CoThink: Token-Efficient Reasoning via Instruct Models Guiding Reasoning Models

Large language models (LLMs) benefit from increased test-time compute, a phenomenon known as test-time scaling. However, reasoning-optimized models often overthink even simple problems, producing excessively verbose outputs and leading to low token efficiency. By comparing these models with equally sized instruct models, we identify two key causes of this verbosity: (1) reinforcement learning reduces the information density of forward reasoning, and (2) backward chain-of thought training encourages redundant and often unnecessary verification steps. Since LLMs cannot assess the difficulty of a given problem, they tend to apply the same cautious reasoning strategy across all tasks, resulting in inefficient overthinking. To address this, we propose CoThink, an embarrassingly simple pipeline: an instruct model first drafts a high-level solution outline; a reasoning model then works out the solution. We observe that CoThink enables dynamic adjustment of reasoning depth based on input difficulty. Evaluated with three reasoning models DAPO, DeepSeek-R1, and QwQ on three datasets GSM8K, MATH500, and AIME24, CoThink reduces total token generation by 22.3% while maintaining pass@1 accuracy within a 0.42% margin on average. With reference to the instruct model, we formally define reasoning efficiency and observe a potential reasoning efficiency scaling law in LLMs.

TTPA: Token-level Tool-use Preference Alignment Training Framework with Fine-grained Evaluation

Existing tool-learning methods usually rely on supervised fine-tuning, they often overlook fine-grained optimization of internal tool call details, leading to limitations in preference alignment and error discrimination. To overcome these challenges, we propose Token-level Tool-use Preference Alignment Training Framework (TTPA), a training paradigm for constructing token-level tool-use preference datasets that align LLMs with fine-grained preferences using a novel error-oriented scoring mechanism. TTPA first introduces reversed dataset construction, a method for creating high-quality, multi-turn tool-use datasets by reversing the generation flow. Additionally, we propose Token-level Preference Sampling (TPS) to capture fine-grained preferences by modeling token-level differences during generation. To address biases in scoring, we introduce the Error-oriented Scoring Mechanism (ESM), which quantifies tool-call errors and can be used as a training signal. Extensive experiments on three diverse benchmark datasets demonstrate that TTPA significantly improves tool-using performance while showing strong generalization ability across models and datasets.

CulFiT: A Fine-grained Cultural-aware LLM Training Paradigm via Multilingual Critique Data Synthesis

Large Language Models (LLMs) have demonstrated remarkable capabilities across various tasks, yet they often exhibit a specific cultural biases, neglecting the values and linguistic diversity of low-resource regions. This cultural bias not only undermines universal equality, but also risks reinforcing stereotypes and perpetuating discrimination. To address this, we propose CulFiT, a novel culturally-aware training paradigm that leverages multilingual data and fine-grained reward modeling to enhance cultural sensitivity and inclusivity. Our approach synthesizes diverse cultural-related questions, constructs critique data in culturally relevant languages, and employs fine-grained rewards to decompose cultural texts into verifiable knowledge units for interpretable evaluation. We also introduce GlobalCultureQA, a multilingual open-ended question-answering dataset designed to evaluate culturally-aware responses in a global context. Extensive experiments on three existing benchmarks and our GlobalCultureQA demonstrate that CulFiT achieves state-of-the-art open-source model performance in cultural alignment and general reasoning.

If an LLM Were a Character, Would It Know Its Own Story? Evaluating Lifelong Learning in LLMs

Large language models (LLMs) can carry out human-like dialogue, but unlike humans, they are stateless due to the superposition property. However, during multi-turn, multi-agent interactions, LLMs begin to exhibit consistent, character-like behaviors, hinting at a form of emergent lifelong learning. Despite this, existing benchmarks often fail to capture these dynamics, primarily focusing on static, open-ended evaluations. To address this gap, we introduce LIFESTATE-BENCH, a benchmark designed to assess lifelong learning in LLMs. It features two episodic datasets: Hamlet and a synthetic script collection, rich in narrative structure and character interactions. Our fact checking evaluation probes models' self-awareness, episodic memory retrieval, and relationship tracking, across both parametric and non-parametric approaches. Experiments on models like Llama3.1-8B, GPT-4-turbo, and DeepSeek R1, we demonstrate that nonparametric methods significantly outperform parametric ones in managing stateful learning. However, all models exhibit challenges with catastrophic forgetting as interactions extend, highlighting the need for further advancements in lifelong learning.

