diff --git a/examples/on_policy_distillation/qwen3_5_35b_selfdistill/README.md b/examples/on_policy_distillation/qwen3_5_35b_selfdistill/README.md
new file mode 100644
index 00000000000..569ea741906
--- /dev/null
+++ b/examples/on_policy_distillation/qwen3_5_35b_selfdistill/README.md
@@ -0,0 +1,210 @@
+# Qwen3.5-35B-A3B Self-Distillation on a Single Node (RLVR teacher → OPD)
+
+A reproducible two-phase on-policy-distillation (OPD) example for the
+**Qwen3.5-35B-A3B** MoE on a **single 8×H200 node**, using the **in-process
+Megatron teacher** (`--opd-type megatron`, no separate teacher server).
+
+It differs from the sibling examples in three ways:
+
+1. **Real MoE at scale on one node.** The 2-node/16-GPU recipe is re-tiled to 8 GPUs.
+2. **A genuinely diverged teacher.** `run-qwen3-8B-opd-megatron.sh` uses `teacher == base`
+   (a mechanism demo where the reverse-KL is ~0). Here Phase 1 *trains* the teacher
+   with RLVR so it is measurably better and more concise than the base — the
+   prerequisite for OPD to actually move the student.
+3. **Self-distillation is the only valid option here.** Qwen3.5 has its own tokenizer
+   (vocab 248320); the smaller Qwen3 models (vocab 151936) are not token-compatible,
+   so a cross-model teacher would be invalid. Teacher and student are the same family.
+
+## Pipeline
+
+```
+Phase 1  (phase1_rlvr_teacher.sh)         Phase 2  (phase2_opd_selfdistill.sh)
+base 35B  --RLVR (GRPO, lr 1e-5)-->  teacher        base 35B (student)
+          better + concise                     |  <-- reverse-KL (--opd-type megatron)
+          (eval 0.83 -> 0.89)                  teacher (Phase-1 ckpt, in-process)
+```
+
+## Single-node parallelism (world = 8)
+
+The original recipe was 2 nodes × 8 GPUs (`TP2 PP1 CP2 EP8 ETP1`, DP4). On one node
+we keep the same dims and only halve DP:
+
+| dim | value | check |
+|-----|-------|-------|
+| TP  | 2 | decoder `TP*PP*CP = 2` ; `8 % 2 = 0` → DP = 4 |
+| PP  | 1 | |
+| CP  | 2 | shards the long (~17k) sequence so 24k context fits |
+| EP  | 8 | `num_experts 256 % 8 = 0` ; expert `ETP*EP*PP = 8` → expert_dp = 1 |
+| ETP | 1 | expert_dp(1) ≠ dp(4) is allowed (miles rank order ends in `pp`) |
+
+`--colocate` time-shares the train and rollout phases (each fits 143 GB separately,
+not summed); `--optimizer-cpu-offload` puts Adam state on host RAM; the model is a
+hybrid linear-attention MoE so the KV cache is small. Peak ≈ 124 GB / 143 GB per GPU.
+
+## Reproduce
+
+**0. Prereqs** — model + torch_dist checkpoint, and the train/eval split:
+
+```bash
+# model (and mcore conversion, see ../README.md for convert_hf_to_torch_dist usage)
+#   ${MODEL_DIR}/Qwen3.5-35B-A3B  and  ${MODEL_DIR}/Qwen3.5-35B-A3B_torch_dist
+# disjoint, seeded train/eval split (eval is held out from BOTH phases):
+python make_split.py --src /path/to/dapo-math-17k.jsonl --out-dir ${DATA_DIR}
+#   -> ${DATA_DIR}/dapo_train.jsonl (16886)  ${DATA_DIR}/dapo_eval.jsonl (512)
+```
+
+**1. Phase 1 — train the teacher** (watch `rollout/raw_reward` climb and
+`eval/dapo_heldout` rise above the base ~0.83):
+
+```bash
+MODEL_DIR=... DATA_DIR=... OUTPUT_DIR=/persistent/ckpt-teacher \
+  bash phase1_rlvr_teacher.sh
+```
+
+**2. Phase 2 — distill the teacher into the base student**:
+
+```bash
+# pure OPD (default): training reward = 0, only the teacher reverse-KL drives learning
+TEACHER_LOAD=/persistent/ckpt-teacher DATA_DIR=... \
+  bash phase2_opd_selfdistill.sh
+
+# grounded OPD: correctness reward (raw_reward == accuracy, climbs) + teacher reverse-KL
+MODE=grounded TEACHER_LOAD=/persistent/ckpt-teacher DATA_DIR=... \
+  bash phase2_opd_selfdistill.sh
+```
+
+`OUTPUT_DIR` / the teacher checkpoint must live on **persistent** storage. On a
+KubeRay pod the head can be recreated and wipe the container overlay (`/root`); a
+node-local disk (e.g. `/node_public`) survives and makes runs resumable.
+
+## Run on GB200 / GB300 (CUDA 13, Blackwell) — `phase2_gb200.sh`
+
+The recipe above targets a single **8×H200** node. Blackwell nodes (GB200/GB300)
+have **4 GPUs/node**, so `world = 8` becomes **2 nodes × 4 GPUs** — same parallel
+dims (`TP2 PP1 CP2 EP8 ETP1`, DP4), only the node tiling changes. `phase2_gb200.sh`
+is the GB200 variant of `phase2_opd_selfdistill.sh`; the deltas (all validated on
+2× GB200, reproducing the base eval `0.84` / `~14k`) are:
+
+- **Tiling** — `--actor-num-nodes 2 --num-gpus-per-node 4` (override via
+  `ACTOR_NUM_NODES` / `GPUS_PER_NODE`). Pin both nodes to one NVLink (MNNVL) domain
+  so the EP8 all-to-all stays on the NVLink fabric.
+- **sglang backends** (cf. `scripts/run_qwen3_5_35b_a3b_mtp_cp2_ep8.py`) —
+  `--sglang-moe-runner-backend flashinfer_cutlass`, `--sglang-attention-backend
+  trtllm_mha`, and `--moe-token-dispatcher-type flex`. The default triton fused-MoE
+  mis-shards routed experts on the megatron→sglang weight sync
+  (`fused_moe_triton ... _load_w13`: `tensor a (64) vs b (2048)`), and FA3 is SM≤90
+  only (Blackwell is SM 10.x).
+- **NCCL** — `NCCL_NVLS_ENABLE=0` (multi-node Blackwell NVLS bind fails
+  `ncclCommInitRank`); keep `NCCL_MNNVL_ENABLE=1`.
+- **k8s** — if a `prometheus` Service exists in the namespace, set
+  `PROMETHEUS_PORT=9090` (kube injects a `tcp://…:9090` URL that breaks miles'
+  `int(PROMETHEUS_PORT)`).
