Recent progress in autonomous code generation has fueled excitement around AI agents capable of accelerating scientific discovery by running experiments. However, there is currently no benchmark that evaluates whether such agents can implement scientific ideas when given varied amounts of code as a starting point, interpolating between reproduction (running code) and from-scratch replication (fully re-implementing and running code). We introduce AutoExperiment, a benchmark that evaluates AI agents' ability to implement and run machine learning experiments based on natural language descriptions in research papers. In each task, agents are given a research paper, a codebase with key functions masked out, and a command to run the experiment. The goal is to generate the missing code, execute the experiment in a sandboxed environment, and reproduce the results. AutoExperiment scales in difficulty by varying the number of missing functions $n$, ranging from partial reproduction to full replication. We evaluate state-of-the-art agents and find that performance degrades rapidly as $n$ increases. Agents that can dynamically interact with the environment (e.g. to debug their code) can outperform agents in fixed "agentless" harnesses, and there exists a significant gap between single-shot and multi-trial success rates (Pass@1 vs. Pass@5), motivating verifier approaches to our benchmark. Our findings highlight critical challenges in long-horizon code generation, context retrieval, and autonomous experiment execution, establishing AutoExperiment as a new benchmark for evaluating progress in AI-driven scientific experimentation. Our data and code are open-sourced at https://github.com/j1mk1m/AutoExperiment .
Large language models (LLMs) have demonstrated impressive capabilities in aiding developers with tasks like code comprehension, generation, and translation. Supporting multilingual programming -- i.e., coding tasks across multiple programming languages -- typically requires either (1) finetuning a single LLM across all programming languages, which is cost-efficient but sacrifices language-specific specialization and performance, or (2) finetuning separate LLMs for each programming language, which allows for specialization but is computationally expensive and storage-intensive due to the duplication of parameters. This paper introduces MoLE (Mix-of-Language-Experts), a novel architecture that balances efficiency and specialization for multilingual programming. MoLE is composed of a base model, a shared LoRA (low-rank adaptation) module, and a collection of language-specific LoRA modules. These modules are jointly optimized during the finetuning process, enabling effective knowledge sharing and specialization across programming languages. During inference, MoLE automatically routes to the language-specific LoRA module corresponding to the programming language of the code token being generated. Our experiments demonstrate that MoLE achieves greater parameter efficiency compared to training separate language-specific LoRAs, while outperforming a single shared LLM finetuned for all programming languages in terms of accuracy.
Large Language Models (LLMs) are increasingly used in applications requiring long context lengths, but the key-value (KV) cache often becomes a memory bottleneck on GPUs as context grows. To address this, we propose Commutative Vector Quantization (CommVQ) to significantly reduce memory usage for long-context LLM inference. We first introduce additive quantization with a lightweight encoder and codebook to compress the KV cache, which can be decoded via simple matrix multiplication. To further reduce computational costs during decoding, we design the codebook to be commutative with Rotary Position Embedding (RoPE) and train it using an Expectation-Maximization (EM) algorithm. This enables efficient integration of decoding into the self-attention mechanism. Our approach achieves high accuracy with additive quantization and low overhead via the RoPE-commutative codebook. Experiments on long-context benchmarks and GSM8K show that our method reduces FP16 KV cache size by 87.5% with 2-bit quantization, while outperforming state-of-the-art KV cache quantization methods. Notably, it enables 1-bit KV cache quantization with minimal accuracy loss, allowing a LLaMA-3.1 8B model to run with a 128K context length on a single RTX 4090 GPU. The source code is available at: https://github.com/UMass-Embodied-AGI/CommVQ.
