level distributions.
  • +4.6 nDCG@6 points separate the proposed method from offline listwise KD on MAIR-11, with controls ruling out offline pairwise and on-policy distribution matching as equivalent explanations.
  • Across MAIR tasks, alternative student architectures, and larger released rerankers, the method shows broad OOD and architectural generalization.
  • 0.7624 nDCG@6 is achieved by the resulting 1B reranker on the 9,861-query validation benchmark while reducing latency relative to 4B alternatives.

2 Related Work

Prior work spans instruction-following reranking, prompted large-model ranking, reinforcement-learning alignment, and offline retrieval distillation. The paper positions its contribution at their intersection by applying reward-based on-policy learning to compact ranking models.

  • Instruction-following rerankers condition ranking decisions on natural-language criteria, extending neural reranking beyond implicit query intent.
  • Prompted and open-source LLM rerankers address ranking at inference time, whereas this work trains a compact 1B reranker for deployment.
  • RL methods such as GRPO optimize ranking behavior, while prior offline-RL perspectives emphasize the distributional coverage limits of fixed data.
  • Retrieval distillation commonly transfers fixed teacher scores offline, although soft supervision can sometimes let students outperform teachers through regularization.
  • The paper instead treats distillation as policy learning: students sample rankings from their own policies and receive teacher-derived rewards.
  • The method combines on-policy teacher-distribution distillation with reward-signal optimization, two directions previously treated separately in reranking.

3 Method

The method uses two GRPO stages over Plackett–Luce rankings: teacher optimization with judge rewards, followed by student-policy exploration evaluated with soft teacher rewards. Stabilization terms are secondary to this policy-gradient signal.

  • Model naming conventions: ZeRank-2 is the 4B teacher backbone, Base-1B is the Llama-Nemotron-Rerank-1B-v2 backbone, and Distilled-1B is the Stage 2 student.
  • Stage 1 — Teacher Training (off-policy GRPO): Stage 1 trains the teacher with off-policy GRPO using sampled rankings and LLM-judge rewards normalized into group-relative advantages.
  • Stage 2 — Student Distillation (on-policy GRPO): Stage 2 samples rankings from the student’s own Plackett–Luce policy parameterized by candidate scores.
  • Stage 2 — Student Distillation (on-policy GRPO): The fixed teacher converts scalar utilities into rank-normalized soft relevances and scores each student-sampled permutation with nDCG@6.
  • Stage 2 — Student Distillation (on-policy GRPO): Teacher rewards are transformed into group-normalized advantages, which weight the student’s policy-gradient update.
  • Stage 2 — Student Distillation (on-policy GRPO): The teacher evaluates student-generated rankings rather than generating the rankings used for student training.
  • Stage 2 — Student Distillation (on-policy GRPO): KL and entropy regularizers use first-selection marginals, avoiding direct regularization over the full combinatorial permutation distribution.
  • 3.4 Hypothesis: Why On-Policy Distillation Generalizes: The central hypothesis is that student-policy supervision expands effective training coverage, producing a larger advantage under distribution shift.

4 Experimental Setup

The experiments use broad instruction-following retrieval data, fixed candidate pools, multiple reranking baselines, and complementary MAIR evaluations. Training combines off-policy teacher optimization with on-policy student distillation, using judge-validated rewards and standard ranking metrics.

  • Datasets: The training split contains 88,755 examples, while validation contains 9,861 queries across web, code, mathematics, news, and multi-hop retrieval.
  • Datasets: MAIR-11 covers 11 heterogeneous subsets and 869 queries for controlled ablations and external-model comparisons.
  • Datasets: MAIR-Full evaluates all 126 MAIR tasks and 9,356 queries using task-macro averages because task sizes vary substantially.
  • Evaluation protocol: Candidate pools are frozen and reused across systems, while model inputs are truncated to 512 tokens to prevent retrieval variation from confounding comparisons.
  • Models and baselines: The study compares a 4B ZeRank-2 teacher and 1B Llama-Nemotron student across supervised, offline, off-policy, on-policy, pairwise, and external reranking baselines.
  • Training: Stage 1 uses off-policy GRPO, while Stage 2 uses on-policy GRPO with teacher-derived soft rewards, KL regularization, entropy regularization, and three training epochs.
  • Reward validation: The Stage 1 judge agrees with human relevance annotations on 92.6% of examples, with Cohen’s κ = 0.84, while five judges show high inter-judge agreement.
  • Metrics: Primary metrics are nDCG@6 and MRR@6, with query-micro averages for validation and MAIR-11 and task-macro averages for MAIR-Full.

5 Main Results

The proposed reward-based on-policy student performs especially well under distribution shift, outperforming offline and distribution-matching distillation controls on MAIR-11 and MAIR-Full. It also exceeds larger released RL-trained rerankers, while repairing concentrated teacher failures rather than uniformly surpassing the teacher.

