English static mirror for SEO/GEO · AI-assisted translation · Read Chinese original

Self-Distilled RLVR: RLSD Decouples Direction and Magnitude for Token-Level Credit Assignment in GRPO

Forum topic · 小凯 · 2026-05-25

Summary

A team from the Chinese Academy of Sciences, UCAS, Microsoft Research Asia, and JD presents RLSD, a self-distilled RLVR framework that fixes GRPO's coarse sequence-level credit assignment. The paper first analyzes why on-policy self-distillation (OPSD) fails: its objective contains an irreducible conditional mutual information term, forcing students to match privileged information and eventually causing reference-solution leakage and accuracy collapse. RLSD's key insight is that update direction and magnitude have asymmetric requirements: direction must come from sparse but reliable verifier rewards, while magnitude can be dense. RLSD computes a privileged information gain per token (log-ratio between teacher and student probabilities) via one extra forward pass, converts it into a positive Bayesian evidence ratio weight, applies sign-aware reweighting, and clips it within a trust region so verifier rewards retain exclusive control of update direction. Experiments on Qwen3-VL-8B-Instruct across MMMU, MathVista, MathVision, ZeroBench, and WeMath show RLSD averaging 56.18, beating GRPO by +2.32 and OPSD-based alternatives, with faster convergence, higher entropy retention, and token-level credit visualization. The paper claims a zero-leakage guarantee satisfying objective stability, sustained improvement, and no leakage.

Self-Distilled RLVR: Using Self-Distillation as GRPO's Credit Assignment Officer

> Paper: Self-Distilled RLVR (arXiv:2604.03128) > Authors: Chenxu Yang, Chuanyu Qin, Qingyi Si, et al. > Institutions: Institute of Information Engineering, CAS; University of Chinese Academy of Sciences; Microsoft Research Asia; JD Group > Link: https://arxiv.org/abs/2604.03128

1. The Problem: GRPO's "Uniform Payroll" Reward

GRPO suffers from unclear credit assignment. For a reasoning chain of hundreds of tokens, the verifier only returns a 0/1 sequence-level signal, meaning:

  • The key reasoning step leading to a correct answer gets the same reward weight as the trailing "Therefore, the answer is..." boilerplate
  • In a wrong answer, the actual erroneous step is penalized equally with correct earlier steps
  • This sequence-level sparse signal undermines convergence for long chain-of-thought reasoning.

    2. The OPSD Trap: Why On-Policy Self-Distillation Collapses Late in Training

    OPSD design: the same model acts as both teacher and student; the teacher receives privileged information (e.g., the reference answer) and provides dense per-token logits supervision.

    Observed pathologies (Figure 3):

  • Frequency of referencing "reference solution" monotonically increases during training (leakage)
  • Validation accuracy peaks at steps 10–20, then continuously declines
  • Teacher–student KL divergence stagnates after initial decline
  • Example leaked outputs: *"Given that the reference solution uses 9 values, I will assume..."*, *"But wait, the reference solution says 'No', which contradicts my calculation..."*

    Theoretical diagnosis (Theorem 1):

    \[L_{OPSD} = L_* + I(Y_t; R \mid X, Y_{<t})\]
  • L*: ideal marginal-matching objective (optimizable)
  • I(Y_t; R | X, Y_{<t}): a conditional mutual information term — an irreducible gap independent of parameters θ
  • OPSD forces a conditionally-independent student to match a conditionally-dependent teacher — an ill-posed problem. Gradient analysis (Proposition 1): g(θ; r) = g*(θ) + δ(θ; r), where δ is an r-specific bias whose variance grows with the teacher's reliance on privileged information.

    Two-phase dynamics: 1. Early: ‖g*‖ ≫ ‖δ‖ — fast learning of general reasoning, accuracy rises 2. Late: as P_S → P̄_T, g* → 0 and δ dominates — the model encodes spurious x → r correlations and leakage erupts

    3. RLSD: Decoupling Direction and Magnitude

    > "Update direction and update magnitude have asymmetric requirements."

    | Signal | Requirement | Source | |---|---|---| | Direction | Sparse but reliable | Environment reward (verifier) | | Magnitude | Denser is better | Self-distillation evidence ratio |

    Three-step algorithm

    Step 1 — Privileged information gain (stop-gradient, weights only):

    \[\Delta_t = \mathrm{sg}(\log P^T(y_t) - \log P^S(y_t))\]

    Step 2 — Sign-aware evidence reweighting:

    \[w_t = \exp(\mathrm{sign}(A) \cdot \Delta_t) = (P^T(y_t)/P^S(y_t))^{\mathrm{sign}(A)}\]

    Bayesian interpretation: the ratio is a Bayesian evidence ratio measuring how much generating y_t makes r consistent with the trajectory.

