PaperlayerView exact revision on arXiv

Source-linked AI summary

Share More, Search Less: Collaborative Parallel Thinking for Efficient Test-Time Scaling

Xinglin Wang, Hao Lin, Shaoxiong Feng, Peiwen Yuan, Yiwei Li, Jiayi Shi, Yueqi Zhang, Chuyi Tan, Ji Zhang, Boyuan Pan, Yao Hu, Kan Li

arXiv:2605.27030v1cs.CL

TL;DR

Existing parallel test-time scaling keeps branch discoveries private, causing redundant exploration and slower accumulation of decision information. CPT enables search-time sharing across branches, establishing a stronger accuracy–latency Pareto frontier than strong baselines across rollout budgets and model scales.

  • Problem

    Existing parallel TTS methods leave branch-level discoveries unshared during search, limiting information utilization and causing redundant rediscovery.

  • Method

    CPT is a training-free inference framework that extracts compact information into a deduplicated query-level pool and adaptively broadcasts it across parallel branches.

  • Results

    CPT establishes a stronger accuracy–latency Pareto frontier than strong baselines across rollout budgets, benchmarks, and model scales.

  • Takeaways & Limitations

    Search-time collaboration is an effective direction for improving the efficiency of parallel TTS.

  • Takeaways & Limitations

    Fixed-step sharing cannot prevent duplicate discoveries within a step, and context broadcasting can add FLOPs and latency through context re-prefilling.

Abstract

from arXiv · show

Test-Time Scaling (TTS) enhances the reasoning capabilities of large language models by allocating additional inference compute to explore the solution space. However, existing parallel TTS methods typically keep branches isolated during search: intermediate discoveries remain branch-private and cannot guide other branches in time. This information isolation causes substantial redundant exploration, as branches repeatedly rediscover information already found elsewhere and require more search steps to collect complete decision information needed to reach correct answers. To bridge this gap, we propose \textbf{Collaborative Parallel Thinking (CPT)}, a training-free inference framework that enables search-time information sharing across parallel branches. CPT extracts compact intermediate information from ongoing branches, maintains a deduplicated query-level information pool, and broadcasts pool entries through the input context, allowing each branch in subsequent search steps to reuse discoveries made by other branches rather than rediscover the same information. Empirically, experiments on HMMT and AIME benchmarks show that CPT establishes a stronger accuracy--latency Pareto frontier than strong baselines across rollout budgets and model scales, highlighting search-time collaboration as an effective direction for efficient parallel TTS.

1 Introduction

The introduction identifies information isolation as a bottleneck in parallel test-time scaling: branch-private discoveries cause redundant rediscovery and slow accumulation of decision information. It proposes Collaborative Parallel Thinking (CPT), a training-free method that shares deduplicated intermediate information across branches and improves the accuracy–latency tradeoff on mathematical reasoning benchmarks.

  • Problem: Parallel TTS accelerates solution-space exploration by generating multiple reasoning branches, but isolated branches repeatedly rediscover information and slow convergence toward correct answers.This bottleneck produces diminishing marginal information gain as search progresses.
  • Method: CPT extracts compact decision-relevant information from ongoing branches, deduplicates it in a query-level pool, and broadcasts pool entries through later input contexts.Branches retain their private reasoning histories while reusing discoveries made elsewhere.
  • Evaluation: CPT is evaluated on HMMT24, HMMT25, AIME24, AIME25, and AIME26 across rollout budgets and policy models.The evaluation targets challenging mathematical reasoning benchmarks.
  • Results: CPT establishes a stronger accuracy–latency Pareto frontier than strong baselines across rollout budgets and model scales.The reported gains include higher accuracy at matched wall-clock latency, attributed to reduced redundant exploration.

2 Related Work

Related work organizes test-time scaling by how additional inference compute is structured, contrasting sequential or guided search with parallel reasoning. Separate work studies collaboration among language-model instances through debate, communication-efficient interaction, and multi-agent aggregation.