+
+`phase2_gb200.sh` already sets the sglang/MoE backends and folds
+`NCCL_NVLS_ENABLE=0` + `PROMETHEUS_PORT=9090` into the Ray runtime env. Run it on the
+Ray head in the CUDA-13 ARM64 miles image, with `MILES_DIR` pointing at the repo:
+
+```bash
+ACTOR_NUM_NODES=2 GPUS_PER_NODE=4 MILES_DIR=/workspace/miles \
+MODEL_DIR=... DATA_DIR=... TEACHER_LOAD=/persistent/ckpt-teacher OUTPUT_DIR=/persistent/ckpt-opd-pure \
+  bash phase2_gb200.sh          # MODE=pure (default) | MODE=grounded
+```
+
+## Run Phase 2 only (skip Phase 1)
+
+If you already have a teacher checkpoint, skip Phase 1 and run Phase 2 directly —
+point `--opd-teacher-load` (`TEACHER_LOAD`) at the teacher's **torch_dist parent
+dir** (the one containing `latest_checkpointed_iteration.txt`). You still need the
+base model (`--hf-checkpoint` + the `--ref-load` torch_dist) and the data split, but
+no Phase-1 run.
+
+If your teacher is in **HuggingFace** format, convert it to torch_dist first with
+`convert_gb200.sh` (a thin wrapper over `tools/convert_hf_to_torch_dist.py` carrying
+the Qwen3.5 `MODEL_ARGS`):
+
+```bash
+# teacher: HF safetensors -> Megatron torch_dist parent dir
+bash convert_gb200.sh /path/to/teacher-hf /persistent/ckpt-teacher
+# (and the base, if you don't have Qwen3.5-35B-A3B_torch_dist yet)
+bash convert_gb200.sh ${MODEL_DIR}/Qwen3.5-35B-A3B ${MODEL_DIR}/Qwen3.5-35B-A3B_torch_dist
+
+TEACHER_LOAD=/persistent/ckpt-teacher MODEL_DIR=... DATA_DIR=... \
+  bash phase2_gb200.sh          # or phase2_opd_selfdistill.sh on 8×H200
+```
+
+> **Teacher expert layout.** The public `Qwen/Qwen3.5-35B-A3B` ships *fused* experts
+> (`mlp.experts.gate_up_proj`); a teacher round-tripped through
+> `convert_torch_dist_to_hf` may ship *unfused* per-expert weights
+> (`mlp.experts.{i}.gate_proj.weight`). `miles_plugins/mbridge/qwen3_5.py` now
+> autodetects both for the main layers (mirroring the existing MTP-expert
+> autodetect), so either layout converts without manual re-fusing.
+
+## Results (DAPO-math, held-out 512, eval @ 24k cap, temp 0.6)
+
+**Phase 1 — RLVR teacher** (lr 1e-5):
+
+| step | eval/dapo_heldout | eval response length |
+|------|-------------------|----------------------|
+| 0 (base) | 0.828 | 14,070 |
+| 5        | **0.887** | **6,248** |
+
+The teacher becomes both more accurate **and** ~2× more concise. This Phase-1
+teacher checkpoint is published at
+[**cm00cm/Qwen3.5-35B-A3B-DAPO-RLVR-teacher**](https://huggingface.co/cm00cm/Qwen3.5-35B-A3B-DAPO-RLVR-teacher)
+(weights only) and can be used directly as the Phase-2 teacher via
+`--opd-teacher-load` after `convert_hf_to_torch_dist.py`.
+
+**Phase 2 — pure OPD** (student = base, teacher = Phase-1 step-5 ckpt; reward = 0):
+
+| step | eval/dapo_heldout | eval response length | opd_reverse_kl |
+|------|-------------------|----------------------|----------------|
+| 0 (base) | 0.840 | 14,070 | — |
+| 5        | 0.852 | **6,132** | 0.045 → 0.013 |
+
+With **zero task reward**, pure reverse-KL distillation transfers the teacher's
+concise behavior to the base student — eval length **−57%** with accuracy
+preserved/slightly up (the +1.2 pt is within the ~1.6 pt eval SE; the robust,
+headline effect is the efficiency transfer). A nonzero, shrinking `opd_reverse_kl`
+confirms the teacher genuinely differs from the student and the student is
+converging onto it.
+
+**Phase 2 — grounded OPD** (correctness reward + teacher reverse-KL):
+
+| step | rollout/raw_reward | train length | opd_reverse_kl |
+|------|--------------------|--------------|----------------|
+| 1 | 0.637 | 18,778 | 0.045 |
+| 2 | **0.910** | **7,665** | 0.014 |
+
+With the correctness reward kept, `rollout/raw_reward` (== accuracy) climbs while
+the student simultaneously adopts the teacher's concise responses (18.8k → 7.7k).
+The shrinking `opd_reverse_kl` (0.045 → 0.014) shows the student converging onto
+the teacher. (At lr 1e-5 the RLVR reward alone also drives accuracy up — Phase 1
+is the controlled view of that — so grounded OPD's `raw_reward` climb reflects
+RLVR + the teacher pull combined; the pure-OPD run above isolates OPD's effect.)
+
+## Gotchas (each cost a wasted run to find)
+
+- **Reward grader.** `--rm-type deepscaler` requires a `</think>` tag and returns 0
+  otherwise; Qwen3.5 reasons inline (no tag) → every reward 0. `--rm-type math`
+  only reads `\boxed{}`; `--rm-type dapo` only `Answer:`. Use the format-agnostic
+  `rm.reward_func` (accepts either). Always pass `--label-key label` for the
+  `{prompt, label}` DAPO jsonl, or `Sample.label` is `None` and reward reads 0.
+- **Context length.** The 35B's DAPO chain-of-thought is ~14–17k tokens. An 8k
+  response cap truncates ~95% of rollouts mid-reasoning → reward ~0. Use ≥24k
+  (CP2 makes 24–32k feasible).
+- **`--opd-teacher-load` path.** Point at the checkpoint **parent** dir (contains
+  `latest_checkpointed_iteration.txt`), not an `iter_XXXXXXX` subdir. The subdir
+  has no metadata → silent fallback to base → teacher == student → `opd_reverse_kl ≈ 0`.
+  Sanity check: in the rollout log, `teacher_log_probs` should differ from
+  `rollout/log_probs`.
+- **Teacher must diverge.** A few RLVR steps at lr 1e-6 barely move the weights, so
+  the teacher ≈ base and OPD is inert (`opd_reverse_kl ≈ 5e-4`). lr 1e-5 diverges it
+  fast (`opd_reverse_kl ≈ 5e-2`). `--opd-kl-coef` cannot amplify a ~0 KL.
+- **Memory.** `with_ref = (--use-kl-loss or --kl-coef≠0)`. Dropping `--use-kl-loss`
+  keeps only student + teacher (2×35B) in memory; the teacher reverse-KL is the
+  regularizer. Adding it loads a 3rd model and risks OOM.
+
+## References
+- Phase-1 teacher checkpoint: https://huggingface.co/cm00cm/Qwen3.5-35B-A3B-DAPO-RLVR-teacher
+- ../README.md (served-teacher OPD), ../run-qwen3-8B-opd-megatron.sh (in-process teacher)
+- https://thinkingmachines.ai/blog/on-policy-distillation/
diff --git a/examples/on_policy_distillation/qwen3_5_35b_selfdistill/convert_gb200.sh b/examples/on_policy_distillation/qwen3_5_35b_selfdistill/convert_gb200.sh
new file mode 100644
index 00000000000..8cd85d2e047
--- /dev/null
+++ b/examples/on_policy_distillation/qwen3_5_35b_selfdistill/convert_gb200.sh
@@ -0,0 +1,39 @@
+#!/bin/bash
+# =============================================================================
+# Convert a HF Qwen3.5-35B-A3B checkpoint -> Megatron torch_dist.
