As a model-agnostic approach to long context modeling, multi-agent systems can process inputs longer than a large language model's context window without retraining or architectural modifications. However, their performance often heavily relies on hand-crafted multi-agent collaboration strategies and prompt engineering, which limit generalizability. In this work, we introduce a principled framework that formalizes the model-agnostic long context modeling problem as a compression problem, yielding an information-theoretic compression objective. Building on this framework, we propose Graph of Agents (GoA), which dynamically constructs an input-dependent collaboration structure that maximizes this objective. For Llama 3.1 8B and Qwen3 8B across six document question answering benchmarks, GoA improves the average $F_1$ score of retrieval-augmented generation by 5.7\% and a strong multi-agent baseline using a fixed collaboration structure by 16.35\%, respectively. Even with only a 2K context window, GoA surpasses the 128K context window Llama 3.1 8B on LongBench, showing a dramatic increase in effective context length. Our source code is available at https://github.com/tjoo512/graph-of-agents.

Model integrity of Large language models (LLMs) has become a pressing security concern with their massive online deployment. Prior Bit-Flip Attacks (BFAs)--a class of popular AI weight memory fault-injection techniques--can severely compromise Deep Neural Networks (DNNs): as few as tens of bit flips can degrade accuracy toward random guessing. Recent studies extend BFAs to LLMs and reveal that, despite the intuition of better robustness from modularity and redundancy, only a handful of adversarial bit flips can also cause LLMs' catastrophic accuracy degradation. However, existing BFA methods typically focus on either integer or floating-point models separately, limiting attack flexibility. Moreover, in floating-point models, random bit flips often cause perturbed parameters to extreme values (e.g., flipping in exponent bit), making it not stealthy and leading to numerical runtime error (e.g., invalid tensor values (NaN/Inf)). In this work, for the first time, we propose SBFA (Sneaky Bit-Flip Attack), which collapses LLM performance with only one single bit flip while keeping perturbed values within benign layer-wise weight distribution. It is achieved through iterative searching and ranking through our defined parameter sensitivity metric, ImpactScore, which combines gradient sensitivity and perturbation range constrained by the benign layer-wise weight distribution. A novel lightweight SKIP searching algorithm is also proposed to greatly reduce searching complexity, which leads to successful SBFA searching taking only tens of minutes for SOT A LLMs. Across Qwen, LLaMA, and Gemma models, with only one single bit flip, SBFA successfully degrades accuracy to below random levels on MMLU and SST -2 in both BF16 and INT8 data formats. Remarkably, flipping a single bit out of billions of parameters reveals a severe security concern of SOT A LLM models.

Large language models (LLMs) transcend passive generation and act as goal-directed agents by invoking external tools. Reinforcement learning (RL) offers a principled framework for optimizing these emergent tool-use policies, yet the prevailing paradigm relies exclusively on sparse outcome rewards and lacks consideration of the particularity of tool-use tasks, inflating policy-gradient variance and resulting in inefficient training. To better understand and address these challenges, we first establish a theoretical link between policy entropy and training stability of tool-use tasks, which reveals that structured, low-entropy tokens are primary determinants of rewards. Motivated by this insight, we propose \textbf{Res}haped \textbf{T}oken-level policy gradients (\textbf{ResT}) for tool-use tasks. ResT reshapes the policy gradient through entropy-informed token reweighting, progressively upweighting reasoning tokens as training proceeds. This entropy-aware scheme enables a smooth shift from structural correctness to semantic reasoning and stabilizes convergence in multi-turn tool-use tasks. Evaluation on BFCL and API-Bank shows that ResT achieves state-of-the-art results, outperforming prior methods by up to $8.76\%$. When fine-tuned on a 4B base LLM, ResT further surpasses GPT-4o by $4.11\%$ on single-turn tasks and $1.50\%$ on multi-turn base tasks.