  • Test-Time Scaling: Test-time scaling enhances reasoning by allocating additional inference compute, with methods differing in how that compute is organized.The passage identifies test-time scaling as a critical avenue for improving reasoning capabilities.
  • Test-Time Scaling: Sequential refinement extends or revises one reasoning trajectory, whereas structured search explores reasoning states using methods such as Tree-of-Thoughts and MCTS-style decoding.These approaches steer computation toward more promising reasoning states.
  • Test-Time Scaling: Sequential refinement, intermediate evaluation, and verifier-guided control can lengthen the critical inference path and add latency overhead, while parallel TTS generates multiple branches simultaneously.Parallel TTS is presented as an approach for accelerating search by processing multiple reasoning branches for the same problem at once.
  • Collaborative Reasoning with LLMs: Collaborative reasoning among LLM instances includes debate and discussion, where agents exchange arguments, critiques, or proposals to support divergent reasoning, consensus, or improved decisions.Related analyses examine when interaction helps relative to single-agent prompting or voting-based aggregation.
  • Collaborative Reasoning with LLMs: Efficiency-oriented collaboration research studies sparse communication, token-efficient debate, group discussion, selective debate triggering, and multi-agent aggregation or orchestration.The passage places these methods beyond debate within broader scalable multi-agent collaboration frameworks.

3 Methodology

CPT turns parallel test-time scaling into a search-time information-sharing process by extracting compact discoveries into a deduplicated pool and broadcasting them across branches. Its methodology combines synchronized parallel search, shared information pooling, and adaptive scheduling while preserving each branch’s private reasoning trajectory.

  • Collaborative Parallel Search: CPT synchronizes K parallel branches at fixed-token boundaries, sharing context between steps while preserving each branch’s private reasoning history.Each search step generates a chunk of length C, after which newly generated segments enter the pooling stage.
  • Shared Information Pooling: CPT extracts compact information units—such as conclusions, constraints, observations, counterexamples, and checks—from otherwise long and noisy reasoning trajectories.The extracted units are designed for pooling and broadcasting under a limited context budget.
  • Shared Information Pooling: A deduplicated query-level pool filters semantically similar discoveries using embedding similarity threshold τdup, preventing repeated information from consuming shared context.When broadcasting is active, CPT samples at most M pool entries and provides the same updated shared-information context to all branches.
  • Adaptive Broadcast Scheduling: CPT schedules broadcasting from the rate of newly admitted pool entries, avoiding premature diversity loss and late synchronization overhead.The method averages new-entry counts over windows of W steps and compares marginal information gain with a first-window reference.
  • Adaptive Broadcast Scheduling: Using τstop < τstart, CPT transitions from independent probing to broadcast when rj < τstart and then to free-run decoding when rj < τstop.The probe phase still writes extracted information into the pool; the free-run phase stops synchronization until branch termination.

4 Experiments

Experiments on HMMT and AIME benchmarks evaluate CPT across reasoning-focused Qwen3 models and parallel rollout budgets. CPT consistently improves the accuracy–latency trade-off, while reducing redundant exploration through adaptive information sharing.

