We thank all the reviewers for excellent questions and many relevant remarks. Thank you for this remark. One of the reason for this is that our method produces interpretations directly in terms of the input features. Thank you for pointing this out, we agree that faithful is not best. This is not the case for local models such as LIME.

A major drawback of reasoning models is their excessive token usage, inflating computational cost, resource demand, and latency. We show this verbosity stems not from deeper reasoning but from reinforcement learning loss minimization when models produce incorrect answers. With unsolvable problems dominating training, this effect compounds into a systematic tendency toward longer outputs. Through theoretical analysis of PPO and GRPO, we prove that incorrect answers inherently drive policies toward verbosity \textit{even when} $γ=1$, reframing response lengthening as an optimization artifact. We further uncover a consistent correlation between conciseness and correctness across reasoning and non-reasoning models. Building on these insights, we propose a two-phase RL procedure where a brief secondary stage, trained on a small set of solvable problems, significantly reduces response length while preserving or improving accuracy. Finally, we show that while GRPO shares properties with PPO, it exhibits collapse modes, limiting its reliability for concise reasoning. Our claims are supported by extensive experiments.

Large language models (LLMs) frequently generate responses that are lengthy and verbose, filled with redundant or unnecessary details. This diminishes clarity and user satisfaction, and it increases costs for model developers, especially with well-known proprietary models that charge based on the number of output tokens. In this paper, we introduce a novel reference-free metric for evaluating the conciseness of responses generated by LLMs. Our method quantifies non-essential content without relying on gold standard references and calculates the average of three calculations: i) a compression ratio between the original response and an LLM abstractive summary; ii) a compression ratio between the original response and an LLM extractive summary; and iii) wordremoval compression, where an LLM removes as many non-essential words as possible from the response while preserving its meaning, with the number of tokens removed indicating the conciseness score. Experimental results demonstrate that our proposed metric identifies redundancy in LLM outputs, offering a practical tool for automated evaluation of response brevity in conversational AI systems without the need for ground truth human annotations.

Reinforcement learning with verifiable rewards (RLVR) has been shown to enhance the reasoning capabilities of large language models (LLMs), enabling the development of large reasoning models (LRMs). However, LRMs such as DeepSeek-R1 and OpenAI o1 often generate verbose responses containing redundant or irrelevant reasoning step-a phenomenon known as overthinking-which substantially increases computational costs. Prior efforts to mitigate this issue commonly incorporate length penalties into the reward function, but we find they frequently suffer from two critical issues: length collapse and training collapse, resulting in sub-optimal performance. To address them, we propose a pipeline for training a Conciseness Reward Model (CRM) that scores the conciseness of reasoning path. Additionally, we introduce a novel reward formulation named Conciseness Reward Function (CRF) with explicit dependency between the outcome reward and conciseness score, thereby fostering both more effective and more efficient reasoning. From a theoretical standpoint, we demonstrate the superiority of the new reward from the perspective of variance reduction and improved convergence properties. Besides, on the practical side, extensive experiments on five mathematical benchmark datasets demonstrate the method's effectiveness and token efficiency, which achieves an 8.1% accuracy improvement and a 19.9% reduction in response token length on Qwen2.5-7B. Furthermore, the method generalizes well to other LLMs including Llama and Mistral. The implementation code and datasets are publicly available for reproduction: https://anonymous.4open.science/r/CRM.

The use of Large Language Models (LLMs) to support tasks in software development has steadily increased over recent years. From assisting developers in coding activities to providing conversational agents that answer newcomers' questions. In collaboration with the Mozilla Foundation, this study evaluates the effectiveness of Retrieval-Augmented Generation (RAG) in assisting developers within the Mozilla Firefox project. We conducted an empirical analysis comparing responses from human developers, a standard GPT model, and a GPT model enhanced with RAG, using real queries from Mozilla's developer chat rooms. To ensure a rigorous evaluation, Mozilla experts assessed the responses based on helpfulness, comprehensiveness, and conciseness. The results show that RAG-assisted responses were more comprehensive than human developers (62.50% to 54.17%) and almost as helpful (75.00% to 79.17%), suggesting RAG's potential to enhance developer assistance. However, the RAG responses were not as concise and often verbose. The results show the potential to apply RAG-based tools to Open Source Software (OSS) to minimize the load to core maintainers without losing answer quality. Toning down retrieval mechanisms and making responses even shorter in the future would enhance developer assistance in massive projects like Mozilla Firefox.

