Chuyi Tan

Xinglin Wang, Hao Lin, Shaoxiong Feng et al.

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.

Xinglin Wang, Shaoxiong Feng, Yiwei Li et al.

Self-consistency (SC), a widely used decoding strategy for chain-of-thought reasoning, shows significant gains across various multi-step reasoning tasks but comes with a high cost due to multiple sampling with the preset size. Its variants, Adaptive self-consistency (ASC) and Early-stopping self-consistency (ESC), dynamically adjust the number of samples based on the posterior distribution of a set of pre-samples, reducing the cost of SC with minimal impact on performance. Both methods, however, do not exploit the prior information about question difficulty. It often results in unnecessary repeated sampling for easy questions that could be accurately answered with just one attempt, wasting resources. To tackle this problem, we propose Difficulty-Adaptive Self-Consistency (DSC), which leverages the difficulty information of batch queries from both prior and posterior perspectives to adaptively allocate inference resources, further reducing the overall cost of SC. To demonstrate the effectiveness of DSC, we conduct extensive experiments on three popular categories of reasoning tasks: arithmetic, commonsense and symbolic reasoning on six benchmarks. The empirical results show that DSC consistently surpasses the strong baseline ASC and ESC in terms of costs by a significant margin, while attaining comparable performances.

Peiwen Yuan, Shaoxiong Feng, Yiwei Li et al.

In-Context Learning (ICL) enables large language models (LLMs) to achieve rapid task adaptation by learning from demonstrations. With the increase in available context length of LLMs, recent experiments have shown that the performance of ICL does not necessarily scale well in many-shot (demonstration) settings. We theoretically and experimentally confirm that the reason lies in more demonstrations dispersing the model attention from the query, hindering its understanding of key content. Inspired by how humans learn from examples, we propose a training-free method FocusICL, which conducts triviality filtering to avoid attention being diverted by unimportant contents at token-level and operates hierarchical attention to further ensure sufficient attention towards current query at demonstration-level. We also design an efficient hyperparameter searching strategy for FocusICL based on model perplexity of demonstrations. Comprehensive experiments validate that FocusICL achieves an average performance improvement of 5.2% over vanilla ICL and scales well with many-shot demonstrations.

Xinglin Wang, Jiayi Shi, Shaoxiong Feng et al.

Test-Time Scaling enhances the reasoning capabilities of Large Language Models by allocating additional inference compute to broaden the exploration of the solution space. However, existing search strategies typically treat rollouts as disposable samples, where valuable intermediate insights are effectively discarded after each trial. This systemic memorylessness leads to massive computational redundancy, as models repeatedly re-derive discovered conclusions and revisit known dead ends across extensive attempts. To bridge this gap, we propose \textbf{Recycling Search Experience (RSE)}, a self-guided, training-free strategy that turns test-time search from a series of isolated trials into a cumulative process. By actively distilling raw trajectories into a shared experience bank, RSE enables positive recycling of intermediate conclusions to shortcut redundant derivations and negative recycling of failure patterns to prune encountered dead ends. Theoretically, we provide an analysis that formalizes the efficiency gains of RSE, validating its advantage over independent sampling in solving complex reasoning tasks. Empirically, extensive experiments on HMMT24, HMMT25, IMO-Bench, and HLE show that RSE consistently outperforms strong baselines with comparable computational cost, achieving state-of-the-art scaling efficiency.

Xinglin Wang, Zishen Liu, Shaoxiong Feng et al.

Agentic systems increasingly solve complex user requests by executing orchestrated workflows, where subtasks are assigned to specialized models or tools and coordinated according to their dependencies. While recent work improves agent efficiency by optimizing the performance--cost--latency frontier, real deployments often impose concrete requirements: a workflow must be completed within a specified budget and before a specified deadline. This shifts the goal from average efficiency optimization to maximizing the probability that the entire workflow completes successfully under explicit budget and deadline constraints. We study \emph{constraint-driven online resource allocation for agentic workflows}. Given a dependency-structured workflow and estimates of success rates and generation lengths for each subtask--model pair, the executor allocates models and parallel samples across simultaneously executable subtasks while managing the remaining budget and time. We formulate this setting as a finite-horizon stochastic online allocation problem and propose \emph{Monte Carlo Portfolio Planning} (MCPP), a lightweight closed-loop planner that directly estimates constrained completion probability through simulated workflow executions and replans after observed outcomes. Experiments on CodeFlow and ProofFlow demonstrate that MCPP consistently improves constrained completion probability over strong baselines across a wide range of budget--deadline constraints.

