While reasoning large language models (LLMs) demonstrate remarkable performance across various tasks, they also contain notable security vulnerabilities. Recent research has uncovered a "thinking-stopped" vulnerability in DeepSeek-R1, where model-generated reasoning tokens can forcibly interrupt the inference process, resulting in empty responses that compromise LLM-integrated applications. However, existing methods triggering this vulnerability require complex mathematical word problems with long prompts--even exceeding 5,000 tokens. To reduce the token cost and formally define this vulnerability, we propose a novel prompt injection attack named "Reasoning Interruption Attack", based on adaptive token compression. We demonstrate that simple standalone arithmetic tasks can effectively trigger this vulnerability, and the prompts based on such tasks exhibit simpler logical structures than mathematical word problems. We develop a systematic approach to efficiently collect attack prompts and an adaptive token compression framework that utilizes LLMs to automatically compress these prompts. Experiments show our compression framework significantly reduces prompt length while maintaining effective attack capabilities. We further investigate the attack's performance via output prefix and analyze the underlying causes of the vulnerability, providing valuable insights for improving security in reasoning LLMs.

Knowledge distillation typically transfers knowledge from a teacher model to a student model by minimizing differences between their output distributions. However, existing distillation approaches largely focus on mimicking absolute probabilities and neglect the valuable relational inductive biases embedded in the teacher's relative predictions, leading to exposure bias. In this paper, we propose Group Relative Knowledge Distillation (GRKD), a novel framework that distills teacher knowledge by learning the relative ranking among classes, rather than directly fitting the absolute distribution. Specifically, we introduce a group relative loss that encourages the student model to preserve the pairwise preference orderings provided by the teacher's outputs. Extensive experiments on classification benchmarks demonstrate that GRKD achieves superior generalization compared to existing methods, especially in tasks requiring fine-grained class differentiation. Our method provides a new perspective on exploiting teacher knowledge, focusing on relational structure rather than absolute likelihood.

In arbitrary-order language models, it is an open question how to sample tokens in parallel from the correct joint distribution. With discrete diffusion models, the more tokens they generate in parallel, the less their predicted distributions adhere to the originally learned data distribution, as they rely on a conditional independence assumption that only works with infinitesimally small timesteps. We find that a different class of models, any-subset autoregressive models (AS-ARMs), holds the solution. As implied by the name, AS-ARMs can generate tokens in any order, and in parallel. Moreover, AS-ARMs support parallelized joint probability density estimation, allowing them to correct their own parallel-generated token distributions, via our Any-Subset Speculative Decoding (ASSD) algorithm. ASSD provably enables generation of tokens from the correct joint distribution, with the number of neural network calls upper bounded by the number of tokens predicted. We empirically verify that ASSD speeds up language generation, without sacrificing quality. Furthermore, we provide a mathematically justified scheme for training AS-ARMs for generation, and show that AS-ARMs achieve state-of-the-art performance among sub-200M parameter models on infilling benchmark tasks, and nearly match the performance of models 50X larger on code generation. Our theoretical and empirical results indicate that the once-forgotten AS-ARMs are a promising direction of language modeling.

Personalized news recommendation systems often struggle to effectively capture the complexity of user preferences, as they rely heavily on shallow representations, such as article titles and abstracts. To address this problem, we introduce a novel method, namely PNR-LLM, for Large Language Models for Personalized News Recommendation. Specifically, PNR-LLM harnesses the generation capabilities of LLMs to enrich news titles and abstracts, and consequently improves recommendation quality. PNR-LLM contains a novel module, News Enrichment via LLMs, which generates deeper semantic information and relevant entities from articles, transforming shallow contents into richer representations. We further propose an attention mechanism to aggregate enriched semantic- and entity-level data, forming unified user and news embeddings that reveal a more accurate user-news match. Extensive experiments on MIND datasets show that PNR-LLM outperforms state-of-the-art baselines. Moreover, the proposed data enrichment module is model-agnostic, and we empirically show that applying our proposed module to multiple existing models can further improve their performance, verifying the advantage of our design.

Translating knowledge-intensive and entity-rich text between English and Korean requires transcreation to preserve language-specific and cultural nuances beyond literal, phonetic or word-for-word conversion. We evaluate 13 models (LLMs and MT models) using automatic metrics and human assessment by bilingual annotators. Our findings show LLMs outperform traditional MT systems but struggle with entity translation requiring cultural adaptation. By constructing an error taxonomy, we identify incorrect responses and entity name errors as key issues, with performance varying by entity type and popularity level. This work exposes gaps in automatic evaluation metrics and hope to enable future work in completing culturally-nuanced machine translation.

