SVR — Adaptive Test-Time Compute: When the Model Decides How Much to Think

The problem: fixed thinking budgets burn tokens

Reasoning models follow a simple rule: the more thinking tokens, the better the answer. The catch is that budgets are usually fixed – the model gets the same allowance for a trivial calculation as for a hard proof. Easy tasks finish after three tokens, yet you keep paying for empty chain-of-thought. The alternative, an external verifier saying “stop”, means another model in the loop: extra cost and latency.

SVR (Self-Verifying Refinement) from Sun Yat-sen University and the X-Era AI Lab attacks both problems at once: the model decides when to stop thinking, with no external judge.

Verdict + confidence as a compute-control policy

At each turn the model generates a solution plus two extra signals: a discrete verdict (Correct / Incorrect) and a confidence score. It stops only when the verdict is Correct and confidence exceeds a threshold; otherwise it refines its own solution and the loop repeats.

The key design decision: ground-truth correctness is used only to construct training rewards. It never enters the refinement prompt and is not needed at inference. The model must learn to verify its own answers and allocate compute from that – hence the “oracle-free” label.

Training: GRPO on fixed-horizon trajectories

Everything is trained with GRPO on fixed-horizon trajectories, with rewards for three things at once:

• solution correctness – whether the final answer is right,
• calibration-aware self-verification – whether the model’s verdict matches reality,
• a “stop-ready” state when the answer is correct.

The model learns not just to solve tasks, but to know that it solved them. Adaptive stopping is activated only at inference – training always runs the full trajectory.

Results: 2.99 turns instead of 10

On Qwen3.5-2B across seven math benchmarks, SVR reaches 0.563 macro-average accuracy with an average of 2.99 inference turns. In the full comparison it beats standard GRPO, strong multi-turn baselines, and a fixed-budget oracle-guided score-feedback reference – while using substantially fewer turns than a rigid ten.

In practice: better quality than the fixed ten-turn budget variant, at roughly three turns on average. Test-time compute is the dominant cost driver for thinking models today, so that ratio shows up on the inference bill.

What it means for engineers

• Reasoning inference gets cheaper. An adaptive budget saves on easy prompts without touching quality on hard ones. If you host thinking models, watch this direction – and start measuring how many tokens you actually burn on simple queries.
• Fewer external verifiers. As self-verification matures, the extra judge model and the whole best-of-n machinery may go away. One loop, one model.
• Mind the domain. Verdict and confidence are natural for tasks with checkable answers. Outside math (coding, agents) this is a harder problem – do not extrapolate these numbers blindly.

This content was generated using artificial intelligence and reviewed by a human editor.

Sources

• SVR: Self-Verifying Refinement via Joint Verdict-Confidence Reinforcement Learning for Adaptive Test-Time Compute, arXiv:2607.28457 (2026-07-30) – https://arxiv.org/abs/2607.28457
• China AI Weekly Brief, 02.08.2026 (job 3d03e2d4fae9) – weekly review of the Chinese AI scene

---

This article is also available in Polish (PL).