Neural implicit scene representations have recently shown promising results in dense visual SLAM. However, existing implicit SLAM algorithms are constrained to single-agent scenarios, and fall difficulties in large-scale scenes and long sequences. Existing NeRF-based multi-agent SLAM frameworks cannot meet the constraints of communication bandwidth. To this end, we propose the first distributed multi-agent collaborative neural SLAM framework with hybrid scene representation, distributed camera tracking, intra-to-inter loop closure, and online distillation for multiple submap fusion. A novel triplane-grid joint scene representation method is proposed to improve scene reconstruction. A novel intra-to-inter loop closure method is designed to achieve local (single-agent) and global (multi-agent) consistency. We also design a novel online distillation method to fuse the information of different submaps to achieve global consistency. Furthermore, to the best of our knowledge, there is no real-world dataset for NeRF-based/GS-based SLAM that provides both continuous-time trajectories groundtruth and high-accuracy 3D meshes groundtruth. To this end, we propose the first real-world Dense slam (DES) dataset covering both single-agent and multi-agent scenarios, ranging from small rooms to large-scale outdoor scenes, with high-accuracy ground truth for both 3D mesh and continuous-time camera trajectory. This dataset can advance the development of the research in both SLAM, 3D reconstruction, and visual foundation model. Experiments on various datasets demonstrate the superiority of the proposed method in both mapping, tracking, and communication. The dataset and code will open-source on https://github.com/dtc111111/mcnslam.
Language models handle increasingly long contexts for tasks such as book summarization, but this leads to growing memory costs for the key-value (KV) cache. Many prior works have proposed ways of discarding KVs from memory, but their approaches are tailored to favorable settings, obscuring caveats like high peak memory and performance degradation, and a fair comparison between methods is difficult. In this paper, we propose the *KV footprint* as a unified metric, which accounts for both the amount of KV entries stored and their lifespan in memory. We evaluate methods based on the smallest footprint they attain while preserving performance in both long-context understanding and generation, with context lengths of up to 128K tokens. This metric reveals the high peak memory of prior KV eviction methods. One class of methods -- *post-fill eviction* -- has a high footprint due to being incompatible with eviction during pre-filling. We adapt these methods to be able to evict KVs during pre-filling, achieving substantially lower KV footprints. We then turn to *recency eviction* methods, wherein we propose PruLong, an end-to-end optimization method for learning which attention heads need to retain the full KV cache and which do not. PruLong saves memory while preserving long-context performance, achieving 12% smaller KV footprint than prior methods while retaining performance in challenging recall tasks. Our paper clarifies the complex tangle of long-context inference methods and paves the way for future development to minimize the KV footprint.
Foundation models have revolutionized music information retrieval, but questions remain about their ability to generalize across diverse musical traditions. This paper presents a comprehensive evaluation of five state-of-the-art audio foundation models across six musical corpora spanning Western popular, Greek, Turkish, and Indian classical traditions. We employ three complementary methodologies to investigate these models' cross-cultural capabilities: probing to assess inherent representations, targeted supervised fine-tuning of 1-2 layers, and multi-label few-shot learning for low-resource scenarios. Our analysis shows varying cross-cultural generalization, with larger models typically outperforming on non-Western music, though results decline for culturally distant traditions. Notably, our approaches achieve state-of-the-art performance on five out of six evaluated datasets, demonstrating the effectiveness of foundation models for world music understanding. We also find that our targeted fine-tuning approach does not consistently outperform probing across all settings, suggesting foundation models already encode substantial musical knowledge. Our evaluation framework and benchmarking results contribute to understanding how far current models are from achieving universal music representations while establishing metrics for future progress.
LaTeX's precision and flexibility in typesetting have made it the gold standard for the preparation of scientific documentation. Large Language Models (LLMs) present a promising opportunity for researchers to produce publication-ready material using LaTeX with natural language instructions, yet current benchmarks completely lack evaluation of this ability. By introducing TeXpert, our benchmark dataset with natural language prompts for generating LaTeX code focused on components of scientific documents across multiple difficulty levels, we conduct an in-depth analysis of LLM performance in this regard and identify frequent error types. Our evaluation across open and closed-source LLMs highlights multiple key findings: LLMs excelling on standard benchmarks perform poorly in LaTeX generation with a significant accuracy drop-off as the complexity of tasks increases; open-source models like DeepSeek v3 and DeepSeek Coder strongly rival closed-source counterparts in LaTeX tasks; and formatting and package errors are unexpectedly prevalent, suggesting a lack of diverse LaTeX examples in the training datasets of most LLMs. Our dataset, code, and model evaluations are available at https://github.com/knowledge-verse-ai/TeXpert.