+# Used to stage both the base (--ref-load) and the teacher (--opd-teacher-load)
+# for phase2_gb200.sh, since neither is pre-staged on /cluster_public.
+#
+# Usage: convert_gb200.sh <hf_checkpoint_dir> <torch_dist_save_dir>
+# =============================================================================
+set -ex
+HF_IN=${1:?hf checkpoint dir}
+SAVE_OUT=${2:?torch_dist save dir}
+MILES_DIR=${MILES_DIR:-/workspace/miles}
+MEGATRON_PATH=${MEGATRON_PATH:-/root/Megatron-LM}
+
+# Identical architecture spec to phase2_gb200.sh's MODEL_ARGS.
+MODEL_ARGS=(
+   --spec miles_plugins.models.qwen3_5 get_qwen3_5_spec
+   --disable-bias-linear --qk-layernorm --group-query-attention
+   --num-attention-heads 16 --num-query-groups 2 --kv-channels 256
+   --num-layers 40 --hidden-size 2048 --ffn-hidden-size 512
+   --normalization RMSNorm --apply-layernorm-1p --position-embedding-type rope
+   --norm-epsilon 1e-6 --rotary-percent 0.25 --swiglu
+   --untie-embeddings-and-output-weights --vocab-size 248320 --rotary-base 10000000
+   --moe-ffn-hidden-size 512 --moe-shared-expert-intermediate-size 512
+   --moe-router-score-function softmax --moe-token-dispatcher-type alltoall
+   --moe-router-topk 8
+   --moe-layer-freq "[1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1]"
+   --num-experts 256 --moe-grouped-gemm --moe-token-drop-policy probs --moe-router-dtype fp32
+   --moe-permute-fusion --moe-aux-loss-coeff 0 --attention-output-gate --moe-shared-expert-gate
+   --mtp-num-layers 1
+)
+
+cd "${MILES_DIR}"
+PYTHONPATH="${MILES_DIR}:${MEGATRON_PATH}" python3 "${MILES_DIR}/tools/convert_hf_to_torch_dist.py" \
+   "${MODEL_ARGS[@]}" \
+   --hf-checkpoint "${HF_IN}" \
+   --save "${SAVE_OUT}"
+echo "CONVERTED ${HF_IN} -> ${SAVE_OUT}"
+ls -la "${SAVE_OUT}"; cat "${SAVE_OUT}/latest_checkpointed_iteration.txt" 2>/dev/null || echo "(no latest_checkpointed_iteration.txt yet)"
diff --git a/examples/on_policy_distillation/qwen3_5_35b_selfdistill/eval_dapo_heldout.yaml b/examples/on_policy_distillation/qwen3_5_35b_selfdistill/eval_dapo_heldout.yaml
new file mode 100644
index 00000000000..217b9aa4256
--- /dev/null
+++ b/examples/on_policy_distillation/qwen3_5_35b_selfdistill/eval_dapo_heldout.yaml
@@ -0,0 +1,17 @@
+# Held-out DAPO eval (disjoint from the training split, see README "Data split").
+# Scored by the example's format-agnostic reward (--custom-rm-path ...rm.reward_func),
+# which reports accuracy. max_response_len 24576 must exceed the model's reasoning
+# length (~14-17k for the base 35B) or accuracy is suppressed by truncation.
+eval:
+  defaults:
+    temperature: 0.6
+    top_p: 0.95
+  datasets:
+    - name: dapo_heldout
+      path: ${DATA_DIR}/dapo_eval.jsonl     # rendered by the launch scripts via envsubst
+      input_key: prompt
+      label_key: label                       # REQUIRED: without it Sample.label=None -> eval reads 0
+      n_samples_per_eval_prompt: 1
+      max_response_len: 24576
+      metadata_overrides:
+        opd_reward_mode: eval_math            # tags eval samples for reward_func_pure_opd
diff --git a/examples/on_policy_distillation/qwen3_5_35b_selfdistill/make_split.py b/examples/on_policy_distillation/qwen3_5_35b_selfdistill/make_split.py
new file mode 100644
index 00000000000..15f9264ff38
--- /dev/null
+++ b/examples/on_policy_distillation/qwen3_5_35b_selfdistill/make_split.py
@@ -0,0 +1,61 @@
+"""Carve a disjoint train/eval split from dapo-math-17k for the self-distillation example.
+
+The file is ordered by difficulty, so we shuffle with a FIXED SEED before splitting
+(a contiguous tail-N split would be systematically easier and bias the eval). The
+512-problem eval split is held out from BOTH phases. Dedup is on prompt text (labels
+are not unique). Usage:
+
+    python make_split.py --src /path/dapo-math-17k.jsonl --out-dir /path/split
+"""
+
+import argparse
+import hashlib
+import json
+import os
+import random
+
+
+def prompt_text(d):
+    p = d["prompt"]
+    return "\n".join(m.get("content", "") for m in p) if isinstance(p, list) else str(p)
+
+
+def main():
+    ap = argparse.ArgumentParser()
+    ap.add_argument("--src", required=True, help="dapo-math-17k.jsonl")
+    ap.add_argument("--out-dir", required=True)
+    ap.add_argument("--eval-n", type=int, default=512)
+    ap.add_argument("--seed", type=int, default=42)
+    args = ap.parse_args()
+    os.makedirs(args.out_dir, exist_ok=True)
+
+    rows, seen = [], set()
+    with open(args.src) as f:
+        for line in f:
+            line = line.strip()
+            if not line:
+                continue
+            key = prompt_text(json.loads(line))
+            if key in seen:
+                continue
+            seen.add(key)
+            rows.append(line)
+
+    random.Random(args.seed).shuffle(rows)  # fixed seed: reproducible, unbiased split
+    eval_rows, train_rows = rows[-args.eval_n :], rows[: -args.eval_n]
+
+    ek = {prompt_text(json.loads(r)) for r in eval_rows}
+    tk = {prompt_text(json.loads(r)) for r in train_rows}
+    assert ek.isdisjoint(tk), "LEAK: eval prompt found in train split"
+
+    with open(os.path.join(args.out_dir, "dapo_train.jsonl"), "w") as f:
+        f.write("\n".join(train_rows) + "\n")
+    with open(os.path.join(args.out_dir, "dapo_eval.jsonl"), "w") as f:
+        f.write("\n".join(eval_rows) + "\n")
+
+    md5 = hashlib.md5("\n".join(eval_rows).encode()).hexdigest()
+    print(f"train={len(train_rows)} eval={len(eval_rows)} seed={args.seed} eval_md5={md5}")
+
+
+if __name__ == "__main__":
+    main()
diff --git a/examples/on_policy_distillation/qwen3_5_35b_selfdistill/phase1_rlvr_teacher.sh b/examples/on_policy_distillation/qwen3_5_35b_selfdistill/phase1_rlvr_teacher.sh
new file mode 100755
index 00000000000..b04c3d5a5af
--- /dev/null
+++ b/examples/on_policy_distillation/qwen3_5_35b_selfdistill/phase1_rlvr_teacher.sh
@@ -0,0 +1,105 @@
+#!/bin/bash
+# =============================================================================
+# Phase 1: RLVR-train a Qwen3.5-35B-A3B *teacher* on DAPO-math (single 8xH200 node)
+# =============================================================================
+# Trains the base model with GRPO + format-agnostic correctness reward so it becomes
+# measurably BETTER and MORE CONCISE than the base. This trained checkpoint is the
+# teacher distilled into the base student in Phase 2.
