Not All Thoughts Are Worth Thinking: Reliability-Aware Test-Time Scaling

Test-time scaling improves LLM reasoning, but concept-guided inference can waste compute on weak or redundant solving directions. We study pre-expansion concept allocation: which explored concepts should receive solution-generation compute? We introduce RATTS, a Reliability-Aware Test-Time Scaling method that combines calibrated predictions of judged concept usefulness with semantic soft coverage. Its objective favors useful, complementary concepts, is monotone submodular, and admits a $(1-1/e)$ greedy approximation guarantee; under an explicit concept-success model, we further derive a conditional bound connecting the surrogate to candidate recall. Across Math500, GPQA Diamond, and HumanEval+ with three generators, RATTS achieves $95.65\%$ macro-average pass-any at $K_{\max }=50$ and $\rho =0.30$, versus $85.73\%$ for score-only selection and $94.45\%$ for uniform-weight semantic coverage. RATTS retains $98.7\%$ of full-expansion pass-any while reducing exploration-plus-expansion tokens by $56.69\%$. Macro-average final accuracy is $74.42\%$ versus $74.08\%$ for full expansion, with a paired difference of $0.34$ pp ($95\%$ CI: $[-0.90,1.62]$). To enable reproduction of these results, we will release the code and experimental artifacts upon acceptance.