Chain-of-thought (CoT) offers a potential boon for AI safety as it allows monitoring a model's CoT to try to understand its intentions and reasoning processes. However, the effectiveness of such monitoring hinges on CoTs faithfully representing models' actual reasoning processes. We evaluate CoT faithfulness of state-of-the-art reasoning models across 6 reasoning hints presented in the prompts and find: (1) for most settings and models tested, CoTs reveal their usage of hints in at least 1% of examples where they use the hint, but the reveal rate is often below 20%, (2) outcome-based reinforcement learning initially improves faithfulness but plateaus without saturating, and (3) when reinforcement learning increases how frequently hints are used (reward hacking), the propensity to verbalize them does not increase, even without training against a CoT monitor. These results suggest that CoT monitoring is a promising way of noticing undesired behaviors during training and evaluations, but that it is not sufficient to rule them out. They also suggest that in settings like ours where CoT reasoning is not necessary, test-time monitoring of CoTs is unlikely to reliably catch rare and catastrophic unexpected behaviors.

Reasoning capabilities of large language models are primarily studied for English, even when pretrained models are multilingual. In this work, we investigate to what extent English reasoning finetuning with long chain-of-thoughts (CoTs) can generalize across languages. First, we find that scaling up inference compute for English-centric reasoning language models (RLMs) improves multilingual mathematical reasoning across many languages including low-resource languages, to an extent where they outperform models twice their size. Second, we reveal that while English-centric RLM's CoTs are naturally predominantly English, they consistently follow a quote-and-think pattern to reason about quoted non-English inputs. Third, we discover an effective strategy to control the language of long CoT reasoning, and we observe that models reason better and more efficiently in high-resource languages. Finally, we observe poor out-of-domain reasoning generalization, in particular from STEM to cultural commonsense knowledge, even for English. Overall, we demonstrate the potentials, study the mechanisms and outline the limitations of crosslingual generalization of English reasoning test-time scaling. We conclude that practitioners should let English-centric RLMs reason in high-resource languages, while further work is needed to improve reasoning in low-resource languages and out-of-domain contexts.

--In conventional end-to-end frameworks for Autonomous driving (AD), the underlying cognitive processes are inadequately addressed. While large language models (LLMs) offer improved understanding and reasoning capabilities, integrating them into AD systems poses two major challenges. First, there is an invariably significant discrepancy between the high computational demands of LLMs and the high efficiency required for autonomous vehicles. Second, although LLMs can generate high-quality semantic reasoning, it remains an open challenge to map high-level textual reasoning to low-level trajectory planning for autonomous vehicles. T o deal with these issues, we present DSDrive, a streamlined end-to-end paradigm tailored for integrating the reasoning and planning of autonomous vehicles into a unified framework. DSDrive leverages a compact LLM that employs a distillation method to preserve the enhanced reasoning capabilities of a larger-sized vision language model (VLM). T o effectively align the reasoning and planning tasks, a waypoint-driven dual-head coordination module is further developed, which synchronizes dataset structures, optimization objectives, and the learning process. DSDrive has been thoroughly tested in closed-loop simulations, where it performs on par with benchmark models and even outperforms in many key metrics, all while being more compact in size. Additionally, the computational efficiency of DSDrive (as reflected in its time and memory requirements during inference) has been significantly enhanced. Evidently thus, this work brings promising aspects and underscores the potential of lightweight systems in delivering interpretable and efficient solutions for AD. Video of the experiments: https://youtu.be/op8PzQurugY Autonomous vehicles (A Vs) have demonstrated pervasive power in modern intelligent transportation systems, which significantly enhances on-road driving safety and travel efficiency [1]. The end-to-end (E2E) autonomous driving (AD) framework [2], leveraging large-scale driving datasets [3] and advanced learning methodologies [4], plays an important role in advancing the forefront of A Vs. This fully differentiable framework processes raw sensor inputs to output low-level control through joint task optimization. Wenru Liu, Pei Liu, and Jun Ma are with The Hong Kong University of Science and Technology, China (e-mail: wliu354@connect.hkust-gz.edu.cn;