  • A > 0 (correct): teacher-supported tokens get larger weights
  • A < 0 (wrong): teacher-opposed tokens get larger penalties
  • Since exp(·) > 0, w_t never flips the sign of the advantage — the environment reward retains exclusive control of update direction; the teacher only modulates magnitude.

    Step 3 — Clipped credit assignment:

    \[\hat{A}_t = A \cdot \mathrm{clip}(w_t, 1-\epsilon_w, 1+\epsilon_w)\]

    with ε_w = 0.2, plus a λ schedule decaying from 0.5 to 0 over the first 50 steps.

    Overhead: only one extra forward pass per response — negligible compared to rollout generation.

    4. Experiments: Multimodal Reasoning on Qwen3-VL-8B

    Setup: Qwen3-VL-8B-Instruct; MMFineReason-123K difficulty-filtered subset (samples where Qwen3-VL-4B-Thinking failed all 4 rollouts); 4 nodes × 8 × NVIDIA H200 140GB.

    | Method | MMMU | MathVista | MathVision | ZeroBench | WeMath | Avg | |---|---:|---:|---:|---:|---:|---:| | Base LLM | 62.44 | 73.80 | 47.37 | 19.76 | 54.10 | 51.49 | | GRPO | 65.11 | 76.20 | 48.82 | 22.60 | 56.57 | 53.86 | | OPSD | 63.82 | 75.10 | 47.53 | 21.06 | 54.95 | 52.49 | | SDPO | 65.11 | 74.00 | 47.27 | 25.15 | 52.19 | 52.74 | | GRPO+OPSD | 63.22 | 75.90 | 48.52 | 22.16 | 54.76 | 52.91 | | RLSD (Ours) | 67.22 | 78.10 | 52.73 | 24.85 | 58.00 | 56.18 |

    Key findings:

  • RLSD achieves the best average, +4.69 over base and +2.32 over GRPO; math gains are largest (MathVision +3.91)
  • OPSD and SDPO both underperform GRPO, validating the leakage theory; naive GRPO+OPSD combination also fails
  • Training dynamics: RLSD converges faster and higher; GRPO shows rapid entropy collapse while RLSD retains higher entropy; clip ratio stays at 3%–6%
  • Token-level visualizations confirm sensible credit: key steps (identifying critical objects, final subtraction) get high credit; generic narration gets little; in wrong trajectories, the actual misreading gets the largest penalty
  • 5. Why RLSD Does Not Leak

    | Method | Token-level advantage | Direction | Magnitude | |---|---|---|---| | GRPO | A (uniform) | Environment reward | None (uniform) | | OPSD | Δ_t | Teacher preference | Teacher preference | | RLSD | A · clip(w_t, ...) | Environment reward | Teacher evidence ratio |

    RLSD satisfies all three conditions of the impossibility trilemma (Appendix A.6): objective stability, sustained improvement, and zero-leakage training — by moving P^T/P^S from gradient direction to gradient magnitude.

    6. Assessment: Elegant Recombination, Not a New Paradigm

    RLSD invents no new components: GRPO's sequence-level advantage, OPSD's self-distillation signal, and PPO-style clipping are all retained. It merely reassigns their roles — the environment reward remains the judge of correctness, while self-distillation becomes the recorder that determines how much reward or punishment each token receives. Distillation acts as RL's co-pilot: direction stays with the verifier, magnitude gets finer-grained.

    7. Limitations and Follow-ups

  • Validated only on Qwen3-VL-8B; generalization to other architectures (e.g., text-only LLMs) unverified
  • Privileged information limited to ground-truth answers; richer forms (intermediate reasoning steps) unexplored
  • Absolute overhead at very large scale still to be evaluated
  • Follow-up arXiv:2605.10781 (Rebellious Student) inverts RLSD's weighting for exploration; the framework may extend to multi-turn dialogue and tool-use chains
Reference: Yang, C., Qin, C., Si, Q., et al. Self-Distilled RLVR. arXiv:2604.03128.

Tags

#rlvr#self-distillation#grpo#credit-assignment#reinforcement-learning#multimodal-reasoning#qwen3-vl#llm-training

This page is an English static mirror generated for search and AI citation. It may be a full translation or structured summary of the Chinese original. Canonical interactive discussion lives on the Chinese page: https://zhichai.net/topic/177620752