STaR, V-STaR, Quiet-STaR — Self-Taught Reasoning

> The smallest possible self-improvement loop sits inside the rationale. A model generates a chain of thought, keeps the ones that land on correct answers, and fine-tunes on those. That is STaR. V-STaR adds a verifier so inference-time selection is better. Quiet-STaR pushes the rationale down to every token. All three work. None of them are magic — the loop preserves any shortcut that happened to reach the right answer.

Type: Learn

Languages: Python (stdlib, bootstrap-loop simulator)

Prerequisites: Phase 13 · 01-03 (Reasoning and CoT), Phase 15 · 01 (long-horizon framing)

Time: ~60 minutes

The Problem

The straightforward way to teach a model to reason is to collect human-written reasoning traces. That is expensive, slow, and bounded by how much high-quality chain-of-thought humans are willing to write.

STaR (Self-Taught Reasoner, Zelikman et al., 2022) asks: what if the model writes its own rationales and grades them against known answers? The loop is:

  1. Sample a reasoning trace plus answer.
  2. If the final answer is correct, keep the trace.
  3. Fine-tune on the kept traces.
  4. Repeat.

It works. GSM8K and CommonsenseQA both improved without new human annotation. But the loop has a built-in bias: any rationale that produced the right answer is retained, regardless of whether the reasoning itself was sound. V-STaR (Hosseini et al., 2024) patches this with a learned verifier; Quiet-STaR (Zelikman et al., 2024) generalizes the idea to per-token internal rationales.

The Concept

STaR: bootstrap on what worked

Start from a base model with some weak reasoning ability. On each training problem, sample a rationale plus answer. If the answer matches the label, keep the (problem, rationale, answer) triple. Fine-tune the model on the kept set. Repeat.

One twist matters. If the model can never get a problem right, the loop cannot learn on it. STaR adds rationalization: for problems the model fails, inject the correct answer as a hint and re-prompt the model to produce a rationale that leads to it. Rationalized rationales are added to the training set.

Result in the original paper (Zelikman et al., 2022): a GPT-J base model improved on GSM8K from 5.8% to 10.7% through repeated STaR rounds with rationalization — about 5 percentage points absolute. On CommonsenseQA, STaR-trained GPT-J 6B reached 72.5%, comparable to a fine-tuned GPT-3 175B (~73%) — a roughly 30x larger model trained on hand-annotated rationales.

V-STaR: train a verifier with DPO

STaR throws away incorrect rationales. Hosseini et al. (2024) observed those are also data: every pair of (rationale, "is this correct") can train a verifier. They use Direct Preference Optimization over both correct and incorrect solutions to build a ranker. At inference time, sample N rationales and pick the verifier's top choice.

Reported delta: +4 to +17 percentage points over prior self-improvement baselines on GSM8K and MATH, with most of the gain coming from using the verifier for inference-time selection rather than for additional generator fine-tuning.

Quiet-STaR: per-token internal rationales

Zelikman et al. (2024) asked: what if the model learns to generate a short internal rationale at every token position, not just between problem and answer? Quiet-STaR trains a model to emit a hidden "thought" before each predicted token, then mixes the thought-aware prediction with the baseline prediction via a learned weight.

Result: Mistral 7B gained absolute zero-shot improvements on GSM8K from 5.9% to 10.9% and CommonsenseQA from 36.3% to 47.2% without task-specific fine-tuning. The model learned "when to think" — hard tokens get longer internal rationales; easy ones get almost none.

Why all three share a safety concern

All three methods use the final answer as the gradient signal. A rationale that reaches the right answer via flawed reasoning — exploiting a shortcut, guessing, or using a non-generalizing pattern — gets positively reinforced. On in-distribution problems the shortcut works. On out-of-distribution problems it breaks silently.

V-STaR's verifier mitigates by learning to rank rationales, but the verifier is trained on the same label set. It can learn to prefer well-formatted wrong reasoning over honest uncertainty. The safer design is to combine STaR-style data with (a) process-supervised reward models (rewarding intermediate steps, not just answers) and (b) held-out OOD evaluation that breaks simple shortcuts.

Comparison

Method Training signal Inference cost Data waste Known failure mode
STaR keep (rationale, answer) if correct 1x discards all incorrect rationales shortcut rationales
STaR + rationalization above + correct-answer hinted retries 1x less rationalized rationales may be implausible
V-STaR STaR + DPO verifier from both classes Nx (best-of-N) minimal verifier can reinforce confident wrongness
Quiet-STaR per-token rationale + mixing weight 1.5-3x minimal still answer-conditioned gradient

Where this sits in the 2026 stack

STaR is old. But the pattern reappears everywhere in 2025-2026. RL on verifiable math problems (DeepSeek-R1, Kimi-k1.5, o1) is STaR's answer-conditioned gradient signal, scaled up. Process reward models (Lightman et al., 2023; OpenAI's "Let's verify step by step") are the process-supervised alternative. AlphaEvolve (Lesson 3) is STaR for code, with a program evaluator instead of a label. Darwin Godel Machine (Lesson 4) is STaR for the agent scaffolding itself.

Understanding STaR makes all of these click. It is the minimum-viable self-improvement loop.

Use It

code/main.py runs a simulated STaR loop on a toy arithmetic task. You can watch:

Ship It

outputs/skill-star-loop-reviewer.md helps you audit a proposed self-taught-reasoning pipeline before you train on it.

Exercises

  1. Run the simulator. Set the shortcut frequency to zero, then to 0.4. How much does final accuracy diverge between the two runs, even though both hit >90% on the training distribution?
  1. Add a held-out OOD test to the simulator. Draw problems from a different distribution and evaluate the bootstrapped model on both in-distribution and OOD sets. Quantify the gap.
  1. Read the Quiet-STaR paper (arXiv:2403.09629) Section 3. Explain the "end-of-thought" token and the mixing-weight head in three sentences each.
  1. Compare STaR's keep-if-correct filter to a process-supervised alternative that rewards each rationale step independently. Identify the labelling cost difference and the plausible quality difference.
  1. Design one evaluation that would catch shortcut rationales in a deployed model. It does not have to be perfect — it has to break the simplest shortcuts a STaR loop would reinforce.

Key Terms

Term What people say What it actually means
STaR "Self-Taught Reasoner" Fine-tune on model-generated rationales that land correct answers; repeat
Rationalization "Hinted retry" Inject the correct answer and re-prompt for a rationale on problems the base model fails
V-STaR "Verifier STaR" DPO-train a verifier on both correct and incorrect rationales, use it for inference-time selection
Quiet-STaR "Per-token rationales" Generate hidden thoughts at every token position; mix with baseline prediction
Answer-conditioned gradient "Outcome-based signal" The training loop rewards final answers, not reasoning steps
Process reward model "Step-level verifier" Reward model trained on per-step correctness, not outcome — contrasts with STaR
Shortcut rationale "Right answer, wrong reasoning" A rationale that reaches the label via a non-generalizing pattern; STaR keeps these

Further Reading