  • Benchmarks and Models: CPT is evaluated on HMMT24, HMMT25, AIME25, and AIME26 using QWEN3-4B-THINKING-2507 and QWEN3-30B-A3B-THINKING-2507.These benchmarks test complex multi-step reasoning under practical latency budgets.
  • Baselines and Protocol: The evaluation compares CPT with Base Parallel Sampling, DeepConf, and LeaP across rollout budgets K ∈ {8, 16, 32, 64, 128}.Latency is measured from parallel decoding start until all branches finish, with Pass@1 and MV@K reported after answer normalization.
  • Main Results: CPT consistently establishes a stronger accuracy–latency Pareto frontier than the baselines across rollout budgets, models, and AIME and HMMT benchmarks.Simultaneous gains in Pass@1 and MV@K indicate improvements to individual branch quality and parallel exploration effectiveness.
  • Information Redundancy Reduction: CPT reduces per-step duplicate information while preserving newly discovered information, limiting later computation spent rediscovering findings from other branches.Independent parallel sampling produces more duplicate units and fewer newly discovered units as search proceeds.
  • Adaptive Broadcast Scheduling: Moderate early stopping of synchronization generally provides a better accuracy–latency trade-off than never stopping or stopping too aggressively.Immediate broadcasting also does not yield the best accuracy, indicating that overly early sharing can prematurely align branch trajectories.
  • Sensitivity and Runtime: Increasing broadcast size M generally improves Pass@1 and reduces generated tokens and latency, but the benefit saturates at larger M; the default is M = 512 and τdup = 0.75.CPT’s prompt-level broadcasting adds a FLOPs trade-off because updated contexts may require re-prefilling in standard decoding implementations.

5 Conclusions

The paper identifies information isolation as a bottleneck in parallel test-time scaling and introduces CPT to enable branches to share and reuse decision-relevant information during search. Evaluations show that CPT achieves a stronger accuracy–latency Pareto frontier than strong baselines across rollout budgets and model scales.

  • 5 Conclusions: CPT addresses branch-private discoveries that cause redundant rediscovery and slower convergence toward correct answers.The framework enables information sharing across parallel branches during search.
  • 5 Conclusions: CPT establishes a stronger accuracy–latency Pareto frontier than strong baselines across rollout budgets and model scales.This result comes from evaluations on challenging mathematical reasoning benchmarks.
  • 5 Conclusions: Parallel TTS efficiency depends on sharing and reusing decision-relevant information, not only on launching more reasoning branches.The conclusion emphasizes search-time collaboration as an efficiency factor.

Limitations … B.1 Information Statistics during Parallel Search

CPT combines fixed-token parallel generation with a deduplicated shared information pool, broadcasting discoveries across branches during search. Its information-statistics experiment measures newly discovered and duplicate information under a controlled 64-branch, 1024-token-step protocol, while synchronization and context-update overhead limit immediate reuse.

  • Limitations: The fixed-token synchronous protocol delays within-step sharing until the next search step, allowing duplicate discoveries before the next broadcast.The protocol balances timely information reuse against synchronization and context-update overhead.
  • A Collaborative Parallel Thinking Algorithm: CPT maintains private branch histories and a query-level shared pool that stores compact, semantically deduplicated information extracted from generated segments.At each search step, unfinished branches decode up to C tokens conditioned on the current broadcast set, then the same policy model extracts candidate information units.
  • A Collaborative Parallel Thinking Algorithm: CPT alternates between PROBE, BROADCAST, and FREERUN modes according to windowed relative information gain, eventually stopping synchronization while preserving private reasoning histories.Broadcasting begins below τstart, and FREERUN begins below τstop; remaining branches decode until EOS or Lmax.
  • B.1 Information Statistics during Parallel Search: The information-statistics experiment uses HMMT24–25, QWEN3-30B-A3B-THINKING-2507, 64 parallel reasoning branches, and 1024-token generation chunks per search step.The experiment applies CPT’s information-extraction prompt and embedding-based semantic deduplication protocol.
  • B.1 Information Statistics during Parallel Search: Candidate units below the deduplication similarity threshold count as newly discovered information, whereas semantically similar units count as duplicate information for that step.Admission depends on whether maximum similarity to existing entries is below the deduplication threshold.
  • B.1 Information Statistics during Parallel Search: 64 × 1024 generated tokens are normalized to 10,240 generated tokens by multiplying each per-step count by 10,240/(64 × 1024).The normalized measure reports information units per 10,240 generated tokens.

B.2 Offline Information Injection

The offline information-injection experiment tests whether information scattered across reasoning branches remains decision-useful when shared through a deduplicated pool. It varies the proportion of injected information from 0% to 100% using CPT’s extraction and semantic de-duplication procedures.

