Notes · Method note 02

Self-play for reasoning

Every self-play method is the same loop — something proposes a task, something solves it, something verifies, and both sides are trained on the result. The four papers below differ in exactly two slots: where the task comes from and who is allowed to be the verifier. Ordered by how much external grounding they keep, from a labeled dataset down to nothing and back up to a corpus.

Fig 1 · The loop, four times

proposersolververifierexternal groundingreward

STaR 2022

arXiv · alphaXiv
labeled questions dataset (fixed) question + answer key model → rationale answer match keep correct rationales → SFT — STaR
  • Sample a rationale + answer; keep the ones whose answer matches; fine-tune on them; repeat.
  • Rationalization: for failures, show the answer as a hint and keep the rationale that reaches it.
  • Not really self-play — the questions and answers are given. The self-generated part is the reasoning trace.
  • Failure mode: nothing forces diversity or difficulty; the loop can converge on easy items.

Absolute Zero May 2025

arXiv · alphaXiv
no external data proposer (same model) Python program task solver (same model) code executor correctness learnability (mid success rate) Absolute Zero
  • One model plays both roles. Tasks are programs in three modes: deduction (program+input → output), abduction (program+output → input), induction (I/O pairs → program).
  • The verifier is a Python executor, so correctness is exact and no data is needed.
  • Proposer is rewarded for tasks the solver sometimes solves — not always, not never.
  • Grounded by execution, but only in the space of things a program can express.

R-Zero Aug 2025

arXiv · alphaXiv
no external data Challenger generated question Solver majority vote (pseudo-label) 1 if matches majority peak at 50% + repetition penalty R-Zero
  • Two copies of the base model co-evolve with GRPO. Challenger is rewarded when Solver's accuracy is \(\approx 50\%\) — maximum uncertainty.
  • No verifier: the label is the Solver's own majority vote over 10 samples. Keep only questions with 3–7 of 10 agreeing.
  • Qwen3-4B-Base: +6.5 math, +7.5 general reasoning after three iterations.
  • Collapses: pseudo-label accuracy decays 79% → 63% by iteration 3; 0.6B models collapse after one round.

SPICE Oct 2025

arXiv · alphaXiv
corpus documents Challenger question w/ extractable answer Reasoner (no doc access) doc-extracted gold 1 if correct variance-based, peak at 50% SPICE
  • Same LLM, two roles, but the Challenger reads a real document and writes a question whose answer is extractable from it. The Reasoner never sees the document — information asymmetry.
  • Grounding fixes R-Zero's failure: the label comes from the text, not from the model agreeing with itself.
  • Both roles trained jointly with Dr. GRPO, separate advantages. Corpora: Nemotron-CC-Math, NaturalReasoning.
  • +8.9 math, +9.8 general reasoning across Qwen3-4B/8B and OctoThinker-3B/8B.

The math in three lines

\[\text{solver:}\qquad \max\ \mathbb{E}_{x\sim\text{proposer}}\big[\,r_{\text{solver}}(y,x)\,\big]\ \ \text{with GRPO}\] // \(r = 1\) if verified correct, else 0 \[\text{proposer:}\qquad \max\ \mathbb{E}_{x}\big[\,f\big(\hat p(x)\big)\,\big]\;-\;\text{repetition penalty}\] // \(\hat{p}\) = the solver's empirical success rate on x; f peaks at \(\hat{p} = \tfrac12\) \[\text{label}(x) \;=\; \text{executor}(x)\ \ \big|\ \ \text{majority vote over solver samples}\ \ \big|\ \ \text{extracted from the source document}\] // the verifier slot — the whole design question

The 50% target is the shared idea. R-Zero rewards the Challenger for questions the Solver gets right about half the time and adds a difficulty band (keep only 3–7 of 10 agreeing); SPICE uses a variance-style reward with the same peak. Absolute Zero calls the same thing learnability. It is the self-play form of the difficulty filters every model report uses — Phi-4's “teachable” seeds, OLMo's 62.5% pass-rate cutoff, Nemotron's “drop what the SFT model always solves.”

Where the four differ

STaRAbsolute ZeroR-ZeroSPICE
Task sourcelabeled datasetself-generated programsself-generated questionsquestions grounded in corpus documents
Verifieranswer keycode executormajority vote of the solver itselfanswer extracted from the document
Difficulty targetingnonelearnability (mid success)peak at 50% + band filtervariance-based, peak at 50%
Algorithmrejection-sampling SFTRL, both rolesGRPO, both rolesDr. GRPO, both roles
Failure modeno pressure toward hard itemsconfined to program-expressible taskspseudo-label decay → collapse in 1–3 itersbounded by corpus quality and coverage

Where it appears in the reports

Stage details for each are on the LLM report diff.

Pitfalls

Sources