+#
+#   Single node, 8 GPUs:  TP2 / PP1 / CP2 / EP8 / ETP1  (world=8, DP2)
+#   lr 1e-5 is intentional: it makes the teacher DIVERGE from base fast enough to
+#   matter within ~5-10 steps. lr 1e-6 (few steps) barely moves the weights, leaving
+#   the teacher ~= base, which makes the Phase-2 OPD reverse-KL ~0 (inert).
+#
+# Usage:  bash phase1_rlvr_teacher.sh
+# Env (override as needed):
+#   MODEL_DIR   dir holding Qwen3.5-35B-A3B and Qwen3.5-35B-A3B_torch_dist
+#   DATA_DIR    dir holding dapo_train.jsonl / dapo_eval.jsonl (see make_split.py)
+#   OUTPUT_DIR  writable, *node-local-persistent* dir for checkpoints (see README)
+#   EXAMPLE_DIR this directory (for rm.py + eval config on PYTHONPATH)
+# =============================================================================
+set -ex
+export PYTHONUNBUFFERED=16
+
+MODEL_DIR=${MODEL_DIR:-/cluster_public/miles_data/models}
+DATA_DIR=${DATA_DIR:-/node_public/maocheng-qwen35/data}
+OUTPUT_DIR=${OUTPUT_DIR:-/node_public/maocheng-qwen35/ckpt-teacher}
+EXAMPLE_DIR=${EXAMPLE_DIR:-$(cd "$(dirname "$0")" && pwd)}
+MILES_DIR=${MILES_DIR:-/root/miles}
+RAY_ADDRESS=${RAY_ADDRESS:-http://127.0.0.1:8265}
+mkdir -p "${OUTPUT_DIR}"
+
+# Render the eval config (substitutes ${DATA_DIR}).
+EVAL_CONFIG="${OUTPUT_DIR}/eval_dapo_heldout.yaml"
+DATA_DIR="${DATA_DIR}" envsubst < "${EXAMPLE_DIR}/eval_dapo_heldout.yaml" > "${EVAL_CONFIG}"
+
+# Qwen3.5-35B-A3B architecture (no scripts/models/*.sh ships for it).
+MODEL_ARGS=(
+   --spec miles_plugins.models.qwen3_5 get_qwen3_5_spec
+   --disable-bias-linear --qk-layernorm --group-query-attention
+   --num-attention-heads 16 --num-query-groups 2 --kv-channels 256
+   --num-layers 40 --hidden-size 2048 --ffn-hidden-size 512
+   --normalization RMSNorm --apply-layernorm-1p --position-embedding-type rope
+   --norm-epsilon 1e-6 --rotary-percent 0.25 --swiglu
+   --untie-embeddings-and-output-weights --vocab-size 248320 --rotary-base 10000000
+   --moe-ffn-hidden-size 512 --moe-shared-expert-intermediate-size 512
+   --moe-router-score-function softmax --moe-token-dispatcher-type alltoall
+   --moe-router-topk 8
+   --moe-layer-freq "[1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1]"
+   --num-experts 256 --moe-grouped-gemm --moe-token-drop-policy probs --moe-router-dtype fp32
+   --moe-permute-fusion --moe-aux-loss-coeff 0 --attention-output-gate --moe-shared-expert-gate
+   --mtp-num-layers 1
+)
+CKPT_ARGS=(
+   --hf-checkpoint ${MODEL_DIR}/Qwen3.5-35B-A3B
+   --ref-load ${MODEL_DIR}/Qwen3.5-35B-A3B_torch_dist
+   --load ${OUTPUT_DIR} --save ${OUTPUT_DIR} --save-interval 5
+)
+ROLLOUT_ARGS=(
+   --prompt-data ${DATA_DIR}/dapo_train.jsonl --input-key prompt --label-key label
+   --apply-chat-template --rollout-shuffle
+   --num-rollout 20 --rollout-batch-size 32 --n-samples-per-prompt 8
+   --rollout-max-response-len 24576 --rollout-temperature 1 --num-steps-per-rollout 1
+   --over-sampling-batch-size 32 --global-batch-size 256 --balance-data
+)
+# Format-agnostic correctness reward -> rollout/raw_reward == accuracy.
+RM_ARGS=( --custom-rm-path examples.on_policy_distillation.qwen3_5_35b_selfdistill.rm.reward_func )
+EVAL_ARGS=( --eval-interval 5 --eval-config ${EVAL_CONFIG} )
+GRPO_ARGS=(
+   --advantage-estimator grpo --kl-loss-type low_var_kl --entropy-coef 0.00
+   --eps-clip 0.2 --eps-clip-high 0.28 --use-tis
+)
+OPTIMIZER_ARGS=(
+   --optimizer adam --lr 1e-5 --lr-decay-style constant --weight-decay 0.1
+   --adam-beta1 0.9 --adam-beta2 0.98
+   --optimizer-cpu-offload --overlap-cpu-optimizer-d2h-h2d --use-precision-aware-optimizer
+)
+PERF_ARGS=(
+   --tensor-model-parallel-size 2 --sequence-parallel --pipeline-model-parallel-size 1
+   --context-parallel-size 2 --expert-model-parallel-size 8 --expert-tensor-parallel-size 1
+   --recompute-granularity full --recompute-method uniform --recompute-num-layers 1
+   --use-dynamic-batch-size --max-tokens-per-gpu 16384 --log-probs-chunk-size 4096
+)
+SGLANG_ARGS=(
+   --rollout-num-gpus-per-engine 8 --sglang-mem-fraction-static 0.8 --sglang-ep-size 8
+   --sglang-watchdog-timeout 1800 --sglang-enable-metrics --sglang-attention-backend fa3
+   --sglang-cuda-graph-bs 1 2 4 8 16 32 --use-rollout-routing-replay
+   --sglang-mamba-scheduler-strategy extra_buffer
+)
+MISC_ARGS=(
+   --attention-dropout 0.0 --hidden-dropout 0.0 --accumulate-allreduce-grads-in-fp32
+   --attention-softmax-in-fp32 --attention-backend flash
+)
+WANDB_ARGS=( --use-wandb --wandb-project miles-opd --wandb-group qwen3.5-35b-rlvr-teacher )
+
+RUNTIME_ENV_JSON="{\"env_vars\": {\"PYTHONPATH\": \"${MILES_DIR}:/root/Megatron-LM/\", \"CUDA_DEVICE_MAX_CONNECTIONS\": \"1\", \"WANDB_API_KEY\": \"${WANDB_API_KEY}\"}}"
+
+cd "${MILES_DIR}"
+ray job submit --address="${RAY_ADDRESS}" --submission-id qwen3.5-rlvr-teacher --no-wait \
+   --runtime-env-json="${RUNTIME_ENV_JSON}" \
+   -- python3 ${MILES_DIR}/train.py \
+   --actor-num-nodes 1 --actor-num-gpus-per-node 8 --num-gpus-per-node 8 --colocate \
+   ${MODEL_ARGS[@]} ${CKPT_ARGS[@]} ${ROLLOUT_ARGS[@]} ${OPTIMIZER_ARGS[@]} ${GRPO_ARGS[@]} \
+   ${WANDB_ARGS[@]} ${PERF_ARGS[@]} ${EVAL_ARGS[@]} ${SGLANG_ARGS[@]} ${MISC_ARGS[@]} ${RM_ARGS[@]}
+echo "Submitted Phase-1 RLVR teacher training (submission-id qwen3.5-rlvr-teacher)."