Generating editable, parametric CAD models from a single image holds great potential to lower the barriers of industrial concept design. However, current multi-modal large language models (MLLMs) still struggle with accurately inferring 3D geometry from 2D images due to limited spatial reasoning capabilities. We address this limitation by introducing GACO-CAD, a novel two-stage post-training framework. It is designed to achieve a joint objective: simultaneously improving the geometric accuracy of the generated CAD models and encouraging the use of more concise modeling procedures. First, during supervised fine-tuning, we leverage depth and surface normal maps as dense geometric priors, combining them with the RGB image to form a multi-channel input. In the context of single-view reconstruction, these priors provide complementary spatial cues that help the MLLM more reliably recover 3D geometry from 2D observations. Second, during reinforcement learning, we introduce a group length reward that, while preserving high geometric fidelity, promotes the generation of more compact and less redundant parametric modeling sequences. A simple dynamic weighting strategy is adopted to stabilize training. Experiments on the DeepCAD and Fusion360 datasets show that GACO-CAD achieves state-of-the-art performance under the same MLLM backbone, consistently outperforming existing methods in terms of code validity, geometric accuracy, and modeling conciseness.

Concise reasoning in large language models seeks to generate only essential intermediate steps needed to arrive at a final answer, thereby alleviating issues of "over-thinking". Most proposed approaches hinge on carefully hand-crafted heuristics, struggling to balance concision with performance, often failing to adapt across domains and model scales. In this work, we address these challenges by introducing a principled and pragmatic strategy, performance-aware length updating (P ALU). As a principled algorithm, P ALU formulates concise reasoning as a constrained optimization problem, minimizing response length subject to a performance constraint, and then applies Lagrangian optimization to convert it into a tractable unconstrained problem. As a pragmatic solution, P ALU streamlines complicated update rules through three approximations: (i) estimating performance with off-policy rollouts, (ii) truncating the Lagrange multiplier to two extremes, and (iii) replacing gradient-based updates with quantile-driven length adjustments. Furthermore, P ALU is demonstrated to adapt across both domain (logic, STEM and math) and model scale (1.5B, 7B, 14B) entrenching the algorithm as a practical and effective concise reasoning approach. Reasoning, requiring large language models (LLMs) to work through intermediate steps before producing a final answer, substantially improves performance on complex tasks such as mathematics (Jaech et al., 2024; Shao et al., 2024), programming (Lambert et al., 2024), and value alignment (Guo et al., 2025). Y et this benefit is often accompanied by overthinking: redundant self-reflection, backtracking, and validation (Chen et al., 2024; Zhang et al., 2024; Fatemi et al., 2025). These limitations inflate inference costs and hampers user experience, motivating the need for concise reasoning--the production of only the essential steps required to reach a correct answer.

Automatic summarization systems have advanced rapidly with large language models (LLMs), yet they still lack reliable guarantees on inclusion of critical content in high-stakes domains like healthcare, law, and finance. In this work, we introduce Conformal Importance Summarization, the first framework for importance-preserving summary generation which uses conformal prediction to provide rigorous, distribution-free coverage guarantees. By calibrating thresholds on sentence-level importance scores, we enable extractive document summarization with user-specified coverage and recall rates over critical content. Our method is model-agnostic, requires only a small calibration set, and seamlessly integrates with existing black-box LLMs. Experiments on established summarization benchmarks demonstrate that Conformal Importance Summarization achieves the theoretically assured information coverage rate. Our work suggests that Conformal Importance Summarization can be combined with existing techniques to achieve reliable, controllable automatic summarization, paving the way for safer deployment of AI summarization tools in critical applications. Code is available at https://github.com/layer6ai-labs/conformal-importance-summarization.

Large language models (LLMs) require careful alignment to balance competing objectives - factuality, safety, conciseness, proactivity, and diversity. Existing studies focus on individual techniques or specific dimensions, lacking a holistic assessment of the inherent trade-offs. We propose a unified evaluation framework that compares LLM alignment methods (PPO, DPO, ORPO, KTO) across these five axes, using both in-distribution and out-of-distribution datasets. Leveraging a specialized LLM-as-Judge prompt, validated through human studies, we reveal that DPO and KTO excel in factual accuracy, PPO and DPO lead in safety, and PPO best balances conciseness with proactivity. Our findings provide insights into trade-offs of common alignment methods, guiding the development of more balanced and reliable LLMs.