PaperlayerView exact revision on arXiv

Source-linked AI summary

HRBench: Benchmarking and Understanding Thinking-Mode Switch Strategies in Hybrid-Reasoning LLMs

Yansong Ning, Mianpeng Liu, Jingwen Ye, Weidong Zhang, Hao Liu

arXiv:2605.28398v1cs.AI

TL;DR

Adaptive thinking-mode methods are hard to compare because prior evaluations use incompatible models, datasets, and implementation settings. HRBench unifies three switching strategies and four training regimes across models and reasoning tasks, finding distinct trade-offs and strategy preferences that shift with model scale and task domain.

  • Problem

    Existing adaptive thinking-mode methods are evaluated under incomparable models, datasets, metrics, and decoding configurations, obscuring cross-strategy and training comparisons.

  • Method

    HRBench evaluates 12 strategy–training configurations across six LLMs and five benchmarks while integrating 12+ existing methods into one pipeline.

  • Results

    Prompt-Tuning tends toward favorable accuracy–token trade-offs, Routing toward stable cost reduction, and Speculative toward higher accuracy at greater token cost, with preferences shifting by scale and domain.

  • Takeaways & Limitations

    Choosing a switching strategy should account for both model scale and expected task domain because no single strategy dominates universally.

  • Takeaways & Limitations

    Training-based evaluations are limited to Qwen3.5-9B, and multi-turn, agentic, creative-writing, and multilingual settings remain unexplored.

Abstract

from arXiv · show

Hybrid-reasoning large language models (LLMs) expose explicit controls over reasoning effort, allowing users or systems to trade off answer quality against inference cost. However, existing methods for adaptive thinking-mode selection are typically evaluated under different models, datasets, and implementation assumptions, making it difficult to compare their practical behavior. We introduce HRBench, a unified evaluation framework for studying thinking-mode switching in hybrid-reasoning LLMs. HRBench organizes the design space along two axes: three switching strategy families, prompt-based selection, external routing, and speculative execution, and four training regimes, training-free, SFT, offline and online RL, yielding 12 controlled evaluation settings. We evaluate these settings across 6 LLMs, from Qwen3.5-2B to Kimi-K2.5-1.1T, and 5 reasoning benchmarks covering mathematics, science, and code, while reimplementing 12+ representative prior methods within the same pipeline. Our analysis characterizes how different switching strategies occupy distinct effectiveness-efficiency trade-off regions: prompt-based methods often provide favorable token-accuracy trade-offs, routing methods offer more stable cost reduction, and speculative methods tend to improve accuracy at higher token cost. We further find that training affects strategies differently, and that the preferred strategy varies with model scale and task domain. HRBench provides reference implementations and a unified evaluation platform to support more controlled research on efficient reasoning in hybrid-reasoning LLMs. Our data, code and repository are available at https://github.com/usail-hkust/HRBench.

1 Introduction

Hybrid-reasoning LLMs expose controllable reasoning depth, but adaptive switching methods lack comparable evaluation. HRBench unifies strategies, training regimes, models, and tasks to characterize their efficiency–effectiveness trade-offs.

  • Hybrid-reasoning LLMs let users choose between deep reasoning and direct answering or specify reasoning effort and token budgets.
  • The benchmark evaluates six LLMs from 2B to 1.1T across five datasets and 527 experiment runs.
  • Existing adaptive switching methods are difficult to compare because they use different models, datasets, metrics, and decoding configurations.
  • HRBench combines three switching strategies with four training regimes, producing 12 controlled evaluation configurations.
  • Prompt-Tuning offers higher accuracy with fewer tokens, Routing provides moderate savings with preserved accuracy, and Speculative improves accuracy at additional token cost.
  • Training effects and strategy preferences vary with model scale and task domain: Speculative surpasses Prompt-Tuning at 20B and 671B, while Prompt-Tuning favors math and Speculative favors code.

2 Related Work

Prior adaptive thinking-mode methods fall into Prompt-Tuning, Routing, and Speculative strategies, distinguished by when and how mode decisions are made. HRBench addresses their lack of controlled cross-strategy comparison.