+echo "Watch rollout/raw_reward climb and eval/dapo_heldout rise above the base ~0.83."
diff --git a/examples/on_policy_distillation/qwen3_5_35b_selfdistill/phase2_gb200.sh b/examples/on_policy_distillation/qwen3_5_35b_selfdistill/phase2_gb200.sh
new file mode 100755
index 00000000000..512abe575c3
--- /dev/null
+++ b/examples/on_policy_distillation/qwen3_5_35b_selfdistill/phase2_gb200.sh
@@ -0,0 +1,140 @@
+#!/bin/bash
+# =============================================================================
+# Phase 2: On-policy distillation of the Phase-1 teacher into the BASE student
+# =============================================================================
+# Student = base Qwen3.5-35B-A3B.  Teacher = the Phase-1 RLVR checkpoint (in-process,
+# --opd-type megatron). The student trains only on its own rollouts; the teacher's
+# token-level reverse-KL is folded into the GRPO advantages.
+#
+# Two modes (MODE env var):
+#   pure      (default) — training task reward = 0; ONLY the teacher reverse-KL drives
+#                         learning. Cleanest attribution: any change is distillation.
+#                         eval/dapo_heldout still measures real accuracy.
+#   grounded  — keep the correctness reward (rollout/raw_reward == accuracy, climbs)
+#                         AND add the teacher reverse-KL on top.
+#
+# Memory note: we do NOT pass --use-kl-loss, so the reference model is NOT loaded
+# (with_ref = use_kl_loss or kl_coef!=0). That keeps only student + teacher in memory
+# (2 x 35B, ~124/143 GB per GPU). Adding --use-kl-loss would load a 3rd model and risk OOM.
+# The teacher's reverse-KL is the regularizer.
+#
+# Usage:  TEACHER_LOAD=/path/to/teacher_ckpt_parent bash phase2_opd_selfdistill.sh
+# IMPORTANT: --opd-teacher-load must point at the checkpoint PARENT directory (the one
+# containing latest_checkpointed_iteration.txt), NOT an iter_XXXXXXX subdir. Pointing at
+# the subdir prints "could not find metadata file" and silently falls back to base ->
+# teacher == student -> opd_reverse_kl ~= 0 (inert).
+# =============================================================================
+set -ex
+export PYTHONUNBUFFERED=16
+
+MODE=${MODE:-pure}
+# GB200 tiling: this cluster is 4 GPUs/node, so world=8 = 2 nodes x 4 GPUs.
+# (Upstream phase2_opd_selfdistill.sh assumes a single 8xH200 node.) The parallel
+# dims (TP2/PP1/CP2/EP8/ETP1, DP4) are unchanged; only the node tiling differs.
+# Both nodes must share one NVLink (MNNVL) domain -- launch with MSC_RACK pinned.
+ACTOR_NUM_NODES=${ACTOR_NUM_NODES:-2}
+GPUS_PER_NODE=${GPUS_PER_NODE:-4}
+MODEL_DIR=${MODEL_DIR:-/cluster_public/miles_data/models}
+DATA_DIR=${DATA_DIR:-/node_public/maocheng-qwen35/data}
+OUTPUT_DIR=${OUTPUT_DIR:-/node_public/maocheng-qwen35/ckpt-opd-${MODE}}
+TEACHER_LOAD=${TEACHER_LOAD:-/node_public/maocheng-qwen35/ckpt-teacher}   # parent dir!
+EXAMPLE_DIR=${EXAMPLE_DIR:-$(cd "$(dirname "$0")" && pwd)}
+MILES_DIR=${MILES_DIR:-/workspace/miles}
+RAY_ADDRESS=${RAY_ADDRESS:-http://127.0.0.1:8265}
+OPD_KL_COEF=${OPD_KL_COEF:-0.2}
+mkdir -p "${OUTPUT_DIR}"
+
+EVAL_CONFIG="${OUTPUT_DIR}/eval_dapo_heldout.yaml"
+DATA_DIR="${DATA_DIR}" envsubst < "${EXAMPLE_DIR}/eval_dapo_heldout.yaml" > "${EVAL_CONFIG}"
+
+if [ "${MODE}" = "pure" ]; then
+   RM_FUNC=examples.on_policy_distillation.qwen3_5_35b_selfdistill.rm.reward_func_pure_opd
+else
+   RM_FUNC=examples.on_policy_distillation.qwen3_5_35b_selfdistill.rm.reward_func
+fi
+
+MODEL_ARGS=(
+   --spec miles_plugins.models.qwen3_5 get_qwen3_5_spec
+   --disable-bias-linear --qk-layernorm --group-query-attention
+   --num-attention-heads 16 --num-query-groups 2 --kv-channels 256
+   --num-layers 40 --hidden-size 2048 --ffn-hidden-size 512
+   --normalization RMSNorm --apply-layernorm-1p --position-embedding-type rope
+   --norm-epsilon 1e-6 --rotary-percent 0.25 --swiglu
+   --untie-embeddings-and-output-weights --vocab-size 248320 --rotary-base 10000000
+   --moe-ffn-hidden-size 512 --moe-shared-expert-intermediate-size 512
+   --moe-router-score-function softmax --moe-token-dispatcher-type flex
+   --moe-router-topk 8
+   --moe-layer-freq "[1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1]"
+   --num-experts 256 --moe-grouped-gemm --moe-token-drop-policy probs --moe-router-dtype fp32
+   --moe-permute-fusion --moe-aux-loss-coeff 0 --attention-output-gate --moe-shared-expert-gate
+   --mtp-num-layers 1
+)
+CKPT_ARGS=(
+   --hf-checkpoint ${MODEL_DIR}/Qwen3.5-35B-A3B
+   --ref-load ${MODEL_DIR}/Qwen3.5-35B-A3B_torch_dist
+   --load ${OUTPUT_DIR} --save ${OUTPUT_DIR} --save-interval 5
+)
+OPD_ARGS=(
+   --use-opd --opd-type megatron --opd-teacher-load ${TEACHER_LOAD} --opd-kl-coef ${OPD_KL_COEF}
+)
+ROLLOUT_ARGS=(
+   --prompt-data ${DATA_DIR}/dapo_train.jsonl --input-key prompt --label-key label
+   --apply-chat-template --rollout-shuffle
+   --num-rollout 12 --rollout-batch-size 32 --n-samples-per-prompt 8
+   --rollout-max-response-len 24576 --rollout-temperature 1 --num-steps-per-rollout 1
+   --over-sampling-batch-size 32 --global-batch-size 256 --balance-data
+)
+RM_ARGS=( --custom-rm-path ${RM_FUNC} )
+EVAL_ARGS=( --eval-interval 5 --eval-config ${EVAL_CONFIG} )
+GRPO_ARGS=(
+   --advantage-estimator grpo --kl-loss-type low_var_kl --entropy-coef 0.00
+   --eps-clip 0.2 --eps-clip-high 0.28 --use-tis
+)
+OPTIMIZER_ARGS=(
+   --optimizer adam --lr 1e-5 --lr-decay-style constant --weight-decay 0.1
+   --adam-beta1 0.9 --adam-beta2 0.98
+   --optimizer-cpu-offload --overlap-cpu-optimizer-d2h-h2d --use-precision-aware-optimizer
+)
+PERF_ARGS=(
+   --tensor-model-parallel-size 2 --sequence-parallel --pipeline-model-parallel-size 1
+   --context-parallel-size 2 --expert-model-parallel-size 8 --expert-tensor-parallel-size 1
+   --recompute-granularity full --recompute-method uniform --recompute-num-layers 1
+   --use-dynamic-batch-size --max-tokens-per-gpu 16384 --log-probs-chunk-size 4096
+)
+# Blackwell (GB200 sm100 / B300 sm103) backends, per scripts/run_qwen3_5_35b_a3b_mtp_cp2_ep8.py:
+# - moe-runner-backend flashinfer_cutlass: the default triton fused-MoE mis-shards
+#   routed experts on the megatron->sglang weight sync (w13 reshape 64 vs 2048).
