AIAny
Icon for item

FlowBalance: Verifier-Grounded Self-Improvement from On-Policy Reasoning Experience

Combines sparse verifier outcomes with dense privileged‑hindsight token scoring to learn an outcome‑calibrated, normalized distribution over complete responses for on‑policy self‑improvement. Key features: sign‑gated guidance (retain/reverse/disable per verifier advantage), profiled trajectory balance with one log‑partition per rollout group, and explicit correction against false‑positive self‑guidance.

Introduction

Why this matters

Most on‑policy self‑improvement loops for reasoning models suffer from a tension: terminal verifiers give sparse but reliable rewards, while dense same‑model guidance can accidentally reinforce false confidence or collapse solution diversity. FlowBalance targets that tension by making dense within‑trajectory guidance conditional on verifier outcomes and converting the composite signal into a normalized next‑policy target via trajectory balance.

Key Findings
  • Outcome‑calibrated dense guidance: token‑level log‑probability gains from a frozen privileged‑hindsight view are aggregated per trajectory and sign‑gated by the verifier advantage (retain on positive advantage, reverse on negative, disable when advantage is zero), preventing confident failures from becoming self‑reinforcing.
  • Profiled trajectory balance: the composite energy (verifier + guided gain) is converted into a normalized distribution over complete responses with one log‑partition estimate per rollout group, keeping within‑group probability contrasts and limiting drift from a reference policy.
  • Empirical gains on math reasoning: improves average performance over FlowRL on Qwen3‑4B and Qwen3‑8B (core four‑benchmark average +1.67 points on Qwen3‑4B and +1.98 on Qwen3‑8B), reaches 0.50 AIME24 validation accuracy in ~100 steps on Qwen3‑8B versus ~143 for GRPO, and avoids OPSD’s response‑length collapse while increasing correct‑strategy diversity.
Who it's for and tradeoffs

Great fit if you work on iterative self‑improvement for reasoning LLMs and need a principled way to combine sparse terminal verification with dense within‑trajectory evidence without imposing a separate token‑level imitation loss. It is particularly relevant for research or training setups that can generate rollout groups and run verifiers (e.g., mathematical or code reasoning benchmarks).

Look elsewhere if your setting lacks reliable terminal verifiers, cannot store or compute privileged hindsight scores, or if per‑rollout grouping and log‑partition estimates are impractical for your compute budget. FlowBalance shapes which response distribution is learned via a dense branch that is not optimized as a standalone imitation objective, so its benefits depend on verifier quality and rollout group diversity.

Method snapshot

FlowBalance builds a composite trajectory energy E(y) = η_A·A(y) + β_G·G_H(y)·sgn(A(y)), where A(y) is the verifier advantage and G_H(y) the privileged‑hindsight trajectory gain. That energy reweights a reference policy into a Gibbs target; profiled trajectory balance then fits the normalized target with one log‑partition per rollout group, yielding a minimum‑change reverse‑KL displacement from the reference at the attained energy level.

Information

  • Websitearxiv.org
  • AuthorsZixun Huang, Kishan Panaganti, Haitao Mi, Leowei Liang
  • Published date2026/09/03

More Items

Selectively admits dense token-level teacher supervision only after a prompt-level verifier audit, routing prompts that fail the audit to verifier-grounded trajectory supervision instead — reducing harmful updates from confidently wrong teachers and improving teacher GPU utilization.

Analyzes on-policy self-distillation for language-model reasoning, diagnosing “collapse” and framing it as controlled by three levers: where token-level signals apply, what privileged information the teacher sees, and how teacher dynamics evolve.

Uses lightweight discrete-diffusion adapters to draft blocks of tokens in parallel while keeping the original autoregressive model as the final arbiter, enabling lossless speedups (up to 3×) and compatibility with existing AR weights via a Diffusion Distillation phase.