  • Prompt-Tuning: Prompt-Tuning lets the model select its reasoning mode within one inference pass through prompt instructions.
  • Routing: Routing uses a two-stage process in which a router assesses query difficulty, selects a mode, and the model generates under that mode.
  • Speculative: Speculative methods begin in a fast mode and switch to deep reasoning during inference after detecting uncertainty signals.
  • Prior methods were evaluated in isolation, leaving no unified framework for controlled comparison across strategies.

3 Preliminary

The preliminary section formalizes thinking modes and adaptive switching, then distinguishes prompt-based, routed, and speculative decisions. HRBench organizes evaluation as a 12-cell strategy–training taxonomy.

  • Definitions: A thinking mode controls the token budget allocated to intermediate chain-of-thought before the final answer.
  • Definitions: Available modes may be binary, discrete effort levels, or continuous token budgets bounded by a maximum.
  • Problem formulation: Given a query and mode set, the model adaptively selects modes to produce a chain-of-thought and final answer.
  • Switching strategies: Prompt-Tuning implicitly selects a mode through a prompt template during inference.
  • Evaluation taxonomy: The HRBench taxonomy crosses three strategies with four training regimes, with each cell representing one configuration.
  • Switching strategies: Routing explicitly maps a query to a mode before generation, using a router policy.
  • Switching strategies: Speculative switching starts with one mode, monitors partial output, and changes to another after a trigger signal.

4 HRBench Construction

HRBench constructs a unified evaluation across models, tasks, switching strategies, training regimes, metrics, and representative external methods. Its 12 configurations systematically compare adaptive thinking-mode selection under identical evaluation conditions.

  • Evaluation taxonomy: 12 configurations cross three switching strategies with four training regimes in a systematic evaluation taxonomy.The strategies are Prompt-Tuning, Routing, and Speculative; regimes include training-free, SFT, offline RL, and online RL.
  • Datasets and tasks: Five benchmarks cover mathematics, science, and code, including AIME 2025, MATH500, GPQA-Diamond, Live Code Bench, and Codeforces.The benchmarks range from competition-level mathematics and programming to graduate-level science questions and execution-based coding evaluation.
  • Models: Six hybrid-reasoning LLMs span 2B to 1.1T parameters and support binary, discrete, or numeric-token-budget thinking-mode controls.Evaluated models include Qwen3.5-2B, Qwen3.5-9B, gpt-oss-20B, Seed-OSS-36B-Instruct, DeepSeek-V3.1-671B, and Kimi-K2.5-1.1T.
  • Metrics and baselines: Accuracy is measured as Pass@1, while token cost is the average output length including chain-of-thought tokens.Fixed baselines include always-thinking, always-no-thinking, and budget-aware reasoning-effort tiers.
  • Reference implementations: Training-free implementations use prompts, an LLM router, or uncertainty triggers and entropy thresholds to select or escalate thinking modes.Prompt-Tuning maps model-specific prompts to effort levels, Routing assesses query difficulty before dispatch, and Speculative methods trigger re-thinking during inference.
  • Training and comparators: Training variants use a unified rejection-fine-tuning rollout pipeline, while external methods are reimplemented and evaluated under identical decoding conditions.The pipeline covers SFT, preference optimization, and on-policy reinforcement-learning variants, and integrates 12 representative community methods.

5 Effectiveness–Efficiency Trade-off of Switching Strategies

The three switching strategies occupy distinct accuracy–token-cost trade-offs, and their relative effectiveness and efficiency vary with model scale and task domain. PT often offers favorable joint gains, RT provides steadier savings, and SPEC exchanges extra tokens for accuracy.

  • 5.1 Overall Trade-off Patterns: PT-TF improves accuracy over Full-Think while reducing token cost, producing a Pareto-optimal “win-win” trade-off.Across PT implementations, this Pareto-dominant behavior holds robustly.
  • 5.1 Overall Trade-off Patterns: RT-TF preserves accuracy comparable to Full-Think while achieving moderate token savings through selective routing.The router sends easier problems to no_think while conservatively retaining full reasoning for harder benchmarks.
  • 5.1 Overall Trade-off Patterns: SPEC increases accuracy at extra token cost, particularly on code tasks where its try-then-verify mechanism is effective.Re-triggering deep reasoning after an uncertainty signal adds overhead, making SPEC an effectiveness-enhancing rather than efficiency strategy.
  • 5.2 Model Scale Effect: At 9B and 1.1T, PT leads with 47.6% and 80.8% effectiveness, while SPEC leads at 20B with 36.8% and 671B with 75.8%.At 2B, all three strategies perform similarly at 13.2–14.1%, so no strategy universally dominates in effectiveness.
  • 5.2 Model Scale Effect: RT reduces token cost at every scale from 9B onward, while PT’s efficiency changes from 29.2k tokens at 2B to −39% savings at 36B.SPEC consistently incurs extra tokens across scales because of its re-think mechanism.
  • 5.3 Task Domain Effect: PT is the clear winner in Math and Science, whereas SPEC achieves the largest accuracy boost in Code through try-then-verify.The domain-dependent preferences motivate selecting switching strategies according to expected task domain and model scale.