Position-Aware Depth Decay Decoding ($D^3$): Boosting Large Language Model Inference Efficiency

Due to the large number of parameters, the inference phase of Large Language Models (LLMs) is resource-intensive. Unlike traditional model compression, which needs retraining, recent dynamic computation methods show that not all components are required for inference, enabling a training-free pipeline. In this paper, we focus on the dynamic depth of LLM generation. A token-position aware layer skipping framework is proposed to save 1.5x times operations efficiently while maintaining performance. We first observed that tokens predicted later have lower perplexity and thus require less computation. Then, we propose a training-free algorithm called Position-Aware Depth Decay Decoding ($D^3$), which leverages a power-law decay function, $\left\lfloor L \times (\alpha^i) \right\rfloor$, to determine the number of layers to retain when generating token $T_i$. Remarkably, without any retraining, the $D^3$ achieves success across a wide range of generation tasks for the first time. Experiments on large language models (\ie the Llama) with $7 \sim 70$ billion parameters show that $D^3$ can achieve an average 1.5x speedup compared with the full-inference pipeline while maintaining comparable performance with nearly no performance drop ($<1\%$) on the GSM8K and BBH benchmarks.

More is not always better? Enhancing Many-Shot In-Context Learning with Differentiated and Reweighting Objectives

Large language models (LLMs) excel at few-shot in-context learning (ICL) without requiring parameter updates. However, as the number of ICL demonstrations increases from a few to many, performance tends to plateau and eventually decline. We identify two primary causes for this trend: the suboptimal negative log-likelihood (NLL) optimization objective and the incremental data noise. To address these issues, we introduce DR-ICL, a novel optimization method that enhances model performance through Differentiated Learning and advantage-based Reweighting objectives. Globally, DR-ICL utilizes differentiated learning to optimize the NLL objective, ensuring that many-shot performance surpasses zero-shot levels. Locally, it dynamically adjusts the weighting of many-shot demonstrations by leveraging cumulative advantages inspired by reinforcement learning, thereby improving generalization. This approach allows the model to handle varying numbers of shots effectively, mitigating the impact of noisy data. Recognizing the lack of multi-task datasets with diverse many-shot distributions, we develop the Many-Shot ICL Benchmark (MICLB)-a large-scale benchmark covering shot numbers from 1 to 350 within sequences of up to 8,000 tokens-for fine-tuning purposes. MICLB facilitates the evaluation of many-shot ICL strategies across seven prominent NLP tasks and 50 distinct datasets. Experimental results demonstrate that LLMs enhanced with DR-ICL achieve significant improvements in many-shot setups across various tasks, including both in-domain and out-of-domain scenarios. We release the code and benchmark dataset hoping to facilitate further research in many-shot ICL.

RED: Effective Trajectory Representation Learning with Comprehensive Information

Trajectory representation learning (TRL) maps trajectories to vectors that can then be used for various downstream tasks, including trajectory similarity computation, trajectory classification, and travel-time estimation. However, existing TRL methods often produce vectors that, when used in downstream tasks, yield insufficiently accurate results. A key reason is that they fail to utilize the comprehensive information encompassed by trajectories. We propose a self-supervised TRL framework, called RED, which effectively exploits multiple types of trajectory information. Overall, RED adopts the Transformer as the backbone model and masks the constituting paths in trajectories to train a masked autoencoder (MAE). In particular, RED considers the moving patterns of trajectories by employing a Road-aware masking strategy} that retains key paths of trajectories during masking, thereby preserving crucial information of the trajectories. RED also adopts a spatial-temporal-user joint Embedding scheme to encode comprehensive information when preparing the trajectories as model inputs. To conduct training, RED adopts Dual-objective task learning}: the Transformer encoder predicts the next segment in a trajectory, and the Transformer decoder reconstructs the entire trajectory. RED also considers the spatial-temporal correlations of trajectories by modifying the attention mechanism of the Transformer. We compare RED with 9 state-of-the-art TRL methods for 4 downstream tasks on 3 real-world datasets, finding that RED can usually improve the accuracy of the best-performing baseline by over 5%.