  • Validation benchmark: 0.7624 nDCG@6 is achieved by Distilled-1B on the 9,861-query validation benchmark, the best overall nDCG@6 among evaluated models.
  • MAIR-11: 0.7670 nDCG@6 is achieved by A3 on MAIR-11, outperforming offline listwise KD by 4.6 points.
  • Teacher behavior: Teacher GRPO reaches 0.7422 nDCG@6 on validation but 0.6880 on MAIR-11, with concentrated failures on LitSearch and Touche.
  • Teacher behavior: A3 does not uniformly outperform its teacher; its aggregate improvement partly comes from repairing the teacher’s concentrated out-of-distribution failures.
  • MAIR-11: A3 outperforms off-policy sampling, offline listwise KD, on-policy GKD, RankNet pairwise KD, and distillation from the unaligned teacher on MAIR-11.
  • External comparisons: 0.7670 nDCG@6 and 0.8289 MRR@6 allow A3 to exceed Rank-R1-7B and REARANK-7B on the comparable MAIR-11 evaluation.
  • MAIR-Full: 0.6808 nDCG@6 and 0.7865 MRR@6 are A3’s highest task-macro point estimates among evaluated distillation variants on MAIR-Full.

6 Ablation Study

The ablations identify student-driven on-policy distillation with soft teacher rewards as the primary source of robust OOD gains, while teacher strengthening is necessary and auxiliary regularizers are secondary. Across validation and MAIR evaluations, reward-based on-policy distillation outperforms alternative sampling and supervision choices, with robustness–fit tradeoffs under distribution shift.

  • Sampling policy and supervision form: +4.6 nDCG@6 points over offline listwise KD on MAIR-11, reward-based on-policy distillation outperforming both offline pairwise KD and on-policy teacher-distribution matching.The four-way comparison reports validation nDCG@6 values of 0.7608, 0.7514, 0.7476, and 0.7624 for A1, A10, A9, and A3; MAIR-11 values are 0.7212, 0.7416, 0.7386, and 0.7670.
  • Where supervision is evaluated: Teacher-driven off-policy sampling consistently degrades performance versus student sampling, isolating student exploration as the key mechanism when teacher-derived supervision is held fixed.The comparison between A2 and A3 supports evaluating supervision on the student’s own sampled ranking space.
  • Teacher strengthening: Stage 1 teacher strengthening is necessary: on-policy learning from the unaligned teacher improves over alternatives but remains well below A3 on MAIR-11.The result attributes the full two-stage gain to combining student-side on-policy learning with prior instruction-aware teacher optimization.
  • Soft rewards: Hard-label supervision achieves the best validation performance among core variants but substantially reduces MAIR-11 performance, exposing an in-distribution-fit versus OOD-robustness tradeoff.Soft rewards preserve relative preference information, whereas hard labels fit the in-distribution benchmark more aggressively.
  • Auxiliary terms and robustness: Removing KL or entropy leaves performance close to the full objective, while hyperparameter sweeps remain nearly flat on validation and KL has modest OOD sensitivity.These findings position policy-gradient training with soft teacher rewards as dominant and KL and entropy as secondary regularizers.

7 Efficiency Analysis

The efficiency analysis compares single-query latency under fixed hardware and finds that the distilled 1B reranker offers a favorable quality–efficiency tradeoff relative to larger rerankers.

  • Table 4 compares single-query inference latency under a fixed hardware setup.
  • 27.0ms to 9.2ms: the distilled 1B student reduces mean latency relative to the teacher while improving validation performance.
  • The distilled 1B reranker remains competitive with or stronger than evaluated baselines on ranking quality while being substantially faster than 4B rerankers.

8 Discussion and Conclusion

The discussion attributes the strongest distillation behavior to student-driven sampling with permutation-level teacher rewards, and reports broad performance and architectural generalization. It also distinguishes the contributions of teacher strengthening and soft reward supervision.

  • Student-driven sampling with permutation-level teacher rewards produces stronger OOD distillation behavior than on-policy teacher-distribution matching or offline pairwise distillation.The A2–A3 comparison isolates the role of evaluating supervision on the student’s own ranking distribution.
  • Across all 126 MAIR tasks, A3 obtains the highest task-macro point estimates among the evaluated distillation variants.
  • A3 exceeds the evaluated 7B Rank-R1 and REARANK models on comparable MAIR-11 evaluations despite its substantially smaller size.
  • Teacher strengthening and soft reward supervision are complementary: unaligned-teacher distillation remains below the full method under distribution shift, while hard-label training generalizes less robustly.
  • The same Stage 2 procedure improves BGE-Reranker-v2-Gemma, MXBAI-Rerank-Large-v1, and RankZephyr-7B, supporting transfer across three distinct student architectures.

9 Limitations

The study is bounded to English-language, text-only reranking and relies on LLM-judge rewards with limited human validation. Teacher-side generality and some ablation comparisons remain constrained.

