Factual knowledge stored in Large Language Models (LLMs) inevitably becomes outdated or erroneous over time, making it critical to update these models without incurring the high cost of retraining. Existing sequential knowledge editing methods predominantly rely on strict orthogonal projection to preserve previously edited knowledge. However, this excessive constraint limits gradient expressiveness, resulting in a significant degradation of model generalization and overall performance as the number of edits increases. To address this challenge, we propose Dual-Importance Projection Editing (DipEdit). This method leverages Singular Value Decomposition (SVD) to identify critical gradient subspaces and introduces a dual mechanism comprising “accumulated importance” and “projection importance.” Unlike traditional approaches that enforce strict orthogonality, DipEdit dynamically scales gradient components parallel to key subspaces based on their projection importance rather than discarding them directly. This approach enhances the model’s adaptability to new knowledge while maximally preserving historical knowledge. Extensive experiments conducted on five mainstream LLMs using the ZsRE and Counterfact datasets demonstrate that DipEdit effectively handles thousands of sequential edits. The proposed method achieves an average comprehensive performance improvement of 10.36% and effectively maintains the model’s general capabilities on downstream tasks. Code is available at: https://github.com/czhhhla/DipEdit.

Algorithm Visualization (AV) helps students build mental models by animating algorithm execution states. Recent LLM-based systems such as CODE2VIDEO generate AV videos in an end-to-end manner. However, this paradigm requires the system to simultaneously simulate algorithm flow and satisfy video rendering constraints (element layout, color schemes, etc.), a complex task that induces LLM hallucinations. This results in reduced execution success rates, element overlap, and inter-frame inconsistencies.To address these challenges, we propose ALGOGEN, a novel paradigm that decouples algorithm execution from rendering. We first introduce Visualization Trace Algebra (VTA), a monoid over algorithm visual states and operations. The LLM then generates a Python tracker that simulates algorithm flow and outputs VTA-JSON traces, a JSON encoding of VTA. For rendering, we define a Rendering Style Language (RSL) to templatize algorithm layouts. A deterministic renderer then compiles algorithm traces with RSL into Manim, LaTeX/TikZ, or Three.js outputs[Manim, TikZ, and Three.js are respectively a Python animation engine, a LaTeX vector graphics package, and a JavaScript 3D rendering library.].Evaluated on a LeetCode AV benchmark of 200 tasks, ALGOGEN achieves an average success rate improvement of 17.3% compared to end-to-end methods (99.8% vs. 82.5%). These results demonstrate that our decoupling paradigm effectively mitigates LLM hallucinations in complex AV tasks, providing a more reliable solution for automated generation of high-quality algorithm visualizations. Demo videos and code are available at: .

Large language models deployed as multi-step agents are increasingly trained via Direct Preference Optimization (DPO) on trajectory-level comparisons. Yet standard DPO weights every step uniformly, penalizing correct early actions in a losing trajectory as severely as the causal error itself. We propose Divergence-Weighted DPO(dwDPO), a drop-in replacement that detects the first divergence step between competing trajectories using implicit rewards already computed by DPO, then concentrates the loss around it via an asymmetric kernel with a closed-form scale factor that preserves the Bradley-Terry preference model. dwDPO recovers standard DPO as a special case and provably achieves higher gradient concentration at the critical step than both uniform and monotonic-discount baselines. Extensive experimental results across six deep-research benchmarks demonstrate the superiority of dwDPO over state-of-the-art open-source baselines of comparable scale at no additional inference cost.

Long-context LLMs are constrained by the linear growth of key–value (KV) caches during autoregressive decoding, which incurs pronounced latency and memory overhead. KV eviction mitigates this issue, with existing efforts fall into offline policies with fixed eviction patterns and online policies that adaptively discard cache based on attention scores. While online eviction typically preserves accuracy under standard benchmarks, its performance can collapse in practical multi-turn dialogue scenarios where the query positions vary, and integration with pre-fill acceleration remains challenging. In contrast, offline eviction is infrastructure-friendly and generalizable but commonly sacrifices more accuracy. In this paper, we explore Supervised Fine-Tuning (SFT) for offline KV eviction and demonstrate its efficacy as a simple and powerful alternative to the design of complex online eviction metrics. We further propose Hybrid Sparse Supervised Fine-Tuning (HS-SFT) to explore the optimal offline design of KV eviction within SFT. In particular, HS-SFT employs a straight-through estimator to learn discrete local-window allocations of streaming heads across layers with budget-aware balancing loss, such that under high compression ratios—where dense-head capacity is constrained—the budget can be more effectively skewed to capture critical information. Across extensive evaluations on a wide array of LLMs and long-context tasks, HS-SFT delivers substantial performance gains over state-of-the-art eviction baselines. For example, with fewer than 4 hours of SFT of LLaMA-3-8B-1048K using a single 8-GPU node, HS-SFT achieves 5.86% and 38.3% higher average accuracy than Duo-attn on Longbench and Ruler-16K at 10% KV budget, respectively. These results position training-aware offline eviction—achieved with simple SFT—as an effective and practical path to scalable long-context inference.

