High-quality long-context data is essential for training large language models (LLMs) capable of processing extensive documents, yet existing synthesis approaches using relevance-based aggregation face challenges of computational efficiency. We present LiteLong, a resource-efficient method for synthesizing long-context data through structured topic organization and multi-agent debate. Our approach leverages the BISAC book classification system to provide a comprehensive hierarchical topic organization, and then employs a debate mechanism with multiple LLMs to generate diverse, high-quality topics within this structure. For each topic, we use lightweight BM25 retrieval to obtain relevant documents and concatenate them into 128K-token training samples. Experiments on HELMET and Ruler benchmarks demonstrate that LiteLong achieves competitive long-context performance and can seamlessly integrate with other long-dependency enhancement methods. LiteLong makes high-quality long-context data synthesis more accessible by reducing both computational and data engineering costs, facilitating further research in long-context language training.

We use language to communicate our thoughts. But is language merely the expression of thoughts, which are themselves produced by other, nonlinguistic parts of our minds? Or does language play a more transformative role in human cognition, allowing us to have thoughts that we otherwise could (or would) not have? Recent developments in artificial intelligence (AI) and cognitive science have reinvigorated this old question. We argue that language may hold the key to the emergence of both more general AI systems and central aspects of human intelligence. We highlight two related properties of language that make it such a powerful tool for developing domain--general abilities. First, language offers compact representations that make it easier to represent and reason about many abstract concepts (e.g., exact numerosity). Second, these compressed representations are the iterated output of collective minds. In learning a language, we learn a treasure trove of culturally evolved abstractions. Taken together, these properties mean that a sufficiently powerful learning system exposed to language--whether biological or artificial--learns a compressed model of the world, reverse engineering many of the conceptual and causal structures that support human (and human-like) thought.

Reinforcement Learning from Verifiable Rewards (RLVR) has recently shown that large language models (LLMs) can develop their own reasoning without direct supervision. However, applications in the medical domain, specifically for question answering, are susceptible to significant reward hacking during the reasoning phase. Our work addresses two primary forms of this behavior: i) providing a final answer without preceding reasoning, and ii) employing non-standard reasoning formats to exploit the reward mechanism. To mitigate these, we introduce a composite reward function with specific penalties for these behaviors. Our experiments show that extending RLVR with our proposed reward model leads to better-formatted reasoning with less reward hacking and good accuracy compared to the baselines. This approach marks a step toward reducing reward hacking and enhancing the reliability of models utilizing RLVR.

Large language models (LLMs) have become integral to a wide range of applications worldwide, driving an unprecedented global demand for effective multilingual capabilities. Central to achieving robust multilingual performance is the strategic allocation of language proportions within training corpora. However, determining optimal language ratios is highly challenging due to intricate cross-lingual interactions and sensitivity to dataset scale. This paper introduces Climb (Cross-Lingual Interaction-aware Multilingual Balancing), a novel framework designed to systematically optimize multilingual data allocation. At its core, Climb introduces a cross-lingual interaction-aware language ratio, explicitly quantifying each language's effective allocation by capturing inter-language dependencies. Leveraging this ratio, Climb proposes a principled two-step optimization procedure--first equalizing marginal benefits across languages, then maximizing the magnitude of the resulting language allocation vectors--significantly simplifying the inherently complex multilingual optimization problem. Extensive experiments confirm that Climb can accurately measure cross-lingual interactions across various multilingual settings. LLMs trained with Climb-derived proportions consistently achieve state-of-the-art multilingual performance, even achieving competitive performance with open-sourced LLMs trained with more tokens.

Multilingual Instruction Fine-Tuning (IFT) is essential for enabling large language models (LLMs) to generalize effectively across diverse linguistic and cultural contexts. However, the scarcity of high-quality multilingual training data and corresponding building method remains a critical bottleneck. While data selection has shown promise in English settings, existing methods often fail to generalize across languages due to reliance on simplistic heuristics or language-specific assumptions. In this work, we introduce Multilingual Data Quality and Diversity (M-DaQ), a novel method for improving LLMs multilinguality, by selecting high-quality and semantically diverse multilingual IFT samples. We further conduct the first systematic investigation of the Superficial Alignment Hypothesis (SAH) in multilingual setting. Empirical results across 18 languages demonstrate that models fine-tuned with M-DaQ method achieve significant performance gains over vanilla baselines over 60% win rate. Human evaluations further validate these gains, highlighting the increment of cultural points in the response. We release the M-DaQ code to support future research.