  • The study focuses on English-language, text-only reranking.
  • LLM-judge rewards may inherit judge bias and calibration error, despite validation against human relevance annotations.That validation contains only 188 examples.
  • Most controlled ablations use a single random seed, so small differences among closely clustered objective variants require cautious interpretation.Additional multi-seed evidence is provided for A3, A4, and A5.
  • Teacher-side architectural and capability generality remains open because all experiments use the same strengthened ZeRank-2 teacher.The study establishes student-side architectural generality across three student families.
  • The Stage 1 teacher is not uniformly robust: aggregate MAIR-11 degradation is concentrated in LitSearch and Touche.

10 Ethical Considerations

The paper identifies bias, misranking, and judge-calibration risks in instruction-following reranking, while reporting broad empirical robustness and cross-architecture gains. These findings are bounded by the English text-only evaluation setting and a single strengthened teacher family.

  • Ethical considerations: Judge-derived rewards may inherit calibration errors or systematic preferences, potentially amplifying bias or under-ranking relevant material for some domains or groups.The paper partially mitigates this risk through comparisons with other LLM evaluators and human annotations, but does not eliminate all bias.
  • MAIR-Full results: A3 achieves the highest task-macro point estimates among evaluated distillation variants across all 126 MAIR tasks.MAIR-Full contains 9,356 queries, extending the broader distribution-shift evaluation beyond MAIR-11.
  • Distribution-shift analysis: The Stage 1 teacher is stronger than Base-1B on 7 of 11 MAIR-11 subsets but exhibits concentrated weaknesses on farther-shifted domains.The largest reported weakness is on LitSearch, rather than a uniform degradation across subsets.
  • Cross-architecture generalization: Stage 2 improves all three alternative student backbones on validation and preserves these gains under MAIR-11 distribution shift.The experiments use the same strengthened ZeRank-2 teacher across architectures.
  • Robustness: Validation performance remains nearly flat across hyperparameter sweeps, while MAIR-11 shows modest sensitivity, especially to the KL coefficient.The sweep results suggest the reported gains are not dependent on brittle tuning around one configuration.

F.2 MAIR-11 Results

On MAIR-11, the method remains broadly stable across hyperparameter settings, with somewhat greater sensitivity to the KL coefficient than on validation. The benchmark contains 869 queries across 11 subsets.

  • KL sensitivity: Larger KL values improve MAIR-11 performance, although the KL term is characterized as an auxiliary stabilizer rather than the main learning source.MAIR-11 is more sensitive to KL than the validation set.
  • Evaluation setting: MAIR-11 contains 869 queries across 11 subsets.The evaluation is reported as a distribution-shift benchmark in the surrounding results analysis.
  • Hyperparameter robustness: Validation performance is nearly flat across the tested hyperparameter sweeps, while MAIR-11 shows only modest variation.Red stars in Figure 5 mark the default configuration used for the main results.

G Per-Dataset Validation Breakdown Across All Baselines

Across validation datasets and comparisons with larger rerankers, A3 offers a quality-efficiency tradeoff rather than uniform dominance. Its main advantage over offline KD emerges under distribution shift, while objective variants remain closely clustered.

  • Per-dataset results: A3 remains competitive across all eight validation datasets and obtains the highest or tied-highest point estimate on WebQA.Jina is strongest on MS MARCO and Robust04, while Cohere Rerank v4.0-fast is strongest on MetaMath and LeetCode.
  • Cross-model validation: 0.762 nDCG@6 is achieved by A3 across the full 9,861-query validation benchmark, versus 0.750 for Rank-R1-7B and 0.748 for REARANK-7B.A3 uses a substantially smaller model, while the larger baselines remain stronger on some individual subsets.
  • Objective components: A3, A4, and A5 remain tightly clustered, with no consistent winner across MAIR-11 subsets.This supports treating policy-gradient learning with soft teacher rewards as primary, with KL and entropy as auxiliary stabilizers.
  • Paired effects: A3’s paired advantage over A1 is small on validation but substantially larger on MAIR-11.The validation difference is directionally reliable but practically negligible according to the reported paired analysis.
  • Qualitative analysis: A3 often improves ranking by distinguishing among highly similar candidates when the relevant passage is already present.Many observed failures involve plausible candidate sets whose best instruction- or claim-aligned passage is ranked below distractors.
  • Distribution shift: 4.6 points is A3’s nDCG@6 gain over A1 on MAIR-11, increasing from 0.7212 to 0.7670.MRR@6 also increases from 0.7860 to 0.8289, a 4.3-point gain.
  • Judge validation: Judge outputs show strong inter-judge consistency and 92.6% agreement with human labels on 174 of 188 manually labeled examples.The reported Cohen’s κ is 0.84, while the paper still treats judge outputs as approximations rather than substitutes for human annotation.
Loading 2609.01947v1…