Yiwei Li, Ji Zhang, Shaoxiong Feng et al.

Self-consistency improves reasoning by aggregating diverse stochastic samples, yet the dynamics behind its efficacy remain underexplored. We reframe self-consistency as a dynamic distributional alignment problem, revealing that decoding temperature not only governs sampling randomness but also actively shapes the latent answer distribution. Given that high temperatures require prohibitively large sample sizes to stabilize, while low temperatures risk amplifying biases, we propose a confidence-driven mechanism that dynamically calibrates temperature: sharpening the sampling distribution under uncertainty to align with high-probability modes, and promoting exploration when confidence is high. Experiments on mathematical reasoning tasks show this approach outperforms fixed-diversity baselines under limited samples, improving both average and best-case performance across varying initial temperatures without additional data or modules. This establishes self-consistency as a synchronization challenge between sampling dynamics and evolving answer distributions.

Peiwen Yuan, Shaoxiong Feng, Yiwei Li et al.

The rapid advancement of large language models (LLMs) has led to a surge in both model supply and application demands. To facilitate effective matching between them, reliable, generic and efficient benchmark generators are widely needed. However, human annotators are constrained by inefficiency, and current LLM benchmark generators not only lack generalizability but also struggle with limited reliability, as they lack a comprehensive evaluation framework for validation and optimization. To fill this gap, we first propose an automated and unbiased evaluation framework, structured around four dimensions and ten criteria. Under this framework, we carefully analyze the advantages and weaknesses of directly prompting LLMs as generic benchmark generators. To enhance the reliability, we introduce a series of methods to address the identified weaknesses and integrate them as BenchMaker. Experiments across multiple LLMs and tasks confirm that BenchMaker achieves superior or comparable performance to human-annotated benchmarks on all metrics, highlighting its generalizability and reliability. More importantly, it delivers highly consistent evaluation results across 12 LLMs (0.967 Pearson correlation against MMLU-Pro), while taking only $0.005 and 0.38 minutes per sample.

Xinglin Wang, Yiwei Li, Shaoxiong Feng et al.

Test-Time Scaling (TTS) improves the performance of Large Language Models (LLMs) by using additional inference-time computation to explore multiple reasoning paths through search. Yet how to allocate a fixed rollout budget most effectively during search remains underexplored, often resulting in inefficient use of compute at test time. To bridge this gap, we formulate test-time search as a resource allocation problem and derive the optimal allocation strategy that maximizes the probability of obtaining a correct solution under a fixed rollout budget. Within this formulation, we reveal a core limitation of existing search methods: solution-level allocation tends to favor reasoning directions with more candidates, leading to theoretically suboptimal and inefficient use of compute. To address this, we propose Direction-Oriented Resource Allocation (DORA), a provably optimal method that mitigates this bias by decoupling direction quality from candidate count and allocating resources at the direction level. To demonstrate DORA's effectiveness, we conduct extensive experiments on challenging mathematical reasoning benchmarks including MATH500, AIME2024, and AIME2025. The empirical results show that DORA consistently outperforms strong baselines with comparable computational cost, achieving state-of-the-art accuracy. We hope our findings contribute to a broader understanding of optimal TTS for LLMs.

Peiwen Yuan, Yiwei Li, Shaoxiong Feng et al.

LLM-as-Benchmark-Generator methods have been widely studied as a supplement to human annotators for scalable evaluation, while the potential biases within this paradigm remain underexplored. In this work, we systematically define and validate the phenomenon of inflated performance in models evaluated on their self-generated benchmarks, referred to as self-bias, and attribute it to sub-biases arising from question domain, language style, and wrong labels. On this basis, we propose Silencer, a general framework that leverages the heterogeneity between multiple generators at both the sample and benchmark levels to neutralize bias and generate high-quality, self-bias-silenced benchmark. Experimental results across various settings demonstrate that Silencer can suppress self-bias to near zero, significantly improve evaluation effectiveness of the generated benchmark (with an average improvement from 0.655 to 0.833 in Pearson correlation with high-quality human-annotated benchmark), while also exhibiting strong generalizability.