+# - attention-backend trtllm_mha: FA3 is SM<=90 only; flashinfer/fa3 don't fit here.
+SGLANG_ARGS=(
+   --rollout-num-gpus-per-engine 8 --sglang-mem-fraction-static 0.7 --sglang-ep-size 8
+   --sglang-watchdog-timeout 1800 --sglang-enable-metrics
+   --sglang-moe-runner-backend flashinfer_cutlass --sglang-attention-backend trtllm_mha
+   --sglang-cuda-graph-bs 1 2 4 8 16 32 --use-rollout-routing-replay
+   --sglang-mamba-scheduler-strategy extra_buffer
+)
+MISC_ARGS=(
+   --attention-dropout 0.0 --hidden-dropout 0.0 --accumulate-allreduce-grads-in-fp32
+   --attention-softmax-in-fp32 --attention-backend flash
+)
+WANDB_ARGS=( --use-wandb --wandb-project miles-opd --wandb-group qwen3.5-35b-opd-${MODE} )
+
+# PROMETHEUS_PORT override: the contractor k8s namespace has a `prometheus`
+# Service, so kube auto-injects PROMETHEUS_PORT=tcp://<ip>:9090 into every pod.
+# miles' add_prometheus_arguments does int(os.environ["PROMETHEUS_PORT"]) and
+# crashes on the URL; force the plain port back via the ray runtime env.
+# NCCL_NVLS_ENABLE=0: on multi-node GB200 the NVLS (NVLink SHARP) bind fails
+# during cross-node ncclCommInitRank ("unhandled cuda error" / CUDA 999), which
+# kills the EP8 sglang engine spanning both nodes. MNNVL stays enabled (pod env)
+# for the NVLink fabric itself; only NVLS multicast is disabled.
+RUNTIME_ENV_JSON="{\"env_vars\": {\"PYTHONPATH\": \"${MILES_DIR}:/root/Megatron-LM/\", \"CUDA_DEVICE_MAX_CONNECTIONS\": \"1\", \"WANDB_API_KEY\": \"${WANDB_API_KEY}\", \"PROMETHEUS_PORT\": \"9090\", \"NCCL_NVLS_ENABLE\": \"0\"}}"
+
+cd "${MILES_DIR}"
+ray job submit --address="${RAY_ADDRESS}" --submission-id qwen3.5-opd-${MODE} --no-wait \
+   --runtime-env-json="${RUNTIME_ENV_JSON}" \
+   -- python3 ${MILES_DIR}/train.py \
+   --actor-num-nodes ${ACTOR_NUM_NODES} --actor-num-gpus-per-node ${GPUS_PER_NODE} --num-gpus-per-node ${GPUS_PER_NODE} --colocate \
+   ${MODEL_ARGS[@]} ${CKPT_ARGS[@]} ${OPD_ARGS[@]} ${ROLLOUT_ARGS[@]} ${OPTIMIZER_ARGS[@]} ${GRPO_ARGS[@]} \
+   ${WANDB_ARGS[@]} ${PERF_ARGS[@]} ${EVAL_ARGS[@]} ${SGLANG_ARGS[@]} ${MISC_ARGS[@]} ${RM_ARGS[@]}
+echo "Submitted Phase-2 OPD (${MODE}) self-distillation (submission-id qwen3.5-opd-${MODE})."
+echo "Watch opd_reverse_kl (>>0 means the teacher differs from the student) and"
+echo "eval/dapo_heldout (student moving toward the teacher's accuracy, with shorter responses)."
diff --git a/examples/on_policy_distillation/qwen3_5_35b_selfdistill/phase2_opd_selfdistill.sh b/examples/on_policy_distillation/qwen3_5_35b_selfdistill/phase2_opd_selfdistill.sh
new file mode 100755
index 00000000000..0c3109a1744
--- /dev/null
+++ b/examples/on_policy_distillation/qwen3_5_35b_selfdistill/phase2_opd_selfdistill.sh
@@ -0,0 +1,121 @@
+#!/bin/bash
+# =============================================================================
+# Phase 2: On-policy distillation of the Phase-1 teacher into the BASE student
+# =============================================================================
+# Student = base Qwen3.5-35B-A3B.  Teacher = the Phase-1 RLVR checkpoint (in-process,
+# --opd-type megatron). The student trains only on its own rollouts; the teacher's
+# token-level reverse-KL is folded into the GRPO advantages.
+#
+# Two modes (MODE env var):
+#   pure      (default) — training task reward = 0; ONLY the teacher reverse-KL drives
+#                         learning. Cleanest attribution: any change is distillation.
+#                         eval/dapo_heldout still measures real accuracy.
+#   grounded  — keep the correctness reward (rollout/raw_reward == accuracy, climbs)
+#                         AND add the teacher reverse-KL on top.
+#
+# Memory note: we do NOT pass --use-kl-loss, so the reference model is NOT loaded
+# (with_ref = use_kl_loss or kl_coef!=0). That keeps only student + teacher in memory
+# (2 x 35B, ~124/143 GB per GPU). Adding --use-kl-loss would load a 3rd model and risk OOM.
+# The teacher's reverse-KL is the regularizer.
+#
+# Usage:  TEACHER_LOAD=/path/to/teacher_ckpt_parent bash phase2_opd_selfdistill.sh
+# IMPORTANT: --opd-teacher-load must point at the checkpoint PARENT directory (the one
+# containing latest_checkpointed_iteration.txt), NOT an iter_XXXXXXX subdir. Pointing at
+# the subdir prints "could not find metadata file" and silently falls back to base ->
+# teacher == student -> opd_reverse_kl ~= 0 (inert).