  • Experiment setup: The experiment evaluates whether information originally scattered across reasoning branches is decision-useful after sharing.Parallel reasoning trajectories are collected before information is merged for injection.
  • Experiment setup: Compact information units are extracted with CPT’s prompt and merged into a deduplicated information pool using the same semantic de-duplication protocol.This mirrors CPT’s information-processing procedure.
  • Injection ratio: The injection ratio ρ is varied across 0%, 20%, 40%, 60%, 80%, and 100% through random sampling.For each problem, the experiment randomly samples the selected percentage of information for injection.

C Analysis of Generated Tokens and FLOPs · C.1 Additional Accuracy–Cost Results

CPT generally improves the accuracy–tokens trade-off by sharing information across parallel branches, while prompt-level broadcasts add prefilling computation that can reduce its FLOPs efficiency. Together, these results show that CPT’s computational benefits depend on the cost metric.

  • C.1 Additional Accuracy–Cost Results: CPT generally achieves a stronger accuracy–tokens trade-off by reducing redundant decoding from repeated rediscovery across parallel branches.The token-based comparison complements latency evaluation by measuring generated computation.
  • C.1 Additional Accuracy–Cost Results: Prompt-level broadcasts introduce additional prefilling computation, making FLOPs efficiency less favorable than token efficiency.Updating the input context may require standard decoding implementations to recompute the prefix.
  • C.1 Additional Accuracy–Cost Results: CPT’s search procedure initializes parallel branch histories and a shared pool, then broadcasts selected pool entries to unfinished branches.The algorithm samples at most M pool entries when the pool exceeds the broadcast size.
  • C.1 Additional Accuracy–Cost Results: CPT transitions from probing to broadcasting and then freerunning when measured reference gains fall below configured thresholds.In freerun mode, unfinished branches decode independently until EOS or Lmax without further synchronization.
  • C.1 Additional Accuracy–Cost Results: CPT can be less favorable under a pure FLOPs budget because context changes may trigger prefix recomputation.This limitation motivates cache-aware implementations as an important direction for improving FLOPs efficiency.
  • C.1 Additional Accuracy–Cost Results: Figure 4 compares accuracy and generated tokens across models and benchmarks under different rollout budgets.The figure provides the token-based comparison supporting CPT’s token-efficiency analysis.

C.2 FLOPs Estimation … D.2 From path-wise information to pooled global information

The paper estimates inference FLOPs with a unified request-level decomposition and formalizes why pooling information across parallel paths can improve decision efficiency. Its analysis separates pooled global decision information from redundant path-wise information and relates redundancy to conditional dependence.

  • C.2 FLOPs Estimation: C.2 defines a unified FLOPs estimator over model requests, separating effective linear token costs from quadratic causal-attention costs.The estimator follows the methodology used in the NVIDIA NeMo framework.
  • C.2 FLOPs Estimation: CPT’s cost includes normal reasoning generation and blackboard information extraction, with extraction estimated under cached decoding so only newly generated tokens contribute to the linear term.The framework separately defines normal reasoning requests and information-extraction requests.
  • D Theory: Collaborative Sharing Improves Global Decision Information: D’s central observation is that isolated parallel paths can rediscover overlapping conclusions, constraints, counterexamples, or failure patterns instead of increasing pooled global decision information.The analysis abstracts away implementation details such as broadcast timing and the number of broadcast items.
  • D.1 Decision information in parallel reasoning: D.1 defines pooled global decision information as G(Z1:K | S, q) := I(Y ; Z1:K | S, q), measuring information gained after pooling all paths into shared context.The target answer Y is uncertain to the solver before reasoning, despite being deterministic in ground truth.
  • D.1 Decision information in parallel reasoning: D.1 contrasts pooled information with aggregate path-wise information, which counts each path separately and therefore includes repeated information across isolated paths.Redundant local information is defined as R(Z1:K | S, q) := L(Z1:K | S, q)−G(Z1:K | S, q), under the redundancy-dominated assumption.
  • D.1 Decision information in parallel reasoning: Under log-loss, maximizing pooled global decision information is equivalent to maximizing the expected log-loss reduction of an optimal Bayesian decision rule.The equivalence follows from H(Y | S, q) −H(Y | A, S, q) = I(Y ; A | S, q).
  • D.2 From path-wise information to pooled global information: D.2 decomposes aggregate path-wise information into pooled global decision information plus redundant local information, explaining why repeated discoveries inflate local totals without improving the global decision.Proposition 6 identifies redundant local information with conditional total correlation under the clean model of mutual conditional independence given Y, S, q.

