Pith Number

pith:R2MVQSCX

pith:2026:R2MVQSCXWXV7LP42AYNCDGFLR6

not attested not anchored not stored refs pending

Reinforcement-aware Knowledge Distillation for LLM Reasoning

Dhananjay Ram, Shuli Jiang, Shuo Yang, Stefano Soatto, Wei Xia, Yantao Shen, Yuting Zhang, Zhaoyang Zhang, Zhuowen Tu

RLAD enables better distillation of reasoning LLMs by imitating the teacher selectively during policy updates.

arxiv:2602.22495 v3 · 2026-02-26 · cs.LG · cs.AI

Open paper page JSON Open Graph Bundle Merged state Verified badge What is a Pith Number?

Add to your LaTeX paper

\usepackage{pith}
\pithnumber{R2MVQSCXWXV7LP42AYNCDGFLR6}

Prints a linked badge after your title and injects PDF metadata. Compiles on arXiv. Learn more · Embed verified badge

Record completeness

1 Bitcoin timestamp

2 Internet Archive

3 Author claim open · sign in to claim

4 Citations open

5 Replications open

✓ Portable graph bundle live · download bundle · merged state

The bundle contains the canonical record plus signed events. A mirror can host it anywhere and recompute the same current state with the deterministic merge algorithm.

Claims

C1strongest claim

Across diverse logic reasoning and math benchmarks, RLAD consistently outperforms offline distillation, standard GRPO, and KL-based on-policy teacher-student knowledge distillation.

C2weakest assumption

That guiding the student toward the teacher only when it improves the current policy update will reliably avoid distribution mismatch and objective interference without introducing new instabilities or requiring additional hyperparameter tuning.

C3one line summary

RLAD replaces standard KL-based distillation with Trust Region Ratio Distillation, a PPO-style likelihood ratio objective that performs advantage-aware imitation on student rollouts and outperforms offline KD, GRPO, and KL on-policy KD on logic and math benchmarks.

Formal links

2 machine-checked theorem links

Cited by

4 papers in Pith

Multi-Rollout On-Policy Distillation via Peer Successes and Failures

Self-Distilled RLVR

Skill-SD: Skill-Conditioned Self-Distillation for Multi-turn LLM Agents

Structured Role-Aware Policy Optimization for Multimodal Reasoning

Receipt and verification

First computed	2026-06-19T16:12:18.946766Z
Builder	pith-number-builder-2026-05-17-v1
Signature	Pith Ed25519 (`pith-v1-2026-05`) · public key
Schema	pith-number/v1.0

Canonical hash

8e99584857b5ebf5bf9a061a2198ab8f92e0fefc8f5199cd514f398f85b9462c

Aliases

arxiv: 2602.22495 · arxiv_version: 2602.22495v3 · doi: 10.48550/arxiv.2602.22495 · pith_short_12: R2MVQSCXWXV7 · pith_short_16: R2MVQSCXWXV7LP42 · pith_short_8: R2MVQSCX

Agent API

Resolver JSON Graph JSON Events JSON Schema Signing key

Verify this Pith Number yourself

curl -sH 'Accept: application/ld+json' https://pith.science/pith/R2MVQSCXWXV7LP42AYNCDGFLR6 \
  | jq -c '.canonical_record' \
  | python3 -c "import sys,json,hashlib; b=json.dumps(json.loads(sys.stdin.read()), sort_keys=True, separators=(',',':'), ensure_ascii=False).encode(); print(hashlib.sha256(b).hexdigest())"
# expect: 8e99584857b5ebf5bf9a061a2198ab8f92e0fefc8f5199cd514f398f85b9462c

Canonical record JSON

{
  "metadata": {
    "abstract_canon_sha256": "9fc3211235a91ef9f4e24e5340f6e9b9d74918b26850234a2a52ae235f7949f6",
    "cross_cats_sorted": [
      "cs.AI"
    ],
    "license": "http://creativecommons.org/licenses/by/4.0/",
    "primary_cat": "cs.LG",
    "submitted_at": "2026-02-26T00:20:39Z",
    "title_canon_sha256": "5c443d568afefd33c13d07cff1913eafe48bc17f95dc977340c2973acbd21553"
  },
  "schema_version": "1.0",
  "source": {
    "id": "2602.22495",
    "kind": "arxiv",
    "version": 3
  }
}