Low-bit quantization is crucial for deploying Large Language Models (LLMs) on resource-constrained hardware. However, existing Post-Training Quantization (PTQ) methods adopt a monolithic view of outliers, overlooking their global spatial patterns within LLM architectures and weights that are quantization-sensitive but not numerically large. To overcome these challenges, we introduce MOAQ, a novel finetuning-free weight-only PTQ framework based on a new perspective for outlier analysis. MOAQ holistically addresses the outlier problem by three core components: (1) Spatial Calibration (SC) uses a clustering-optimized Hadamard matrix to reorganize disordered outliers into both discrete and clustered patterns. (2) Local Awareness and Suppression (LAS) adopts a dual strategy guided by our outlier metric. It assigns higher bit-widths to clustered outliers via Clustered Outlier Bit Search (COBS), enabling flexible fractional-bit quantization, while further refining quantization parameters for local discrete outliers through Local Outlierness-Aware Calibration (LOAC). (3) Global Feedback and Denoising (GFD) introduces a dynamic correction mechanism informed by our outlier metric to mitigate the progressive cross-layer enrichment of outliers. Extensive experiments show that MOAQ achieves state-of-the-art performance, outperforming both finetuning-free and many finetuning-based methods with substantially lower computational cost.

The Mixture of Experts (MoE) architecture enables efficient model scaling through conditional computation, where only subset of parameters are activated per input. However, this distributed architecture poses unprecedented challenges for model compression, as conventional quantization methods optimized for dense networks prove inadequate. This paper introduces a specialized quantization framework for MoE architectures, motivated by our discovery that weight matrices across expert networks exhibit distinctive channel-wise outlier distributions, necessitating a more nuanced compression approach. Through theoretical analysis incorporating Fisher Information matrices and condition number characteristics, we establish a fundamental relationship between layer functionality and quantization sensitivity, demonstrating that down-projection layers inherently demand higher precision compared to up-projection layers. Leveraging these insights, we develop an automated channel-wise quantization framework that dynamically determines optimal bit-width allocations while maintaining minimal computational overhead through efficient statistical approximations. When evaluated on the Mixtral-8x7b-v0.1 architecture, our methodology demonstrates a 3.96% improvement over existing state-of-the-art approaches across natural language understanding benchmarks, while achieving superior compression ratios.

Large Language Model (LLM)-based agents have excelled in various domains but face significant challenges when applied to data science workflows due to their complex, multi-stage nature. Current LLM-based agents struggle with non-linear relationships, recursive dependencies, implicit data- and logic-dependent reasoning, and managing extensive context. In this paper, we introduce Data Interpreter, an LLM-based agent that addresses these challenges through hierarchical graph-based modeling to represent the complexity and a progressive strategy for step-by-step verification, refinement, and consistent context management. Extensive experiments confirm the effectiveness of Data Interpreter. On InfiAgent-DABench, it boosts performance by 25% (from 75.9% to 94.9%), and on machine learning and open-ended tasks, it lifts accuracy from 88% to 95% and from 60% to 97%, respectively. Moreover, our method surpasses state-of-the-art baselines by 26% on the MATH dataset. We will release the code upon publication.

Large Language Models (LLMs) have demonstrated powerful performance in sequential recommendation due to their robust language modeling and comprehension capabilities. In such paradigms, the item texts of interaction sequences are formulated as sentences and LLMs are utilized to learn language representations or directly generate target item texts by incorporating instructions. Despite their promise, these methods solely focus on modeling the mapping from sequential texts to target items, neglecting the relationship between the items in an interaction sequence. This results in a failure to learn the transition patterns between items, which reflect the dynamic change in user preferences and are crucial for predicting the next item. To tackle this issue, we propose a novel framework for mapping the sequential item texts to the sequential item IDs, named ST2SI. Specifically, we first introduce multi-query input and item linear projection (ILP) to model the conditional probability distribution of items. Then, we further propose ID alignment to address misalignment between item texts and item IDs by instruction tuning. Finally, we propose efficient ILP tuning to adapt flexibly to different scenarios, requiring only training a linear layer to achieve competitive performance. Extensive experiments on six real-world datasets show our approach outperforms the best baselines by 7.33% in NDCG@10, 4.65% in Recall@10, and 8.42% in MRR.