Aspect sentiment triplet extraction (ASTE) aims to extract triplets composed of aspect terms, opinion terms, and sentiment polarities from given sentences. The table tagging method is a popular approach to addressing this task, which encodes a sentence into a 2-dimensional table, allowing for the tagging of relations between any two words. Previous efforts have focused on designing various downstream relation learning modules to better capture interactions between tokens in the table, revealing that a stronger capability to capture relations can lead to greater improvements in the model. Motivated by this, we attempt to directly utilize transformer layers as downstream relation learning modules. Due to the powerful semantic modeling capability of transformers, it is foreseeable that this will lead to excellent improvement. However, owing to the quadratic relation between the length of the table and the length of the input sentence sequence, using transformers directly faces two challenges: overly long table sequences and unfair local attention interaction. To address these challenges, we propose a novel Table-Transformer (T-T) for the tagging-based ASTE method. Specifically, we introduce a stripe attention mechanism with a loop-shift strategy to tackle these challenges. The former modifies the global attention mechanism to only attend to a 2-dimensional local attention window, while the latter facilitates interaction between different attention windows. Extensive and comprehensive experiments demonstrate that the T-T, as a downstream relation learning module, achieves state-of-the-art performance with lower computational costs.

Few-shot tabular learning, in which machine learning models are trained with a limited amount of labeled data, provides a cost-effective approach to addressing real-world challenges. The advent of Large Language Models (LLMs) has sparked interest in leveraging their pre-trained knowledge for few-shot tabular learning. Despite promising results, existing approaches either rely on test-time knowledge extraction, which introduces undesirable latency, or text-level knowledge, which leads to unreliable feature engineering. To overcome these limitations, we propose Latte, a training-time knowledge extraction framework that transfers the latent prior knowledge within LLMs to optimize a more generalized downstream model. Latte enables general knowledge-guided downstream tabular learning, facilitating the weighted fusion of information across different feature values while reducing the risk of overfitting to limited labeled data. Furthermore, Latte is compatible with existing unsupervised pre-training paradigms and effectively utilizes available unlabeled samples to overcome the performance limitations imposed by an extremely small labeled dataset.

Large Language Models (LLMs) assist in specialized tasks but struggle to align with evolving domain knowledge without costly fine-tuning. Domain knowledge consists of: Knowledge: Immutable facts (e.g., 'A stone is solid') and generally accepted principles (e.g., ethical standards); Refined Memory: Evolving insights shaped by business needs and real-world changes. However, a significant gap often exists between a domain expert's deep, nuanced understanding and the system's domain knowledge, which can hinder accurate information retrieval and application. Our Memory-Augmented Refinement of Knowledge (MARK) framework enables LLMs to continuously learn without retraining by leveraging structured refined memory, inspired by the Society of Mind. MARK operates through specialized agents, each serving a distinct role: Residual Refined Memory Agent: Stores and retrieves domain-specific insights to maintain context over time; User Question Refined Memory Agent: Captures user-provided facts, abbreviations, and terminology for better comprehension; LLM Response Refined Memory Agent: Extracts key elements from responses for refinement and personalization. These agents analyse stored refined memory, detect patterns, resolve contradictions, and improve response accuracy. Temporal factors like recency and frequency prioritize relevant information while discarding outdated insights. MARK enhances LLMs in multiple ways: Ground Truth Strategy: Reduces hallucinations by establishing a structured reference; Domain-Specific Adaptation: Essential for fields like healthcare, law, and manufacturing, where proprietary insights are absent from public datasets; Personalized AI Assistants: Improves virtual assistants by remembering user preferences, ensuring coherent responses over time.

