Linear attention is an efficient alternative that maintains a constant memory footprint, even on infinite context lengths. While this is a potential candidate for lifelong learning, it falls short in memory capacity. In this paper, we propose LoLA, a training-free augmentation to linear attention that boosts associative recall. LoLA distributes past key-value pairs from context into three memory systems: (i) recent pairs in a local sliding window cache; (ii) difficult-to-memorize pairs in a sparse, global cache; and (iii) generic pairs in the recurrent hidden state of linear attention. We show through ablations that our self-recall error metric is crucial to efficiently manage long-term associative memories. On pass-key retrieval tasks, LoLA improves the base model's performance from 0.6% to 97.4% accuracy. This is achieved with a 4.6 smaller cache than Llama-3.1 8B on 4K context length. LoLA also outperforms other 1B and 8B parameter subquadratic models on zero-shot commonsense reasoning tasks. Transformer-based large language models (LLMs) rely on storing all past tokens in an ever-growing key-value (KV) cache (V aswani et al., 2017). This allows future query tokens to access past memories with associative recall, which enables in-context learning (Olsson et al., 2022). Since no previous information is discarded, the KV cache continues to grow with context length. This eventually leads to a memory bottleneck on long context tasks, such as lifelong in-context learning. Alternative architectures to transformers have been proposed--such as Mamba (Gu & Dao, 2024), DeltaNet (Schlag et al., 2021), linear attention (Katharopoulos et al., 2020), and others (Y ang et al., 2024a; Behrouz et al., 2024; Sun et al., 2024)--to reduce the compute complexity from quadratic to linear. Additionally, these approaches reduce the memory cost from linear to constant. In particular, linear attention removes the exponential dot product in softmax (Katharopoulos et al., 2020).

Large language models (LLMs) have improved significantly in their reasoning through extensive training on massive datasets. However, relying solely on additional data for improvement is becoming increasingly impractical, highlighting the need for models to autonomously enhance their reasoning without external supervision. In this paper, we propose Debate, Train, Evolve (DTE), a novel ground truth-free training framework that uses multi-agent debate traces to evolve a single language model. We also introduce a new prompting strategy Reflect-Critique-Refine, to improve debate quality by explicitly instructing agents to critique and refine their reasoning. Extensive evaluations on seven reasoning benchmarks with six open-weight models show that our DTE framework achieve substantial improvements, with an average accuracy gain of 8.92% on the challenging GSM-PLUS dataset. Furthermore, we observe strong cross-domain generalization, with an average accuracy gain of 5.8% on all other benchmarks, suggesting that our method captures general reasoning capabilities. Our framework code and trained models are publicly available at https://github.com/ctrl-gaurav/Debate-Train-Evolve

Large language models (LLMs) exhibit strong capabilities as decision-making agents by interleaving reasoning and actions, as seen in ReAct-style frameworks. Yet, their practical deployment is constrained by high inference costs and large model sizes. We propose Structured Agent Distillation, a framework that compresses large LLM-based agents into smaller student models while preserving both reasoning fidelity and action consistency. Unlike standard token-level distillation, our method segments trajectories into {[REASON]} and {[ACT]} spans, applying segment-specific losses to align each component with the teacher's behavior. This structure-aware supervision enables compact agents to better replicate the teacher's decision process. Experiments on ALFWorld, HotPotQA-ReAct, and WebShop show that our approach consistently outperforms token-level and imitation learning baselines, achieving significant compression with minimal performance drop. Scaling and ablation results further highlight the importance of span-level alignment for efficient and deployable agents.

