Flux-OPD: evolving contexts and conflict-weighted distillation¶

Fidelity: Implements a context-free anchor, geometric teacher, contextual correction, and conflict weighting.

Paper information¶

Field	Details
Paper	Flux-OPD (arXiv 2607.28022)
Organization	Peking University / Kling Team / Tsinghua University / Shanghai Jiao Tong University
First public release	2026-07-30
Authors' code	No official author implementation was found when checked on 2026-08-01.
Adapter / method	`flux-opd`
Local reproduction code	`src/auto_research/post_training/`

Original paper¶

Motivation and method¶

Students absorb fixed contexts quickly, but replacing a contextual teacher outright produces discontinuous targets. Flux-OPD holds the context-free teacher fixed as an anchor and injects only log-probability differences from multiple evolving contextual teachers. The geometric-mean normalization constant captures conflict; higher conflict weakens the correction.

flowchart LR
    A["Context-free anchor"] --> C["Log-space correction"]
    E["Evolving contexts"] --> G["Geometric teacher"]
    G --> D["Conflict δ"]
    G --> C
    D --> W["Correction weight λ"]
    W --> C
    C --> O["Reverse-KL OPD"]

Key figure from the paper¶

Key figure: Original motivation diagram. From the original paper; copyright remains with the authors.

Core equations¶

\[ q_k^{\mathrm{flux}}=\operatorname{softmax}(\log q_0+\lambda_k(\log q_{\mathrm{geo},k}-\log q_0)),\quad \delta_k=-\log Z_k. \]

Paper results¶

In video prompt optimization, the student averages VBench 80.18 versus 79.28 for OPD, and Video-Bench 3.61 versus 3.54. HealthBench also beats the listed OPD approaches. No production A/B is reported.

Local reproduction¶

Local comparison: Same GSM8K candidate policy, 120 steps, seed 42. Accuracy goes from 0.1719 to 0.8438 (+390.91%).

auto-research post-train --algorithm flux-opd --dataset gsm8k-candidate --maximum-examples 256 --steps 120 --seed 42

See the fixed metrics.

Reproduction boundaries¶

Two local reward axes stand in for evolving contextual teachers. No video model, medical teacher, or large-model context extraction is called.