In this paper, we introduce a combination of novel and exciting tasks: the solution and generation of linguistic puzzles. We focus on puzzles used in Linguistic Olympiads for high school students. We first extend the existing benchmark for the task of solving linguistic puzzles. We explore the use of Large Language Models (LLMs), including recent state-of-the-art models such as OpenAI's o1, for solving linguistic puzzles, analyzing their performance across various linguistic topics. We demonstrate that LLMs outperform humans on most puzzles types, except for those centered on writing systems, and for the understudied languages. We use the insights from puzzle-solving experiments to direct the novel task of puzzle generation. We believe that automating puzzle generation, even for relatively simple puzzles, holds promise for expanding interest in linguistics and introducing the field to a broader audience. This finding highlights the importance of linguistic puzzle generation as a research task: such puzzles can not only promote linguistics but also support the dissemination of knowledge about rare and understudied languages.

Accurately forecasting chaotic systems, prevalent in domains such as weather prediction and fluid dynamics, remains a significant scientific challenge. The inherent sensitivity of these systems to initial conditions, coupled with a scarcity of observational data, severely constrains traditional modeling approaches. Since these models are typically trained for a specific system, they lack the generalization capacity necessary for real-world applications, which demand robust zero-shot or few-shot forecasting on novel or data-limited scenarios. To overcome this generalization barrier, we propose ChaosNexus, a foundation model pre-trained on a diverse corpus of chaotic dynamics. ChaosNexus employs a novel multi-scale architecture named ScaleFormer augmented with Mixture-of-Experts layers, to capture both universal patterns and system-specific behaviors. The model demonstrates state-of-the-art zero-shot generalization across both synthetic and real-world benchmarks. On a large-scale testbed comprising over 9,000 synthetic chaotic systems, it improves the fidelity of long-term attractor statistics by more than 40% compared to the leading baseline. This robust performance extends to real-world applications with exceptional data efficiency. For instance, in 5-day global weather forecasting, ChaosNexus achieves a competitive zero-shot mean error below 1 degree, a result that further improves with few-shot fine-tuning. Moreover, experiments on the scaling behavior of ChaosNexus provide a guiding principle for scientific foundation models: cross-system generalization stems from the diversity of training systems, rather than sheer data volume.

Simultaneous Machine Translation (SiMT) requires high-quality translations under strict real-time constraints, which traditional encoder-decoder policies with only READ/WRITE actions cannot fully address. We extend the action space of SiMT with four adaptive actions: SENTENCE_CUT, DROP, PAR-TIAL_SUMMARIZATION and PRONOMINALIZATION, which enable real-time restructuring, omission, and simplification while preserving semantic fidelity. We implement these actions in a decoder-only large language model (LLM) framework and construct training references through action-aware prompting. To evaluate both quality and latency, we further develop a latency-aware TTS pipeline that maps textual outputs to speech with realistic timing. Experiments on the ACL60/60 English-Chinese and English-German benchmarks show that our framework consistently improves semantic metrics (e.g., COMET-KIWI) and achieves lower delay (measured by Average Lagging) compared to reference translations and salami-based baselines. Notably, combining DROP and SEN-TENCE_CUT yields the best overall balance between fluency and latency. These results demonstrate that enriching the action space of LLM-based SiMT provides a promising direction for bridging the gap between human and machine interpretation.

Group relative policy optimization (GRPO) has demonstrated significant potential in improving the reasoning capabilities of large language models (LLMs) via reinforcement learning. However, its practical deployment is impeded by an excessively slow training process, primarily attributed to the computationally intensive autoregressive generation of multiple responses per query, which makes the generation phase the primary performance bottleneck. Although speculative decoding presents a promising direction for acceleration, its direct application in GRPO achieves limited speedup under high-concurrency training conditions. To overcome this limitation, we propose a concurrency-aware speculative decoding framework that dynamically adjusts the drafting and verification strategy according to real-time concurrency levels, thereby maximizing the acceleration of the generation process. Furthermore, to address performance degradation arising from distributional drift between the evolving target model and the fixed draft model during training, we introduce an online draft learning mechanism that enables the draft model to continuously adapt using feedback signals from the target model. Experimental results across multiple mathematical reasoning datasets and models demonstrate that the proposed method achieves end-to-end speedups of 2.35x to 2.72x, significantly surpassing baseline approaches in efficiency. The code is available at https://github.com/yedaotian9/GRPO Group relative policy optimization (GRPO) has recently emerged as a promising framework for enhancing the reasoning capabilities of large language models (LLMs) through reinforcement learning Team (2025a). In each training iteration, the LLM generates a group of responses to a given query. These responses are subsequently evaluated using a predefined rule-based reward function, and the resulting rewards are standardized prior to model updates via policy optimization Shao et al. (2024).