Semi-supervised learning (SSL) has emerged as a practical solution for addressing data scarcity challenges by leveraging unlabeled data. Recently, vision-language models (VLMs), pre-trained on massive image-text pairs, have demonstrated remarkable zero-/few-shot performance that often surpasses SSL approaches due to their exceptional generalization capabilities. This gap motivates us to question: how can we effectively harness the powerful generalization capabilities of VLMs into task-specific models? Knowledge distillation (KD) offers a natural framework for transferring VLM capabilities, but we identify that it suffers from gradient conflicts between supervised and distillation losses. To address this challenge, we propose Dual-Head Optimization (DHO), which introduces dual prediction heads for each distinct signal. We observe that DHO resolves gradient conflicts, enabling improved feature learning compared to single-head KD baselines, with practical benefits of minimal computational overhead and test-time hyperparameter tuning without retraining. Extensive experiments across 15 datasets show that DHO consistently outperforms KD baselines, often outperforming teacher models with smaller student models. DHO also achieves new state-of-the-art performance on both in-distribution ImageNet semi-supervised learning and out-of-distribution generalization across ImageNet variants. We publicly release our code and model checkpoints to facilitate future research at https://github.com/erjui/DHO.

--High storage and computational costs obstruct deep neural networks to be deployed on resource-constrained devices. Knowledge distillation aims to train a compact student network by transferring knowledge from a larger pre-trained teacher model. However, most existing methods on knowledge distillation ignore the valuable information among training process associated with training results. In this paper, we provide a new Collaborative T eaching Knowledge Distillation (CTKD) strategy which employs two special teachers. Specifically, one teacher trained from scratch (i.e., scratch teacher) assists the student step by step using its temporary outputs. It forces the student to approach the optimal path towards the final logits with high accuracy. The other pre-trained teacher (i.e., expert teacher) guides the student to focus on a critical region which is more useful for the task. The combination of the knowledge from two special teachers can significantly improve the performance of the student network in knowledge distillation. The results of experiments on CIF AR-10, CIF AR-100, SVHN and Tiny ImageNet datasets verify that the proposed knowledge distillation method is efficient and achieves state-of-the-art performance. ECENTL Y, deep neural networks achieved superior performance in a variety of applications such as computer vision [1][2][3][4] and natural language processing [5][6]. However, along with high-performance, the deep neural network's architecture becomes much deeper and wider which requires a high cost of computation and memory in inference. It is a great burden to deploy these models on edge-computing systems such as embedded devices and mobile-phones. Therefore, many methods [7][8][9][10][11] are proposed to reduce the deep neural network's computational complexity and high storage. Some lightweight networks like Inception [12], MobileNet [13], ShuffleNet [14], SqueezeNet [15] and Condense-Net [16] have been proposed to reduce the network size as much as possible under the condition of keeping a high recognition accuracy. All the above mentioned methods focus on physically reducing internal redundancy of the model to obtain a shallow and thin architecture.

The rapid advancement of LLMs sparked significant interest in their potential to augment or automate managerial functions. One of the most recent trends in AI benchmarking is performance of Large Language Models (LLMs) over longer time horizons. While LLMs excel at tasks involving natural language and pattern recognition, their capabilities in multi-step, strategic business decision-making remain largely unexplored. Few studies demonstrated how results can be different from benchmarks in short-term tasks, as Vending-Bench revealed. Meanwhile, there is a shortage of alternative benchmarks for long-term coherence. This research analyses a novel benchmark using a business game for the decision making in business. The research contributes to the recent literature on AI by proposing a reproducible, open-access management simulator to the research community for LLM benchmarking. This novel framework is used for evaluating the performance of five leading LLMs available in free online interface: Gemini, ChatGPT, Meta AI, Mistral AI, and Grok. LLM makes decisions for a simulated retail company. A dynamic, month-by-month management simulation provides transparently in spreadsheet model as experimental environment. In each of twelve months, the LLMs are provided with a structured prompt containing a full business report from the previous period and are tasked with making key strategic decisions: pricing, order size, marketing budget, hiring, dismissal, loans, training expense, R&D expense, sales forecast, income forecast The methodology is designed to compare the LLMs on quantitative metrics: profit, revenue, and market share, and other KPIs. LLM decisions are analyzed in their strategic coherence, adaptability to market changes, and the rationale provided for their decisions. This approach allows to move beyond simple performance metrics for assessment of the long-term decision-making.

