|
Download docs/paper_reconstruction/experiment_plot_plan.md from Jack04810/agentic-rl-main: direct link, hf CLI and curl.
- Browser
- Download file 11 kB
-
https://huggingface.co/Jack04810/agentic-rl-main/resolve/main/docs/paper_reconstruction/experiment_plot_plan.md
- Command line
-
hf download hf://Jack04810/agentic-rl-main/docs/paper_reconstruction/experiment_plot_plan.md
-
curl -L -o experiment_plot_plan.md https://huggingface.co/Jack04810/agentic-rl-main/resolve/main/docs/paper_reconstruction/experiment_plot_plan.md
11 kB
| # OPD for Sub-Billion VLM 实验与图表设计计划 | |
| 更新时间:2026-07-13 | |
| 本文件只定义论文图表需要回答的问题、数据字段和准入条件。实验命令与 run 状态见 | |
| `docs/opd_experiment_plan.md`,已完成数字见 `experiment_ledger.md`。 | |
| ## 1. 图表叙事原则 | |
| 论文的证据链按以下顺序展开: | |
| 1. matched no-OPD 对照证明 OPD 在统一 4epoch 预算下具有净收益。 | |
| 2. OPD-only、GRPO-only、fallback-only 与联合训练证明信号互补性。 | |
| 3. unconditional OPD 与 verifier-routed OPD 证明可靠性路由不是额外 teacher loss 的包装。 | |
| 4. 全局控制器和动作拆分解释如何让 OPD 随学生自主性介入和退出,并报告 teacher compute。 | |
| 5. token-selective OPD 近邻基线与训练尺度分析界定本文方法边界。 | |
| 任何 oracle/privileged 结果都必须在主图中显式标记,不与 gold-hidden-teacher 方法合并平均。`Teacher sees gold` 与 `Verifier uses reference` 必须分列。 | |
| ## 1.1 Motivation-to-Evidence Contract | |
| 每条 motivation 必须产生一个可证伪预测。若对应实验不成立,应收缩论文 claim,不能仅用 final accuracy 覆盖机制失败。 | |
| | Motivation | CLRC prediction | Required comparison | Supporting evidence | Falsification condition | | |
| |---|---|---|---|---| | |
| | total reward 可掩盖 task signal 消失 | realized GRPO route 比 total-reward mixed/variance 更准确地反映学生自主覆盖率 | controller state: total-reward proxy vs task mixed/zero proxy vs global final GRPO route | state 与 next-window task accuracy、all-wrong、zero-loss 的相关/滞后预测;global/local disagreement | total-reward proxy 在多 seed 上同样或更好地预测 task recovery 与 final accuracy | | |
| | OPD 填补 SFT 与 RLVR 之间的学生错误状态 | matched verifier-routed OPD 优于 no-OPD | no-OPD vs verifier-routed OPD,固定其余训练流程与 teacher budget accounting | final accuracy、zero-loss、all-wrong recovery、OPD loss/coverage | matched no-OPD 相同或更好 | | |
| | student-state guidance 比 teacher-sequence imitation 更适合小模型 | OPD-only guidance 优于或至少不弱于 OPD + hard trajectory,并减少空模板塌缩 | verifier-routed OPD without hard imitation vs matched OPD + full trajectory | held-out accuracy、train accuracy、full-template/empty-skeleton/malformed-answer rates | hard trajectory 在相同预算下提高 held-out accuracy且不增加模板失败 | | |
| | OPD 与现有信号互补 | 完整联合训练优于任一单信号或两信号组合 | OPD-only、GRPO-only、fallback-only、OPD+GRPO、full | final accuracy、route occupancy、各状态分层收益 | 任一单信号在相同预算下达到同等或更高效果 | | |
| | all-wrong 中只有部分状态适合 OPD | verifier-routed OPD 优于 unconditional OPD | all-wrong 全 OPD vs verifier-confirmed OPD | accepted teacher accuracy、coverage、teacher tokens、final accuracy | unconditional OPD 在 matched compute 下同样或更好 | | |
| | fixed clock 不代表学生能力 | 改变 batch/data scale 后,CLRC 在相近 autonomy state 触发动作,而 fixed-step 在不同能力状态触发 | fixed step vs normalized progress vs CLRC,至少两种 scale | trigger step、normalized progress、GRPO EMA、task accuracy、final accuracy | CLRC trigger autonomy state 方差不低于 fixed/progress,且无 final/compute 收益 | | |