Peiwen Yuan, Yueqi Zhang, Shaoxiong Feng et al.

Evaluating models on large benchmarks is very resource-intensive, especially during the period of rapid model evolution. Existing efficient evaluation methods estimate the performance of target models by testing them only on a small and static coreset of the benchmark, which is derived from the publicly available evaluation results of source models. These methods rely on the assumption that target models have high prediction consistency with source models. However, we demonstrate that it doesn't generalize well in practice. To alleviate the inconsistency issue, we present TailoredBench, a method that conducts customized evaluation tailored to each target model. Specifically, a Global-coreset is first constructed as a probe to identify the most consistent source models for each target model with an adaptive source model selection strategy. Afterwards, a scalable K-Medoids clustering algorithm is proposed to extend the Global-coreset to a tailored Native-coreset for each target model. According to the predictions on Native-coresets, we obtain the performance of target models on the whole benchmark with a calibrated estimation strategy. Comprehensive experiments on 5 benchmarks across over 300 models demonstrate that compared to best performing baselines, TailoredBench achieves an average reduction of 31.4% in MAE of accuracy estimates under the same inference budgets, showcasing strong effectiveness and generalizability.

Yiwei Li, Jiayi Shi, Shaoxiong Feng et al.

We propose a novel speculative decoding method tailored for multi-sample reasoning scenarios, such as self-consistency and Best-of-N sampling. Our method exploits the intrinsic consensus of parallel generation paths to synthesize high-quality draft tokens without requiring auxiliary models or external databases. By dynamically analyzing structural patterns across parallel reasoning paths through a probabilistic aggregation mechanism, it identifies consensus token sequences that align with the decoding distribution. Evaluations on mathematical reasoning benchmarks demonstrate a substantial improvement in draft acceptance rates over baselines, while reducing the latency in draft token construction. This work establishes a paradigm shift for efficient multi-sample inference, enabling seamless integration of speculative decoding with sampling-based reasoning techniques.

Chuyi Tan, Peiwen Yuan, Xinglin Wang et al.

Reinforcement learning with verifiable rewards (RLVR) scales the reasoning ability of large language models (LLMs) but remains bottlenecked by limited labeled samples for continued data scaling. Reinforcement learning with intrinsic rewards (RLIR), where the policy model assigns rewards to its own rollouts, enables sustainable scaling in unlabeled settings, yet its performance and stability lag behind RLVR. We trace this gap to a system bias: the model tends to overestimate its high-confidence rollouts, leading to biased and unstable reward estimation. This bias accumulates as training progresses, with deviations from the oracle drifting toward over-reward, causing unstable training. We characterize this bias using three metrics: $ρ_{\text{noise}}$, $ρ_{\text{selfbias}}$, and $ρ_{\text{symbias}}$. We find that $ρ_{\text{noise}}$ and $ρ_{\text{symbias}}$ impact convergence, while $ρ_{\text{selfbias}}$ amplifies both correct and incorrect updates, leading to instability. To mitigate this, we propose reinforcement learning with ensembled rewards (RLER), which aggregates diverse models and adapts reward interpolation and rollout selection. Extensive experiments show that RLER improves by +13.6% over RLIR and is only 3.6% below RLVR, achieving stable scaling on unlabeled samples, making it highly applicable.

Ji Zhang, Yiwei Li, Shaoxiong Feng et al.

KV cache in autoregressive LLMs eliminates redundant recomputation but has emerged as the dominant memory and bandwidth bottleneck during inference, notably with long contexts and test-time scaling. KV quantization is a key lever for reducing cache cost, but accuracy drops sharply as the native KV distribution lacks flatness and thus maintains a wide quantization range. Prior work focuses on isolating outliers, which caps their error but fails to flatten the overall distribution, leaving performance fragile under low-bit settings. In this work, we show that the K cache maintains a stable structure that evolves gradually with context, while the V cache carries latent semantic regularities. Building on these insights, we propose PatternKV, a pattern-aligned residual quantization scheme. It mines representative pattern vectors online, aligns each KV vector to its nearest pattern, and quantizes only the residual. This reshaping of the KV distribution flattens the quantization target and narrows its range, thereby improving the fidelity of low-bit KV quantization. Across long-context and test-time scaling settings on multiple backbones, PatternKV delivers consistent 2-bit gains, with a 0.08% average 4-bit drop relative to FP16, improves test-time scaling accuracy by 10% on average, and raises throughput by 1.4x while supporting 1.25x larger batches.