Large audio-language models are advancing rapidly, yet most evaluations emphasize speech or globally sourced sounds, overlooking culturally distinctive cues. This gap raises a critical question: can current models generalize to localized, non-semantic audio that communities instantly recognize but outsiders do not? To address this, we present TAU (Taiwan Audio Understanding), a benchmark of everyday Taiwanese "soundmarks." TAU is built through a pipeline combining curated sources, human editing, and LLM-assisted question generation, producing 702 clips and 1,794 multiple-choice items that cannot be solved by transcripts alone. Experiments show that state-of-the-art LALMs, including Gemini 2.5 and Qwen2-Audio, perform far below local humans. TAU demonstrates the need for localized benchmarks to reveal cultural blind spots, guide more equitable multimodal evaluation, and ensure models serve communities beyond the global mainstream.

Current density modeling approaches suffer from at least one of the following shortcomings: expensive training, slow inference, approximate likelihood, mode collapse or architectural constraints like bijective mappings. We propose a simple yet powerful framework that overcomes these limitations altogether. We define our model $q_θ(x)$ through a parametric distribution $q(x|w)$ with latent parameters $w$. Instead of directly optimizing the latent variables $w$, our idea is to marginalize them out by sampling $w$ from a learnable distribution $q_θ(w)$, hence the name Marginal Flow. In order to evaluate the learned density $q_θ(x)$ or to sample from it, we only need to draw samples from $q_θ(w)$, which makes both operations efficient. The proposed model allows for exact density evaluation and is orders of magnitude faster than competing models both at training and inference. Furthermore, Marginal Flow is a flexible framework: it does not impose any restrictions on the neural network architecture, it enables learning distributions on lower-dimensional manifolds (either known or to be learned), it can be trained efficiently with any objective (e.g. forward and reverse KL divergence), and it easily handles multi-modal targets. We evaluate Marginal Flow extensively on various tasks including synthetic datasets, simulation-based inference, distributions on positive definite matrices and manifold learning in latent spaces of images.

As the mathematical capabilities of large language models (LLMs) improve, it becomes increasingly important to evaluate their performance on research-level tasks at the frontier of mathematical knowledge. However, existing benchmarks are limited, as they focus solely on final-answer questions or high-school competition problems. To address this gap, we introduce IMProofBench, a private benchmark consisting of 39 peer-reviewed problems developed by expert mathematicians. Each problem requires a detailed proof and is paired with subproblems that have final answers, supporting both an evaluation of mathematical reasoning capabilities by human experts and a large-scale quantitative analysis through automated grading. Furthermore, unlike prior benchmarks, the evaluation setup simulates a realistic research environment: models operate in an agentic framework with tools like web search for literature review and mathematical software such as SageMath. Our results show that current LLMs can succeed at the more accessible research-level questions, but still encounter significant difficulties on more challenging problems. Quantitatively, Grok-4 achieves the highest accuracy of 52% on final-answer subproblems, while GPT-5 obtains the best performance for proof generation, achieving a fully correct solution for 22% of problems. IMProofBench will continue to evolve as a dynamic benchmark in collaboration with the mathematical community, ensuring its relevance for evaluating the next generation of LLMs.

The primary objective of Chinese grammatical error correction (CGEC) is to detect and correct errors in Chinese sentences. Recent research shows that large language models (LLMs) have been applied to CGEC with significant results. For LLMs, selecting appropriate reference examples can help improve their performance. However, existing methods predominantly rely on text similarity for example retrieval, a strategy that frequently mismatches actual error patterns and retrieves lexically similar yet grammatically irrelevant sentences. To address this problem, we propose a method named RE$^2$, which retrieves appropriate examples with explanations of grammatical errors. Instead of using text similarity of the input sentence, we use explanations of grammatical errors to select reference examples, which are used by LLMs to improve the performance of CGEC. We conduct experiments on two CGEC datasets and create a high-quality grammatical error explanation (GEE) dataset, which is not only used in our research but also serves as a valuable resource for future studies in both CGEC and GEE. The experimental results on the two datasets indicate that our proposed method effectively improves the performance of CGEC.