6 Effect of Training Pipeline on Switching Strategies

Training improves switching capacity across strategies, primarily by reducing unnecessary reasoning tokens rather than substantially increasing accuracy. The efficiency benefit is strongest for RT, while DPO, GRPO, and SFT emphasize effectiveness, efficiency, and balance respectively.

  • 6 Effect of Training Pipeline on Switching Strategies: Training maintains or slightly improves accuracy while reducing tokens by 12% to 65% across strategies and training methods.Accuracy gains remain within 1–2 percentage points of Training-Free baselines.
  • 6 Effect of Training Pipeline on Switching Strategies: GRPO achieves 65% token reduction for RT, compared with 31% for PT and 23% for SPEC.RT’s binary routing decision aligns closely with the training signal, enabling a sharper difficulty boundary.
  • 6 Effect of Training Pipeline on Switching Strategies: DPO yields the largest accuracy improvements, GRPO achieves the greatest efficiency gains, and SFT provides a balanced middle ground.These regimes therefore offer distinct trade-off profiles for effectiveness, efficiency, and balanced optimization.

7 Fair Comparison of Existing Methods

HRBench’s unified evaluation finds that external methods reproduce strategy-level trade-offs, while no single method dominates across all task domains.

  • External methods exhibit the same strategy-level trade-off patterns identified by HRBench’s controlled evaluation.The comparison re-implements 12 representative methods and evaluates them against training-free baselines under identical conditions.
  • Prompt-Tuning spans the widest Pareto frontier, from RASC at 53.4% accuracy with 5× more tokens to CoD at 34.0% accuracy with +80% token reduction.PT-TF reaches 47.6% accuracy with +24% token reduction, placing it competitively within this range.
  • Routing methods consistently provide 18–21% token savings with 43–45% preserved accuracy.RT-TF achieves 44.1% accuracy with +12.5% token reduction, aligning with the routing pattern.
  • Speculative methods mostly show −11% to −30% token reduction, functioning primarily as accuracy-boosting rather than efficiency-oriented approaches.ADR is the only external speculative method reported to achieve token reduction.
  • No external method dominates every task domain, with RASC best on AIME, Budget-Guidance best on MATH500, and SoT best on GPQA.The reported results support selecting methods according to the target task domain rather than aggregate score alone.

8 Conclusion

HRBench unifies evaluation of thinking-mode switching across strategies, training regimes, models, and datasets. Its findings show distinct trade-offs, strategy-dependent training gains, and choices that shift with scale and domain.

  • HRBench evaluates 12 configurations, combining 3 strategies with 4 training regimes across 6 models and 5 datasets.The framework also integrates 12+ external methods into the same pipeline.
  • The three thinking-mode switching strategies exhibit fundamentally different trade-off profiles.The conclusion identifies strategy differences as a central empirical finding of the benchmark.
  • Training gains are strongly strategy-dependent.The conclusion treats training effects as varying across switching strategies rather than uniformly improving them.
  • The optimal strategy–training combination shifts with model scale and task domain.HRBench releases its implementations and unified evaluation framework for further research on efficient hybrid reasoning.

Limitations

The study’s conclusions are bounded by its training scale, single-turn setting, and coverage of only mathematics, science, and code.

  • Training-based evaluations are limited to Qwen3.5-9B because of computational constraints.Scaling training evaluations to 20B+ models would strengthen conclusions about training-scale interactions.
  • The evaluation focuses on single-turn reasoning, leaving multi-turn and agentic mode-switching scenarios unexplored.The limitation specifically concerns decisions that compound across multiple steps.
  • The benchmark covers mathematics, science, and code, but other domains may exhibit different trade-off patterns.Examples include creative writing and multilingual tasks.