We study the primacy effect in three commercial LLMs: ChatGPT, Gemini and Claude. We do this by repurposing the famous experiment Asch (1946) conducted using human subjects. The experiment is simple, given two candidates with equal descriptions which one is preferred if one description has positive adjectives first before negative ones and another description has negative adjectives followed by positive ones. We test this in two experiments. In one experiment, LLMs are given both candidates simultaneously in the same prompt, and in another experiment, LLMs are given both candidates separately. We test all the models with 200 candidate pairs. We found that, in the first experiment, ChatGPT preferred the candidate with positive adjectives listed first, while Gemini preferred both equally often. Claude refused to make a choice. In the second experiment, ChatGPT and Claude were most likely to rank both candidates equally. In the case where they did not give an equal rating, both showed a clear preference to a candidate that had negative adjectives listed first. Gemini was most likely to prefer a candidate with negative adjectives listed first.

Large language models (LLMs) have revolutionized natural language understanding and generation but face significant memory bottlenecks during training. GaLore, Gradient Low-Rank Projection, addresses this issue by leveraging the inherent low-rank structure of weight gradients, enabling substantial memory savings without sacrificing performance. Recent works further extend GaLore from various aspects, including low-bit quantization and higher-order tensor structures. However, there are several remaining challenges for GaLore, such as the computational overhead of SVD for subspace updates and the integration with state-of-the-art training parallelization strategies (e.g., FSDP). In this paper, we present GaLore 2, an efficient and scalable GaLore framework that addresses these challenges and incorporates recent advancements. In addition, we demonstrate the scalability of GaLore 2 by pre-training Llama 7B from scratch using up to 500 billion training tokens, highlighting its potential impact on real LLM pre-training scenarios.

The advancement of reasoning capabilities in Large Language Models (LLMs) requires substantial amounts of high-quality reasoning data, particularly in mathematics. Existing data synthesis methods, such as data augmentation from annotated training sets or direct question generation based on relevant knowledge points and documents, have expanded datasets but face challenges in mastering the inner logic of the problem during generation and ensuring the verifiability of the solutions. To address these issues, we propose RV-Syn, a novel Rational and Verifiable mathematical Synthesis approach. RV-Syn constructs a structured mathematical operation function library based on initial seed problems and generates computational graphs as solutions by combining Python-formatted functions from this library. These graphs are then back-translated into complex problems. Based on the constructed computation graph, we achieve solution-guided logic-aware problem generation. Furthermore, the executability of the computational graph ensures the verifiability of the solving process. Experimental results show that RV-Syn surpasses existing synthesis methods, including those involving human-generated problems, achieving greater efficient data scaling. This approach provides a scalable framework for generating high-quality reasoning datasets.

Scripting interfaces enable users to automate tasks and customize software workflows, but creating scripts traditionally requires programming expertise and familiarity with specific APIs, posing barriers for many users. While Large Language Models (LLMs) can generate code from natural language queries, runtime code generation is severely limited due to unverified code, security risks, longer response times, and higher computational costs. To bridge the gap, we propose an offline simulation framework to curate a software-specific skillset, a collection of verified scripts, by exploiting LLMs and publicly available scripting guides. Our framework comprises two components: (1) task creation, using top-down functionality guidance and bottom-up API synergy exploration to generate helpful tasks; and (2) skill generation with trials, refining and validating scripts based on execution feedback. To efficiently navigate the extensive API landscape, we introduce a Graph Neural Network (GNN)-based link prediction model to capture API synergy, enabling the generation of skills involving underutilized APIs and expanding the skillset's diversity. Experiments with Adobe Illustrator demonstrate that our framework significantly improves automation success rates, reduces response time, and saves runtime token costs compared to traditional runtime code generation. This is the first attempt to use software scripting interfaces as a testbed for LLM-based systems, highlighting the advantages of leveraging execution feedback in a controlled environment and offering valuable insights into aligning AI capabilities with user needs in specialized software domains.

In recent years, the use of prompts to guide the output of Large Language Models have increased dramatically. However, even the best of experts struggle to choose the correct words to stitch up a prompt for the desired task. To solve this, LLM driven prompt optimization emerged as an important problem. Existing prompt optimization methods optimize a prompt globally, where in all the prompt tokens have to be optimized over a large vocabulary while solving a complex task. The large optimization space (tokens) leads to insufficient guidance for a better prompt. In this work, we introduce Local Prompt Optimization (LPO) that integrates with any general automatic prompt engineering method. We identify the optimization tokens in a prompt and nudge the LLM to focus only on those tokens in its optimization step. We observe remarkable performance improvements on Math Reasoning (GSM8k and MultiArith) and BIG-bench Hard benchmarks across various automatic prompt engineering methods. Further, we show that LPO converges to the optimal prompt faster than global methods.