The dominant retrieve-then-rank pipeline in large-scale recommender systems suffers from mis-calibration and engineering overhead due to its architectural split and differing optimization objectives. While recent generative sequence models have shown promise in unifying retrieval and ranking by auto-regressively generating ranked items, existing solutions typically address either personalized search or query-free recommendation, often exhibiting performance trade-offs when attempting to unify both. We introduce SynerGen, a novel generative recommender model that bridges this critical gap by providing a single generative backbone for both personalized search and recommendation, while simultaneously excelling at retrieval and ranking tasks. Trained on behavioral sequences, our decoder-only Transformer leverages joint optimization with InfoNCE for retrieval and a hybrid pointwise-pairwise loss for ranking, allowing semantic signals from search to improve recommendation and vice versa. We also propose a novel time-aware rotary positional embedding to effectively incorporate time information into the attention mechanism. SynerGen achieves significant improvements on widely adopted recommendation and search benchmarks compared to strong generative recom-mender and joint search and recommendation baselines. This work demonstrates the viability of a single generative foundation model for industrial-scale unified information access. Large-scale search and recommendation systems in e-commerce, short video, and food-delivery platforms are typically deployed as multi-stage cascades.

Recent Large Audio-Language Models (LALMs) have shown strong performance on various audio understanding tasks such as speech translation and Audio Q\&A. However, they exhibit significant limitations on challenging audio reasoning tasks in complex acoustic scenarios. These situations would greatly benefit from the use of acoustic tools like noise suppression, source separation, and precise temporal alignment, but current LALMs lack access to such tools. To address this limitation, we introduce Thinking-with-Sound (TwS), a framework that equips LALMs with Audio CoT by combining linguistic reasoning with on-the-fly audio-domain analysis. Unlike existing approaches that treat audio as static input, TwS enables models to actively think with audio signals, performing numerical analysis and digital manipulation through multimodal reasoning. To evaluate this approach, we construct MELD-Hard1k, a new robustness benchmark created by introducing various acoustic perturbations. Experiments reveal that state-of-the-art LALMs suffer dramatic performance degradation on MELD-Hard1k, with accuracy dropping by more than $50\%$ compared to clean audio. TwS achieves substantial improvements in robustness, demonstrating both effectiveness and scalability: small models gain $24.73\%$ absolute accuracy, with improvements scaling consistently up to $36.61\%$ for larger models. Our findings demonstrate that Audio CoT can significantly enhance robustness without retraining, opening new directions for developing more robust audio understanding systems.

Large reasoning models improve accuracy by producing long reasoning traces, but this inflates latency and cost, motivating inference-time efficiency. We propose Retrieval-of-Thought (RoT), which reuses prior reasoning as composable "thought" steps to guide new problems. RoT organizes steps into a thought graph with sequential and semantic edges to enable fast retrieval and flexible recombination. At inference, RoT retrieves query-relevant nodes and applies reward-guided traversal to assemble a problem-specific template that guides generation. This dynamic template reuse reduces redundant exploration and, therefore, reduces output tokens while preserving accuracy. We evaluate RoT on reasoning benchmarks with multiple models, measuring accuracy, token usage, latency, and memory overhead. Findings show small prompt growth but substantial efficiency gains, with RoT reducing output tokens by up to 40%, inference latency by 82%, and cost by 59% while maintaining accuracy. RoT establishes a scalable paradigm for efficient LRM reasoning via dynamic template construction through retrieval. Large Reasoning Models (LRMs) have demonstrated impressive capabilities in solving complex tasks by producing outputs accompanied by detailed reasoning trajectories (Xu et al., 2025a). These models adopt an intentionally slower and more deliberative inference process, mimicking human-like reasoning. This approach typically involves generating longer outputs and consuming increased inference-time compute to effectively address reasoning-intensive queries. Recent efforts to improve reasoning in LLMs have primarily focused on generating more output tokens to simulate thoughtful, multi-step reasoning (Snell et al., 2024). A common approach involves guiding generation using external reward models Zhang et al. (2024). These include outcome-based reward models, such as Best-of-N (BoN) sampling.