| | teacher support 应随真实自主性减少 | joint controller 在保持或提升 accuracy 时降低 teacher calls/tokens | fixed support vs OPD-weight-only vs cap-only vs joint actions | accuracy/teacher-token Pareto、last50 routes、controller trajectory | joint controller 被 fixed support Pareto 支配 | | |
| 主结果超过 baseline 但上述预测均不成立时,只能将方法描述为一个有效 recipe,不能声称 recoverability curriculum 或 closed-loop mechanism 得到验证。 | |
| ## 2. Figure 1:方法总览 | |
| ### 左侧:Completion-Level Recoverability Routing | |
| 显示同一 prompt 的 `K` 个 student completions: | |
| - correct completion -> GRPO; | |
| - wrong + gold-hidden teacher answer verified correct -> OPD; | |
| - wrong + teacher unrecoverable/quality fail -> fallback or skip。 | |
| 图中 teacher 输入只画 image、question、DePlot/visual facts 和 format constraint,不画 gold answer;teacher 输出后单独画 verifier,并明确其在当前 RLVR 设置下可访问 reference。 | |
| ### 右侧:Global Realized-Autonomy Feedback | |
| 显示各 rank 最终互斥 route counts 经 all-reduce 得到: | |
| ```text | |
| a_t = N_grpo / N_total | |
| ``` | |
| 再经 EMA、monotonic mastery、smoothstep 输出 OPD weight 和 OPD cap。图旁明确标注 | |
| 主方法的 hard-trajectory weight 恒为零;用箭头标明 step `t` snapshot 控制 step `t+1`。 | |
| ## 3. Table 1:统一主结果 | |
| 列定义: | |
| | Method | Student | Teacher | Epoch | Evidence | Teacher sees gold | Verifier uses reference | Controller | Teacher calls | Teacher tokens | ChartQA acc | Processed | | |
| |---|---|---|---:|---|---|---|---|---|---:|---:|---:|---:| | |
| 最低行集合: | |
| - Base/SFT; | |
| - DyME official; | |
| - gold-hidden-teacher PCD aligned; | |
| - gold-hidden-teacher fixed schedule; | |
| - gold-hidden-teacher recoverability-only; | |
| - gold-hidden-teacher full CLRC; | |
| - oracle student_hint_short; | |
| - oracle official; | |
| - oracle CLRC upper bound。 | |
| 主表或紧邻的核心消融表必须直接出现 matched `no-OPD` 与 | |
| `verifier-routed OPD`;否则主结果无法支撑论文标题中的 OPD claim。 | |
| 主表不使用训练 reward 替代 test accuracy。正式数字要求 `2500/2500`;若先使用 `2496/2500` 快速结果,必须在 Processed 列标记并在最终稿补齐。 | |
| ## 4. Figure 2:Accuracy / Teacher-Compute Pareto | |
| 横轴:每个训练 prompt 的 teacher generated tokens,或总 teacher generated tokens。 | |
| 纵轴:ChartQA held-out accuracy。 | |
| 点大小:GPU hours。点颜色:gold-hidden teacher、oracle、no teacher;点形状区分 verifier 是否使用 reference。比较 fixed schedule、recoverability-only、full CLRC 和 action ablations。 | |
| 该图只有在统一统计 teacher calls/tokens 后才能进入主文。若当前日志缺失总量,先放 appendix 计划,不用 candidate accuracy 代替 compute。 | |
| ## 5. Figure 3:闭环训练动力学 | |
| 使用同一横轴和相同平滑窗口,至少包含四个 panel: | |
| 1. `global_signal/grpo_route_rate`、`opd_route_rate`、`sft_route_rate`; | |
| 2. `adaptive/signal_rate`、`signal_ema`、`mastery`、`supervision`; | |
| 3. OPD weight、OPD cap,以及作为负对照记录的 trajectory weight;主 OPD recipe | |
| 中 trajectory weight 应始终为零; | |
| 4. task all-wrong、task zero-loss、total-reward zero-loss、disagreement、accuracy。 | |
| 辅助 panel:clip、EOS、degenerate 和 teacher candidate/correct/accepted funnel。 | |
| 行为 panel 额外绘制 `full_cot_template_rate`、`partial_cot_template_rate`、 | |
| `goal_without_answer_rate`、`empty_cot_skeleton_rate` 和 | |
| `malformed_answer_section_rate`。完整 CoT 比例本身不作为失败;只有空骨架或异常答案 | |
| 与完整模板共同升高时才定义为 template collapse。来自 teacher-probe candidate JSONL 的 | |
| partial/Goal-without-Answer 曲线必须标注为“conditioned on wrong probed completions”,不得 | |
| 与全体 rollout 比例混用。另以虚线画出 | |