Ethical Statement

The paper uses LLMs for polishing and figure plotting, verifies those outputs, and evaluates only public, non-human-subject data. Its benchmark materials standardize domain-specific prompts and reasoning-mode configurations.

  • LLMs were used for paper polishing and figure plotting, with outputs carefully verified for correctness.
  • The evaluation uses publicly available benchmarks and involves no human subjects, private data, or harmful-content generation.The listed datasets include MATH500, AIME 2025, GPQA-Diamond, LiveCodeBench, and Codeforces.
  • All benchmark strategies share domain-specific answer-format instructions while retaining strategy-specific system prompts.Mathematics uses boxed answers, code requires Python solutions, and science requests an option letter for multiple-choice questions.
  • Prompt-tuning selects reasoning depth within one inference pass, whereas routing uses a judge stage before solving.Speculative strategies instead control switching at the token level through two-pass generation.

C Failure Case Analysis

HRBench identifies distinct failure modes across prompt-based, routing, and speculative switching, showing that errors arise from over-compression, task-specific model limits, or late uncertainty detection.

  • Prompt-Tuning failures: PT-GRPO can over-compress reasoning on hard problems, occasionally skipping intermediate steps required for multi-step proofs.On MATH500, PT-GRPO averages 4.1k tokens versus 6.5k for PT-TF; accuracy is maintained on average, but some problems require four or more reasoning steps.
  • Evaluation framing: HRBench uses accuracy and token cost to evaluate effectiveness-efficiency trade-offs under a unified comparison pipeline.The comparison caption defines Red% as token reduction relative to Full-Think, with positive values indicating savings.
  • Routing failures: RT routing cannot improve tasks where neither think-mode nor no_think-mode responses produce correct answers.On GPQA-Diamond, RT achieves 52.0% accuracy versus 54.5% for Full-Think, despite routing 93% of questions to think-mode.
  • Speculative failures: Speculative triggering can add tokens without correcting a confident initial error when uncertainty is detected only after commitment to a wrong approach.On AIME, all 30 problems entered mixed mode under Spec-Entropy, but retriggering did not correct the underlying reasoning failure.

D.3 Unified Training Data Construction

HRBench constructs unified training data by comparing multiple thinking modes, selecting efficient correct responses, and deriving preference or reinforcement-learning signals for each switching strategy.

  • Shared pipeline: Training-based methods across PT, RT, and speculative strategies use a unified rejection-based data construction process.The pipeline is applied to SFT, DPO, and GRPO methods across all three strategy families.
  • Rollout construction: For each problem, the pipeline generates K responses under every available thinking mode, recording token counts and correctness labels.Available modes can include think, no_think, or intermediate effort levels.
  • SFT construction: SFT targets are the correct responses with minimum token cost, teaching models to produce efficient yet accurate answers.The selected response is converted into training pairs with the original problem.
  • DPO construction: DPO pairs prefer the correct minimum-cost response over longer correct responses or incorrect responses.The chosen response is denoted r+, while rejected responses r− come from either correct-but-longer or incorrect alternatives.
  • GRPO construction: GRPO uses n = 8 rollouts and a reward that weights accuracy and efficiency while gating efficiency bonuses on correctness.The reference is Full-Think token count, with α = 1.0 for accuracy and β = 0.5 for efficiency; advantages use the group mean reward.
  • Strategy-specific construction: Training and inference use identical model-specific system prompts for training-based strategies.PT implementations internalize prompt-guided mode selection, while RT and speculative variants preserve their respective routing or triggering mechanisms.
  • Routing construction: RT training labels select the lowest-cost mode that produces a correct answer, whereas GRPO updates only the router while freezing the backbone.The router learns dispatch decisions from final-answer correctness and token efficiency without modifying the LLM’s reasoning capabilities.
  • Speculative construction: Speculative training evaluates the full two-pass process, rewarding triggers that improve correctness without unnecessary re-thinking overhead.SFT and DPO compare total-token outcomes, while GRPO runs the speculative mechanism end-to-end during rollouts.
Loading 2605.28398v1…