Yueqi Zhang, Peiwen Yuan, Shaoxiong Feng et al.

Human-AI conversation frequently relies on quoting earlier text-"check it with the formula I just highlighted"-yet today's large language models (LLMs) lack an explicit mechanism for locating and exploiting such spans. We formalise the challenge as span-conditioned generation, decomposing each turn into the dialogue history, a set of token-offset quotation spans, and an intent utterance. Building on this abstraction, we introduce a quotation-centric data pipeline that automatically synthesises task-specific dialogues, verifies answer correctness through multi-stage consistency checks, and yields both a heterogeneous training corpus and the first benchmark covering five representative scenarios. To meet the benchmark's zero-overhead and parameter-efficiency requirements, we propose QuAda, a lightweight training-based method that attaches two bottleneck projections to every attention head, dynamically amplifying or suppressing attention to quoted spans at inference time while leaving the prompt unchanged and updating < 2.8% of backbone weights. Experiments across models show that QuAda is suitable for all scenarios and generalises to unseen topics, offering an effective, plug-and-play solution for quotation-aware dialogue.

Peiwen Yuan, Chuyi Tan, Shaoxiong Feng et al.

Despite the rapid progress of large language models (LLMs), their length-controllable text generation (LCTG) ability remains below expectations, posing a major limitation for practical applications. Existing methods mainly focus on end-to-end training to reinforce adherence to length constraints. However, the lack of decomposition and targeted enhancement of LCTG sub-abilities restricts further progress. To bridge this gap, we conduct a bottom-up decomposition of LCTG sub-abilities with human patterns as reference and perform a detailed error analysis. On this basis, we propose MarkerGen, a simple-yet-effective plug-and-play approach that:(1) mitigates LLM fundamental deficiencies via external tool integration;(2) conducts explicit length modeling with dynamically inserted markers;(3) employs a three-stage generation scheme to better align length constraints while maintaining content quality. Comprehensive experiments demonstrate that MarkerGen significantly improves LCTG across various settings, exhibiting outstanding effectiveness and generalizability.

Peiwen Yuan, Shaoxiong Feng, Yiwei Li et al.

Human preference plays a significant role in measuring large language models and guiding them to align with human values. Unfortunately, current comparing-based evaluation (CBE) methods typically focus on a single optimization objective, failing to effectively utilize scarce yet valuable preference signals. To address this, we delve into key factors that can enhance the accuracy, convergence, and scalability of CBE: suppressing sampling bias, balancing descending process of uncertainty, and mitigating updating uncertainty. Following the derived guidelines, we propose UniCBE, a unified uniformity-driven CBE framework which simultaneously optimize these core objectives by constructing and integrating three decoupled sampling probability matrices, each designed to ensure uniformity in specific aspects. We further ablate the optimal tuple sampling and preference aggregation strategies to achieve efficient CBE. On the AlpacaEval benchmark, UniCBE saves over 17% of evaluation budgets while achieving a Pearson correlation with ground truth exceeding 0.995, demonstrating excellent accuracy and convergence. In scenarios where new models are continuously introduced, UniCBE can even save over 50% of evaluation costs, highlighting its improved scalability.

Jiayi Shi, Yiwei Li, Shaoxiong Feng et al.

Large language models (LLMs) typically undergo instruction tuning to enhance alignment. Recent studies emphasize that quality and diversity of instruction data are more crucial than quantity, highlighting the need to select diverse, high-quality subsets to reduce training costs. However, how to evolve these selected subsets alongside the development of new instruction data remains insufficiently explored. To achieve LLMs' ongoing alignment, we introduce Instruction Bank (\textbf{InsBank}), a continuously updated repository that integrates the latest valuable instruction data. We further propose Progressive Instruction Bank Evolution (\textbf{PIBE}), a novel framework designed to evolve InsBank effectively and efficiently over time. PIBE employs a gradual data selection strategy to maintain long-term efficiency, leveraging a representation-based diversity score to capture relationships between data points and retain historical information for comprehensive diversity evaluation. This also allows for flexible combination of diversity and quality scores during data selection and ranking. Extensive experiments demonstrate that PIBE significantly outperforms baselines in InsBank evolution and is able to extract budget-specific subsets, demonstrating its effectiveness and adaptability.