Grid and Road Expressions Are Complementary for Trajectory Representation Learning

Trajectory representation learning (TRL) maps trajectories to vectors that can be used for many downstream tasks. Existing TRL methods use either grid trajectories, capturing movement in free space, or road trajectories, capturing movement in a road network, as input. We observe that the two types of trajectories are complementary, providing either region and location information or providing road structure and movement regularity. Therefore, we propose a novel multimodal TRL method, dubbed GREEN, to jointly utilize Grid and Road trajectory Expressions for Effective representatioN learning. In particular, we transform raw GPS trajectories into both grid and road trajectories and tailor two encoders to capture their respective information. To align the two encoders such that they complement each other, we adopt a contrastive loss to encourage them to produce similar embeddings for the same raw trajectory and design a mask language model (MLM) loss to use grid trajectories to help reconstruct masked road trajectories. To learn the final trajectory representation, a dual-modal interactor is used to fuse the outputs of the two encoders via cross-attention. We compare GREEN with 7 state-of-the-art TRL methods for 3 downstream tasks, finding that GREEN consistently outperforms all baselines and improves the accuracy of the best-performing baseline by an average of 15.99\%.

CausalStock: Deep End-to-end Causal Discovery for News-driven Stock Movement Prediction

There are two issues in news-driven multi-stock movement prediction tasks that are not well solved in the existing works. On the one hand, "relation discovery" is a pivotal part when leveraging the price information of other stocks to achieve accurate stock movement prediction. Given that stock relations are often unidirectional, such as the "supplier-consumer" relationship, causal relations are more appropriate to capture the impact between stocks. On the other hand, there is substantial noise existing in the news data leading to extracting effective information with difficulty. With these two issues in mind, we propose a novel framework called CausalStock for news-driven multi-stock movement prediction, which discovers the temporal causal relations between stocks. We design a lag-dependent temporal causal discovery mechanism to model the temporal causal graph distribution. Then a Functional Causal Model is employed to encapsulate the discovered causal relations and predict the stock movements. Additionally, we propose a Denoised News Encoder by taking advantage of the excellent text evaluation ability of large language models (LLMs) to extract useful information from massive news data. The experiment results show that CausalStock outperforms the strong baselines for both news-driven multi-stock movement prediction and multi-stock movement prediction tasks on six real-world datasets collected from the US, China, Japan, and UK markets. Moreover, getting benefit from the causal relations, CausalStock could offer a clear prediction mechanism with good explainability.

What Affects the Stability of Tool Learning? An Empirical Study on the Robustness of Tool Learning Frameworks

Tool learning methods have enhanced the ability of large language models (LLMs) to interact with real-world applications. Many existing works fine-tune LLMs or design prompts to enable LLMs to select appropriate tools and correctly invoke them to meet user requirements. However, it is observed in previous works that the performance of tool learning varies from tasks, datasets, training settings, and algorithms. Without understanding the impact of these factors, it can lead to inconsistent results, inefficient model deployment, and suboptimal tool utilization, ultimately hindering the practical integration and scalability of LLMs in real-world scenarios. Therefore, in this paper, we explore the impact of both internal and external factors on the performance of tool learning frameworks. Through extensive experiments on two benchmark datasets, we find several insightful conclusions for future work, including the observation that LLMs can benefit significantly from increased trial and exploration. We believe our empirical study provides a new perspective for future tool learning research.