Xinglin Wang

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.

Xinglin Wang, Yiwei Li, Shaoxiong Feng et al.

Self-consistency (SC), leveraging multiple samples from LLMs, shows significant gains on various reasoning tasks but struggles with free-form generation due to the difficulty of aggregating answers. Its variants, UCS and USC, rely on sample selection or voting mechanisms to improve output quality. These methods, however, face limitations due to their inability to fully utilize the nuanced consensus knowledge present within multiple candidate samples, often resulting in suboptimal outputs. We propose Fine-Grained Self-Consistency (FSC) to addresses these limitations by extracting and integrating segment-level commonalities from candidate samples, enhancing the performance of LLMs both in open-ended and reasoning tasks. Based on this, we present two additional strategies: candidate filtering, which enhances overall quality by identifying highly similar candidate sets, and merging, which reduces input token requirements by combining similar samples. The effectiveness of FSC is demonstrated through extensive experiments on various tasks, including summarization, code generation, and mathematical reasoning, using GPT-3.5-turbo and GPT-4. The results indicate significant improvements over baseline methods, showcasing the potential of FSC to optimize output quality by effectively synthesizing fine-grained consensus knowledge from multiple samples.

Yiwei Li, Jiayi Shi, Shaoxiong Feng et al.

Instruction data is crucial for improving the capability of Large Language Models (LLMs) to align with human-level performance. Recent research LIMA demonstrates that alignment is essentially a process where the model adapts instructions' interaction style or format to solve various tasks, leveraging pre-trained knowledge and skills. Therefore, for instructional data, the most important aspect is the task it represents, rather than the specific semantics and knowledge information. The latent representations of instructions play roles for some instruction-related tasks like data selection and demonstrations retrieval. However, they are always derived from text embeddings, encompass overall semantic information that influences the representation of task categories. In this work, we introduce a new concept, instruction embedding, and construct Instruction Embedding Benchmark (IEB) for its training and evaluation. Then, we propose a baseline Prompt-based Instruction Embedding (PIE) method to make the representations more attention on tasks. The evaluation of PIE, alongside other embedding methods on IEB with two designed tasks, demonstrates its superior performance in accurately identifying task categories. Moreover, the application of instruction embeddings in four downstream tasks showcases its effectiveness and suitability for instruction-related tasks.

Xinglin Wang, Peiwen Yuan, Shaoxiong Feng et al.

Piaget's Theory of Cognitive Development (PTC) posits that the development of cognitive levels forms the foundation for human learning across various abilities. As Large Language Models (LLMs) have recently shown remarkable abilities across a wide variety of tasks, we are curious about the cognitive levels of current LLMs: to what extent they have developed and how this development has been achieved. To this end, we construct a benchmark CogLM (Cognitive Ability Evaluation for Language Model) based on PTC to assess the cognitive levels of LLMs. CogLM comprises 1,220 questions spanning 10 cognitive abilities crafted by more than 20 human experts, providing a comprehensive testbed for the cognitive levels of LLMs. Through extensive experiments across multiple mainstream LLMs with CogLM, we find that: (1) In our testing framework, advanced LLMs (such as GPT-4) have demonstrated human-like cognitive abilities, comparable to those of a 20-year-old human. (2) The parameter size and optimization objective are two key factors affecting the cognitive levels of LLMs. (3) The performance on downstream tasks is positively correlated with the level of cognitive abilities. These findings fill the gap in research on the cognitive abilities of LLMs, tracing the development of LLMs from a cognitive perspective and guiding the future direction of their evolution.

Peiwen Yuan, Shaoxiong Feng, Yiwei Li et al.

The guidance from capability evaluations has greatly propelled the progress of both human society and Artificial Intelligence. However, as LLMs evolve, it becomes challenging to construct evaluation benchmarks for them with accurate labels on hard tasks that approach the boundaries of human capabilities. To credibly conduct evaluation without accurate labels (denoted as poor-supervised evaluation), we propose the PoEM framework. We first prove that the capability of a model can be equivalently assessed by the consistency between it and certain reference model, when their prediction distributions are independent and the sample size is infinite. To alleviate the insufficiencies of the conditions in reality, we further introduce an algorithm that treats humans (when available) and the models under evaluation as reference models, alternately conducting model weights calibration and filtering during E-step and M-step. Comprehensive experiments across 3 types of tasks with 16 mainstream LLMs have shown that PoEM under poor supervision can achieve an average of 0.98 Pearson correlation coefficient with supervised evaluation results, demonstrating good effectiveness, efficiency and generalizability. More generally, PoEM has advanced the evaluation paradigm evolution from human-centric to human&model-centric by treating both of them as reference models, mitigating the limitations of human evaluation in the era of LLMs.

Peiwen Yuan, Shaoxiong Feng, Yiwei Li et al.

Significant progress has been made in automatic text evaluation with the introduction of large language models (LLMs) as evaluators. However, current sample-wise evaluation paradigm suffers from the following issues: (1) Sensitive to prompt design; (2) Poor resistance to noise; (3) Inferior ensemble performance with static reference. Inspired by the fact that humans treat both criterion definition and inter sample comparison as references for evaluation, we propose BatchEval, a paradigm that conducts batch-wise evaluation iteratively to alleviate the above problems. We explore variants under this paradigm and confirm the optimal settings are two stage procedure with heterogeneous batch composition strategy and decimal scoring format. Comprehensive experiments across 3 LLMs on 4 text evaluation tasks demonstrate that BatchEval outperforms state-of-the-art methods by 10.5% on Pearson correlations with only 64% API cost on average. Further analyses have been conducted to verify the robustness, generalization, and working mechanism of BatchEval.

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.

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.

Yiwei Li, Peiwen Yuan, Shaoxiong Feng et al.

Self-consistency (SC) has been a widely used decoding strategy for chain-of-thought reasoning. Despite bringing significant performance improvements across a variety of multi-step reasoning tasks, it is a high-cost method that requires multiple sampling with the preset size. In this paper, we propose a simple and scalable sampling process, \textbf{E}arly-Stopping \textbf{S}elf-\textbf{C}onsistency (ESC), to greatly reduce the cost of SC without sacrificing performance. On this basis, one control scheme for ESC is further derivated to dynamically choose the performance-cost balance for different tasks and models. To demonstrate ESC's effectiveness, we conducted extensive experiments on three popular categories of reasoning tasks: arithmetic, commonsense and symbolic reasoning over language models with varying scales. The empirical results show that ESC reduces the average number of sampling of chain-of-thought reasoning by a significant margin on six benchmarks, including MATH (-33.8%), GSM8K (-80.1%), StrategyQA (-76.8%), CommonsenseQA (-78.5%), Coin Flip (-84.2%) and Last Letters (-67.4%), while attaining comparable performances.