Large Language Models (LLMs) have achieved remarkable success in code generation tasks, powering various applications like code completion, debugging, and programming assistance. However, existing benchmarks such as HumanEval, MBPP, and BigCodeBench primarily evaluate LLMs on English-only prompts, overlooking the real-world scenario where multilingual developers often use code-mixed language while interacting with LLMs. To address this gap, we introduce CodeMixBench, a novel benchmark designed to evaluate the robustness of LLMs on code generation from code-mixed prompts. Built upon BigCodeBench, CodeMixBench introduces controlled code-mixing (CMD) into the natural language parts of prompts across three language pairs: Hinglish (Hindi-English), Spanish-English, and Chinese Pinyin-English. We comprehensively evaluate a diverse set of open-source code generation models ranging from 1.5B to 15B parameters. Our results show that code-mixed prompts consistently degrade Pass@1 performance compared to their English-only counterparts, with performance drops increasing under higher CMD levels for smaller models. CodeMixBench provides a realistic evaluation framework for studying multilingual code generation and highlights new challenges and directions for building robust code generation models that generalize well across diverse linguistic settings.

The peer review process in major artificial intelligence (AI) conferences faces unprecedented challenges with the surge of paper submissions (exceeding 10,000 submissions per venue), accompanied by growing concerns over review quality and reviewer responsibility. This position paper argues for the need to transform the traditional one-way review system into a bi-directional feedback loop where authors evaluate review quality and reviewers earn formal accreditation, creating an accountability framework that promotes a sustainable, high-quality peer review system. The current review system can be viewed as an interaction between three parties: the authors, reviewers, and system (i.e., conference), where we posit that all three parties share responsibility for the current problems. However, issues with authors can only be addressed through policy enforcement and detection tools, and ethical concerns can only be corrected through self-reflection. As such, this paper focuses on reforming reviewer accountability with systematic rewards through two key mechanisms: (1) a two-stage bi-directional review system that allows authors to evaluate reviews while minimizing retaliatory behavior, (2)a systematic reviewer reward system that incentivizes quality reviewing. We ask for the community's strong interest in these problems and the reforms that are needed to enhance the peer review process.

Chain-of-thoughts (CoT) requires large language models (LLMs) to generate intermediate steps before reaching the final answer, and has been proven effective to help LLMs solve complex reasoning tasks. However, the inner mechanism of CoT still remains largely unclear. In this paper, we empirically study the role of CoT tokens in LLMs on two compositional tasks: multi-digit multiplication and dynamic programming. While CoT is essential for solving these problems, we find that preserving only tokens that store intermediate results would achieve comparable performance. Furthermore, we observe that storing intermediate results in an alternative latent form will not affect model performance. We also randomly intervene some values in CoT, and notice that subsequent CoT tokens and the final answer would change correspondingly. These findings suggest that CoT tokens may function like variables in computer programs but with potential drawbacks like unintended shortcuts and computational complexity limits between tokens. The code and data are available at https://github.com/solitaryzero/CoTs_are_Variables.

Recent advances in AI have catalyzed the adoption of intelligent educational tools, yet many semantic retrieval systems remain ill-suited to the unique linguistic and structural characteristics of academic content. This study presents two open-source embedding models fine-tuned for educational question answering, particularly in the context of course syllabi. A synthetic dataset of 3,197 sentence pairs, spanning synonymous terminology, paraphrased questions, and implicit-explicit mappings, was constructed through a combination of manual curation and large language model (LLM)-assisted generation. Two training strategies were evaluated: (1) a baseline model fine-tuned using MultipleNegativesRankingLoss (MNRL), and (2) a dual-loss model that combines MNRL with CosineSimilarityLoss to improve both semantic ranking and similarity calibration. Evaluations were conducted on 28 university course syllabi using a fixed set of natural language questions categorized into course, faculty, and teaching assistant information. Results demonstrate that both fine-tuned models outperform strong open-source baselines, including all-MiniLM-L6-v2 and multi-qa-MiniLM-L6-cos-v1, and that the dual-loss model narrows the performance gap with high-performing proprietary embeddings such as OpenAI's text-embedding-3 series. This work contributes reusable, domain-aligned embedding models and provides a replicable framework for educational semantic retrieval, supporting downstream applications such as academic chatbots, retrieval-augmented generation (RAG) systems, and learning management system (LMS) integrations.