D.3 When does collaborative sharing help?

Theorem 7 decomposes collaborative sharing’s gain in pooled global decision information into changes in aggregate path-wise information and redundant local information. It shows collaboration helps when redundancy falls without an offsetting loss in path-level exploration, but provides no unconditional dominance guarantee.

  • D.3 When does collaborative sharing help?: Theorem 7 expresses the difference between collaborative and independent pooled global decision information as a decomposition involving path-wise and redundant local information.The comparison conditions both exploration strategies on the same current state S and query q.
  • D.3 When does collaborative sharing help?: Collaborative sharing improves pooled global decision information when the reduction in redundant local information outweighs any decrease in aggregate path-wise information.The theorem states this as a sufficient condition rather than an unconditional guarantee.
  • D.3 When does collaborative sharing help?: Collaboration is beneficial in particular when it reduces redundant local information without reducing aggregate path-wise information.Under this condition, the decomposition implies a positive collaborative gain in pooled global decision information.
  • D.3 When does collaborative sharing help?: Sharing can help by converting path-wise discoveries into pooled global decision information, but excessive shared-context effects can reduce path-level exploration enough to erase the gain.The decrease in aggregate path-wise information may offset the reduction in redundancy.

D.4 A Gaussian redundancy model · E Prompts

The Gaussian redundancy model shows that correlated path information reduces the effective width of parallel exploration, limiting the value of additional rollouts. The prompts operationalize CPT through worker reasoning, blackboard-note extraction, and broadcasted shared information.

  • D.4 A Gaussian redundancy model: The model represents each path as a latent decision signal plus common and path-specific noise, with independent noise components.Common noise captures shared uncertainty or correlated search behavior, while path-specific noise captures idiosyncratic uncertainty that independent paths can average out.
  • D.4 A Gaussian redundancy model: Proposition 8 defines effective parallel width Keff to quantify how cross-path noise correlation reduces nominal parallel information.The result distinguishes the number of rollout paths from the amount of nonredundant information they provide.
  • D.4 A Gaussian redundancy model: When ρ = 0, paths contribute complementary observations and effective width matches nominal width K.Independent path-level noises prevent redundancy in this model.
  • D.4 A Gaussian redundancy model: As ρ → 1, many parallel paths behave like a single effective information source.High cross-path dependence makes additional paths provide little new decision information.
  • D.4 A Gaussian redundancy model: For fixed ρ > 0, increasing K has diminishing information returns because correlated residual search information does not accumulate linearly.The model motivates collaboration that reduces residual-information correlation after conditioning on shared state.
  • E Prompts: CPT uses a worker prompt for each reasoning branch and a blackboard-write prompt to distill reusable notes from partial branch transcripts.Both prompts are used by the same policy model in the described implementation.
  • E Prompts: Worker inputs may include blackboard broadcasts, serialized as a leading system message before each branch continues its private reasoning trace.This placement ensures shared notes appear before subsequent branch reasoning.
  • E Prompts: The blackboard-write system prompt instructs the model to extract concise, reusable, and conservative notes from partial branch transcripts.The default mathematical answer prompt is appended to the worker prompt configuration.
Loading 2605.27030v1…