| `teacher_traj_effective_weight` 和 `teacher_sft_repair_rate`,证明 OPD isolation run | |
| 没有重新引入 hard imitation。 | |
| 禁止只画 rank-local `routing/*` 来证明全局控制器状态;主曲线必须来自 `global_signal/*`。 | |
| ## 6. Figure 4:训练尺度稳健性 | |
| 比较 fixed-step schedule 与 CLRC,在至少两种训练尺度下记录动作发生时的: | |
| - absolute step; | |
| - normalized progress; | |
| - global GRPO route EMA; | |
| - task accuracy/all-wrong; | |
| - final accuracy。 | |
| 理想证据是 fixed-step 在不同规模下对应不同学生能力,而 CLRC 在接近的 realized autonomy state 下产生相近动作。 | |
| ## 7. Table 2:OPD 核心消融 | |
| ### A. Signal Contribution | |
| - no-OPD; | |
| - OPD-only; | |
| - GRPO-only; | |
| - fallback-only; | |
| - OPD + GRPO; | |
| - OPD + GRPO + fallback。 | |
| 这一 block 是论文最重要的机制表,优先于 controller state/action ablation。 | |
| ### B. OPD Routing | |
| - all-wrong SFT fallback; | |
| - unconditional OPD; | |
| - verifier/reward gate; | |
| - token teachability / position-weighted selective OPD; | |
| - gold-hidden-teacher multimodal recoverability(reference-verifier)。 | |
| ### B2. OPD Token Reliability | |
| - uniform-token OPD; | |
| - non-answer-heading selective OPD; | |
| - token-teachability / position-weighted selective OPD near-neighbor baseline。 | |
| 报告保留 token 比例、non-answer-heading mask rate、答案/视觉/数值 token coverage 与 | |
| final accuracy。该 block 不对模型生成结构标题施加 reward penalty,只改变 teacher | |
| distribution loss 在不同 token 位置的可信度权重。 | |
| ### C. Controller State | |
| - fixed weights; | |
| - fixed step; | |
| - normalized progress; | |
| - mixed-rate × nonzero-loss; | |
| - realized global GRPO route。 | |
| ### D. Controller Actions | |
| - OPD weight only; | |
| - OPD cap only; | |
| - OPD weight + cap。 | |
| `OPD + hard trajectory` 单独作为 supervision-type 负对照,不混入 controller action | |
| 消融,避免把已观察到的模板污染误写成主方法中的课程动作。 | |
| 每个 block 内只能改变对应因素。若多个因素同时变化,该行只能作为 recipe,不作为因果消融。 | |
| ## 8. Figure 5:Failure Taxonomy | |
| 从 eval outputs 统计并交叉: | |
| - correct/incorrect; | |
| - answer_flag/full_cot/other; | |
| - clipped/EOS; | |
| - parse failure; | |
| - empty or malformed `Answer:`; | |
| - repeated section template; | |
| - privileged tag leakage。 | |
| 上述 full/partial/empty/malformed 行为直接读取 final eval 的 `Template behavior counts`, | |
| 并与 correct/incorrect 交叉;不得用 teacher-probe candidate 条件统计替代 held-out 全体输出。 | |
| full-CoT 不是默认错误类。图中应回答“哪一种输出行为与错误相关”,而不是“推理越短越好”。 | |
| ## 9. Qualitative Cases | |
| 至少选择: | |
| 1. student wrong、teacher 从 gold-hidden evidence 正确恢复、训练后 student 修正; | |
| 2. teacher 也不可恢复、fallback 避免错误蒸馏; | |
| 3. oracle hint 能恢复但 gold-hidden evidence 不足的上界差异; | |
| 4. full reasoning 正确案例; | |
| 5. 模板化 reasoning + malformed answer 的失败案例。 | |
| 每例披露 teacher 可见信息,不展示会造成误解的隐藏 gold prompt。 | |
| ## 10. Artifact 路由 | |
| - Run truth: `docs/paper_reconstruction/experiment_ledger.md` | |
| - Claim status: `docs/paper_reconstruction/claim_evidence_matrix.md` | |
| - Training logs: `outputs/test-fast/logs/pcd_no_visual_<run>/.../train_*.log` | |
| - Eval truth: `<run>/<variant>/eval_chartqa/summary.csv` | |
| - Candidate funnel: `<run>/<variant>/training_health/teacher_candidate_funnel.csv` | |
| - Health windows: `<run>/<variant>/training_health/training_health_summary.csv` | |
| 图表脚本必须从这些 artifact 读取,不从论文 Markdown 反向解析数字。 | |