Highly capable AI systems could secretly pursue misaligned goals -- what we call "scheming". Because a scheming AI would deliberately try to hide its misaligned goals and actions, measuring and mitigating scheming requires different strategies than are typically used in ML. We propose that assessing anti-scheming interventions requires at least (1) testing propensity to scheme on far out-of-distribution (OOD) tasks, (2) evaluating whether lack of scheming is driven by situational awareness, and (3) checking for robustness to pre-existing misaligned goals. We use a broad category of "covert actions" -- such as secretly breaking rules or intentionally underperforming in tests -- as a proxy for scheming, and design evaluations for covert actions. We then stress-test deliberative alignment as a case study for anti-scheming. Across 26 OOD evaluations (180+ environments), deliberative alignment reduces covert action rates (OpenAI o3: 13%->0.4%) but does not fully eliminate them. Our mitigation is also able to largely stop agents from pursuing a hidden goal previously trained into the model, but we still find misbehavior after additional red-teaming. We find that models' chain-of-thought (CoT) often demonstrates awareness of being evaluated for alignment, and show causal evidence that this awareness decreases covert behavior, while unawareness increases it. Therefore, we cannot exclude that the observed reductions in covert action rates are at least partially driven by situational awareness. While we rely on human-legible CoT for training, studying situational awareness, and demonstrating clear evidence of misalignment, our ability to rely on this degrades as models continue to depart from reasoning in standard English. We encourage research into alignment mitigations for scheming and their assessment, especially for the adversarial case of deceptive alignment, which this paper does not address.

Slang is a commonly used type of informal language that poses a daunting challenge to NLP systems. Recent advances in large language models (LLMs), however, have made the problem more approachable. While LLM agents are becoming more widely applied to intermediary tasks such as slang detection and slang interpretation, their generalizability and reliability are heavily dependent on whether these models have captured structural knowledge about slang that align well with human attested slang usages. To answer this question, we contribute a systematic comparison between human and machine-generated slang usages. Our evaluative framework focuses on three core aspects: 1) Characteristics of the usages that reflect systematic biases in how machines perceive slang, 2) Creativity reflected by both lexical coinages and word reuses employed by the slang usages, and 3) Informativeness of the slang usages when used as gold-standard examples for model distillation. By comparing human-attested slang usages from the Online Slang Dictionary (OSD) and slang generated by GPT-4o and Llama-3, we find significant biases in how LLMs perceive slang. Our results suggest that while LLMs have captured significant knowledge about the creative aspects of slang, such knowledge does not align with humans sufficiently to enable LLMs for extrapolative tasks such as linguistic analyses.

This paper revisits the LLM cache bandit problem, with a special focus on addressing the query heterogeneity for cost-effective LLM inference. Previous works often assume uniform query sizes. Heterogeneous query sizes introduce a combinatorial structure for cache selection, making the cache replacement process more computationally and statistically challenging. We treat optimal cache selection as a knapsack problem and employ an accumulation-based strategy to effectively balance computational overhead and cache updates. In theoretical analysis, we prove that the regret of our algorithm achieves an $O(\sqrt{MNT})$ bound, improving the coefficient of $\sqrt{MN}$ compared to the $O(MN\sqrt{T})$ result in Berkeley, where $N$ is the total number of queries and $M$ is the cache size. Additionally, we also provide a problem-dependent bound, which was absent in previous works. The experiment rely on real-world data show that our algorithm reduces the total cost by approximately 12\%.

Large Language Models have spread rapidly since the release of ChatGPT in late 2022, accompanied by claims of major productivity gains but also concerns about job displacement. This paper examines the short-run labor market effects of LLM adoption by comparing earnings and unemployment across occupations with differing levels of exposure to these technologies. Using a Synthetic Difference in Differences approach, we estimate the impact of LLM exposure on earnings and unemployment. Our findings show that workers in highly exposed occupations experienced earnings increases following ChatGPT's introduction, while unemployment rates remained unchanged. These results suggest that initial labor market adjustments to LLMs operate primarily through earnings rather than worker reallocation.

Sarcasm detection remains a challenge in natural language understanding, as sarcastic intent often relies on subtle cross-modal cues spanning text, speech, and vision. While prior work has primarily focused on textual or visual-textual sarcasm, comprehensive audio-visual-textual sarcasm understanding remains underexplored. In this paper, we systematically evaluate large language models (LLMs) and multimodal LLMs for sarcasm detection on English (MUStARD++) and Chinese (MCSD 1.0) in zero-shot, few-shot, and LoRA fine-tuning settings. In addition to direct classification, we explore models as feature encoders, integrating their representations through a collaborative gating fusion module. Experimental results show that audio-based models achieve the strongest unimodal performance, while text-audio and audio-vision combinations outperform unimodal and trimodal models. Furthermore, MLLMs such as Qwen-Omni show competitive zero-shot and fine-tuned performance. Our findings highlight the potential of MLLMs for cross-lingual, audio-visual-textual sarcasm understanding.