+# =============================================================================
+set -ex
+export PYTHONUNBUFFERED=16
+
+MODE=${MODE:-pure}
+MODEL_DIR=${MODEL_DIR:-/cluster_public/miles_data/models}
+DATA_DIR=${DATA_DIR:-/node_public/maocheng-qwen35/data}
+OUTPUT_DIR=${OUTPUT_DIR:-/node_public/maocheng-qwen35/ckpt-opd-${MODE}}
+TEACHER_LOAD=${TEACHER_LOAD:-/node_public/maocheng-qwen35/ckpt-teacher}   # parent dir!
+EXAMPLE_DIR=${EXAMPLE_DIR:-$(cd "$(dirname "$0")" && pwd)}
+MILES_DIR=${MILES_DIR:-/root/miles}
+RAY_ADDRESS=${RAY_ADDRESS:-http://127.0.0.1:8265}
+OPD_KL_COEF=${OPD_KL_COEF:-0.2}
+mkdir -p "${OUTPUT_DIR}"
+
+EVAL_CONFIG="${OUTPUT_DIR}/eval_dapo_heldout.yaml"
+DATA_DIR="${DATA_DIR}" envsubst < "${EXAMPLE_DIR}/eval_dapo_heldout.yaml" > "${EVAL_CONFIG}"
+
+if [ "${MODE}" = "pure" ]; then
+   RM_FUNC=examples.on_policy_distillation.qwen3_5_35b_selfdistill.rm.reward_func_pure_opd
+else
+   RM_FUNC=examples.on_policy_distillation.qwen3_5_35b_selfdistill.rm.reward_func
+fi
+
+MODEL_ARGS=(
+   --spec miles_plugins.models.qwen3_5 get_qwen3_5_spec
+   --disable-bias-linear --qk-layernorm --group-query-attention
+   --num-attention-heads 16 --num-query-groups 2 --kv-channels 256
+   --num-layers 40 --hidden-size 2048 --ffn-hidden-size 512
+   --normalization RMSNorm --apply-layernorm-1p --position-embedding-type rope
+   --norm-epsilon 1e-6 --rotary-percent 0.25 --swiglu
+   --untie-embeddings-and-output-weights --vocab-size 248320 --rotary-base 10000000
+   --moe-ffn-hidden-size 512 --moe-shared-expert-intermediate-size 512
+   --moe-router-score-function softmax --moe-token-dispatcher-type alltoall
+   --moe-router-topk 8
+   --moe-layer-freq "[1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1]"
+   --num-experts 256 --moe-grouped-gemm --moe-token-drop-policy probs --moe-router-dtype fp32
+   --moe-permute-fusion --moe-aux-loss-coeff 0 --attention-output-gate --moe-shared-expert-gate
+   --mtp-num-layers 1
+)
+CKPT_ARGS=(
+   --hf-checkpoint ${MODEL_DIR}/Qwen3.5-35B-A3B
+   --ref-load ${MODEL_DIR}/Qwen3.5-35B-A3B_torch_dist
+   --load ${OUTPUT_DIR} --save ${OUTPUT_DIR} --save-interval 5
+)
+OPD_ARGS=(
+   --use-opd --opd-type megatron --opd-teacher-load ${TEACHER_LOAD} --opd-kl-coef ${OPD_KL_COEF}
+)
+ROLLOUT_ARGS=(
+   --prompt-data ${DATA_DIR}/dapo_train.jsonl --input-key prompt --label-key label
+   --apply-chat-template --rollout-shuffle
+   --num-rollout 12 --rollout-batch-size 32 --n-samples-per-prompt 8
+   --rollout-max-response-len 24576 --rollout-temperature 1 --num-steps-per-rollout 1
+   --over-sampling-batch-size 32 --global-batch-size 256 --balance-data
+)
+RM_ARGS=( --custom-rm-path ${RM_FUNC} )
+EVAL_ARGS=( --eval-interval 5 --eval-config ${EVAL_CONFIG} )
+GRPO_ARGS=(
+   --advantage-estimator grpo --kl-loss-type low_var_kl --entropy-coef 0.00
+   --eps-clip 0.2 --eps-clip-high 0.28 --use-tis
+)
+OPTIMIZER_ARGS=(
+   --optimizer adam --lr 1e-5 --lr-decay-style constant --weight-decay 0.1
+   --adam-beta1 0.9 --adam-beta2 0.98
+   --optimizer-cpu-offload --overlap-cpu-optimizer-d2h-h2d --use-precision-aware-optimizer
+)
+PERF_ARGS=(
+   --tensor-model-parallel-size 2 --sequence-parallel --pipeline-model-parallel-size 1
+   --context-parallel-size 2 --expert-model-parallel-size 8 --expert-tensor-parallel-size 1
+   --recompute-granularity full --recompute-method uniform --recompute-num-layers 1
+   --use-dynamic-batch-size --max-tokens-per-gpu 16384 --log-probs-chunk-size 4096
+)
+SGLANG_ARGS=(
+   --rollout-num-gpus-per-engine 8 --sglang-mem-fraction-static 0.8 --sglang-ep-size 8
+   --sglang-watchdog-timeout 1800 --sglang-enable-metrics --sglang-attention-backend fa3
+   --sglang-cuda-graph-bs 1 2 4 8 16 32 --use-rollout-routing-replay
+   --sglang-mamba-scheduler-strategy extra_buffer
+)
+MISC_ARGS=(
+   --attention-dropout 0.0 --hidden-dropout 0.0 --accumulate-allreduce-grads-in-fp32
+   --attention-softmax-in-fp32 --attention-backend flash
+)
+WANDB_ARGS=( --use-wandb --wandb-project miles-opd --wandb-group qwen3.5-35b-opd-${MODE} )
+
+RUNTIME_ENV_JSON="{\"env_vars\": {\"PYTHONPATH\": \"${MILES_DIR}:/root/Megatron-LM/\", \"CUDA_DEVICE_MAX_CONNECTIONS\": \"1\", \"WANDB_API_KEY\": \"${WANDB_API_KEY}\"}}"
+
+cd "${MILES_DIR}"
+ray job submit --address="${RAY_ADDRESS}" --submission-id qwen3.5-opd-${MODE} --no-wait \
+   --runtime-env-json="${RUNTIME_ENV_JSON}" \
+   -- python3 ${MILES_DIR}/train.py \
+   --actor-num-nodes 1 --actor-num-gpus-per-node 8 --num-gpus-per-node 8 --colocate \
+   ${MODEL_ARGS[@]} ${CKPT_ARGS[@]} ${OPD_ARGS[@]} ${ROLLOUT_ARGS[@]} ${OPTIMIZER_ARGS[@]} ${GRPO_ARGS[@]} \
+   ${WANDB_ARGS[@]} ${PERF_ARGS[@]} ${EVAL_ARGS[@]} ${SGLANG_ARGS[@]} ${MISC_ARGS[@]} ${RM_ARGS[@]}
+echo "Submitted Phase-2 OPD (${MODE}) self-distillation (submission-id qwen3.5-opd-${MODE})."
+echo "Watch opd_reverse_kl (>>0 means the teacher differs from the student) and"
+echo "eval/dapo_heldout (student moving toward the teacher's accuracy, with shorter responses)."
diff --git a/examples/on_policy_distillation/qwen3_5_35b_selfdistill/rm.py b/examples/on_policy_distillation/qwen3_5_35b_selfdistill/rm.py
new file mode 100644
index 00000000000..88b8cbd572e
--- /dev/null
+++ b/examples/on_policy_distillation/qwen3_5_35b_selfdistill/rm.py
@@ -0,0 +1,73 @@
+"""Reward functions for the Qwen3.5-35B-A3B self-distillation example.
+
+Two reward functions are provided:
+
+* ``reward_func`` — format-agnostic boxed-answer correctness (0/1). Used as the
+  task reward for Phase 1 (RLVR teacher training) AND as the eval scorer for both
+  phases. It accepts EITHER answer format the DAPO prompt may elicit:
+  ``\\boxed{ANS}`` (via ``grade_answer_verl``) OR ``Answer: ANS`` (via the DAPO
+  ``compute_score``). A clean 0/1 == true accuracy, so ``rollout/raw_reward`` and
+  ``eval/<name>`` are directly interpretable.
+
+* ``reward_func_pure_opd`` — for *pure* on-policy distillation (Phase 2, pure
+  variant). EVAL samples (tagged ``opd_reward_mode=eval_math`` via the eval
+  config's ``metadata_overrides``) are scored for accuracy so ``eval/<name>``
+  still measures real held-out accuracy; TRAINING samples return a constant
+  ``0.0`` so the GRPO task advantage is ~0 and the ONLY learning signal is the
+  teacher's reverse-KL. This isolates OPD's effect for clean attribution.
+
+Why a custom reward instead of ``--rm-type``:
+  - ``--rm-type math`` only extracts ``\\boxed{}``; ``--rm-type dapo`` only the
+    ``Answer:`` form; ``--rm-type deepscaler`` additionally requires a
+    ``</think>`` tag in the response. Qwen3.5 reasons inline (no ``</think>``),
+    so ``deepscaler`` would score every response 0. This format-agnostic reward
+    avoids all three traps.
+
+Place ``examples/on_policy_distillation/qwen3_5_35b_selfdistill`` on PYTHONPATH (or
+run from the miles repo root) and pass e.g.
+``--custom-rm-path examples.on_policy_distillation.qwen3_5_35b_selfdistill.rm.reward_func``.
+"""
+
+from miles.rollout.rm_hub.math_dapo_utils import compute_score as _dapo_score
+from miles.rollout.rm_hub.math_utils import grade_answer_verl
+
+
+def _is_eval_sample(sample) -> bool:
+    md = sample.metadata if isinstance(getattr(sample, "metadata", None), dict) else {}
+    return md.get("opd_reward_mode") == "eval_math"
+
+
+def _is_correct(sample) -> bool:
+    response, label = sample.response or "", sample.label
+    if label is None:
+        return False
+    # boxed form: \boxed{ANS}
+    try:
+        if grade_answer_verl(response, label):
+            return True
+    except Exception:
+        pass
+    # DAPO prompt form: "Answer: ANS"
+    try:
+        result = _dapo_score(response, label)
+        if isinstance(result, dict) and result.get("acc"):
+            return True
+    except Exception:
+        pass
+    return False
+
+
+async def reward_func(args, sample, **kwargs):
+    """Format-agnostic correctness (1.0 / 0.0). Phase-1 task reward + eval scorer."""
+    return 1.0 if _is_correct(sample) else 0.0
+
+
+async def reward_func_pure_opd(args, sample, **kwargs):
+    """Pure-OPD reward: accuracy for eval samples, constant 0.0 for training.
+
+    Training reward is constant so the GRPO task advantage vanishes and the only
+    learning signal is the OPD reverse-KL toward the teacher.
+    """
+    if _is_eval_sample(sample):
+        return 1.0 if _is_correct(sample) else 0.0
+    return 0.0
diff --git a/miles_plugins/mbridge/qwen3_5.py b/miles_plugins/mbridge/qwen3_5.py
index fe1556ef7d9..67d1dcb35f7 100644
--- a/miles_plugins/mbridge/qwen3_5.py
+++ b/miles_plugins/mbridge/qwen3_5.py
@@ -111,6 +111,21 @@ class Qwen3_5Bridge(Qwen2MoEBridge):
         "mlp.experts.linear_fc2": ["mtp.layers.{layer_number}.mlp.experts.down_proj"],
     }
 
+    # Main-layer experts can ALSO ship unfused (per-expert .weight files) — e.g. a
+    # checkpoint round-tripped out through convert_torch_dist_to_hf, whose experts
+    # are stored split rather than as the fused 3-D ``gate_up_proj``/``down_proj``
+    # tensors the public Qwen3.5-35B-A3B release uses. ``_experts_fused()``
+    # autodetects from the safetensor index, mirroring ``_mtp_experts_fused()``.
+    _MLP_EXPERTS_MAPPING_UNFUSED = {
+        "mlp.experts.linear_fc1": [
+            "model.language_model.layers.{layer_number}.mlp.experts.{expert_id}.gate_proj.weight",
+            "model.language_model.layers.{layer_number}.mlp.experts.{expert_id}.up_proj.weight",
+        ],
+        "mlp.experts.linear_fc2": [
+            "model.language_model.layers.{layer_number}.mlp.experts.{expert_id}.down_proj.weight"
+        ],
+    }
+
     # Override to make ffn_hidden_size optional (Qwen3.5 MoE has no intermediate_size)
     _CONFIG_MAPPING = {
         "num_layers": "num_hidden_layers",
@@ -180,11 +195,34 @@ def _get_gptmodel_args(self) -> dict:
             ret["mtp_block_spec"] = mtp_block_spec
         return ret
 
+    def _experts_fused(self) -> bool:
+        """Detect whether MAIN-layer MoE expert weights are fused 3-D tensors.
+
+        Fused (public Qwen3.5-35B-A3B): ``...mlp.experts.gate_up_proj``.
+        Unfused (round-tripped checkpoints): ``...mlp.experts.{i}.gate_proj.weight``.
+        Same lazy/cached resolution from ``safetensor_io.index`` as
+        ``_mtp_experts_fused()``; defaults to fused when the index is unavailable
+        so pre-init access keeps the historical behaviour.
+        """
+        cached = getattr(self, "_experts_fused_cached", None)
+        if cached is not None:
+            return cached
+        io = getattr(self, "safetensor_io", None)
+        index = getattr(io, "index", None) if io is not None else None
+        if not index:
+            return True
+        fused = any("model.language_model.layers." in k and k.endswith("mlp.experts.gate_up_proj") for k in index)
+        self._experts_fused_cached = fused
+        return fused
+
     def _weight_name_mapping_mlp(self, name: str) -> list[str]:
-        """Override to handle fused expert weights."""
+        """Override to handle fused (default) or unfused per-expert weights."""
         layer_number = name.split(".")[2]
+        mapping = self._MLP_MAPPING
+        if "mlp.experts.linear_fc" in name and not self._experts_fused():
+            mapping = {**self._MLP_MAPPING, **self._MLP_EXPERTS_MAPPING_UNFUSED}
         convert_names = []
-        for keyword, mapping_names in self._MLP_MAPPING.items():
+        for keyword, mapping_names in mapping.items():
             if keyword in name:
                 if "{expert_id}" in mapping_names[0]:
                     expert_id = name.split("weight")[-1]