Long story generation remains a challenge for existing large language models (LLMs), primarily due to two main factors: (1) discourse coherence, which requires plot consistency, logical coherence, and completeness in the long-form generation, and (2) narrative complexity, which requires an interwoven and engaging narrative. To address these challenges, we propose StoryWriter, a multi-agent story generation framework, which consists of three main modules: (1) outline agent, which generates event-based outlines containing rich event plots, character, and event-event relationships. (2) planning agent, which further details events and plans which events should be written in each chapter to maintain an interwoven and engaging story. (3) writing agent, which dynamically compresses the story history based on the current event to generate and reflect new plots, ensuring the coherence of the generated story. We conduct both human and automated evaluation, and StoryWriter significantly outperforms existing story generation baselines in both story quality and length. Furthermore, we use StoryWriter to generate a dataset, which contains about $6,000$ high-quality long stories, with an average length of $8,000$ words. We train the model Llama3.1-8B and GLM4-9B using supervised fine-tuning on LongStory and develop StoryWriter_GLM and StoryWriter_GLM, which demonstrates advanced performance in long story generation.
Recent advancements in large language models (LLMs) have demonstrated significant progress in math and code reasoning capabilities. However, existing code benchmark are limited in their ability to evaluate the full spectrum of these capabilities, particularly at the competitive level. To bridge this gap, we introduce OJBench, a novel and challenging benchmark designed to assess the competitive-level code reasoning abilities of LLMs. OJBench comprises 232 programming competition problems from NOI and ICPC, providing a more rigorous test of models' reasoning skills. We conducted a comprehensive evaluation using OJBench on 37 models, including both closed-source and open-source models, reasoning-oriented and non-reasoning-oriented models. Our results indicate that even state-of-the-art reasoning-oriented models, such as o4-mini and Gemini-2.5-pro-exp, struggle with highly challenging competition-level problems. This highlights the significant challenges that models face in competitive-level code reasoning.
Retrieval-Augmented Generation (RAG) has become essential for large-scale code generation, grounding predictions in external code corpora to improve actuality. However, a critical yet underexplored aspect of RAG pipelines is chunking -- the process of dividing documents into retrievable units. Existing line-based chunking heuristics often break semantic structures, splitting functions or merging unrelated code, which can degrade generation quality. We propose chunking via Abstract Syntax Trees (\ourwork), a structure-aware method that recursively breaks large AST nodes into smaller chunks and merges sibling nodes while respecting size limits. This approach generates self-contained, semantically coherent units across programming languages and tasks, improving performance on diverse code generation tasks, e.g., boosting Recall@5 by 4.3 points on RepoEval retrieval and Pass@1 by 2.67 points on SWE-bench generation. Our work highlights the importance of structure-aware chunking for scaling retrieval-enhanced code intelligence.
Heuristic algorithms play a vital role in solving combinatorial optimization (CO) problems, yet traditional designs depend heavily on manual expertise and struggle to generalize across diverse instances. We introduce \textbf{HeurAgenix}, a two-stage hyper-heuristic framework powered by large language models (LLMs) that first evolves heuristics and then selects among them automatically. In the heuristic evolution phase, HeurAgenix leverages an LLM to compare seed heuristic solutions with higher-quality solutions and extract reusable evolution strategies. During problem solving, it dynamically picks the most promising heuristic for each problem state, guided by the LLM's perception ability. For flexibility, this selector can be either a state-of-the-art LLM or a fine-tuned lightweight model with lower inference cost. To mitigate the scarcity of reliable supervision caused by CO complexity, we fine-tune the lightweight heuristic selector with a dual-reward mechanism that jointly exploits singals from selection preferences and state perception, enabling robust selection under noisy annotations. Extensive experiments on canonical benchmarks show that HeurAgenix not only outperforms existing LLM-based hyper-heuristics but also matches or exceeds specialized solvers. Code is available at https://github.com/microsoft/HeurAgenix.
Dense matching methods like DUSt3R regress pairwise pointmaps for 3D reconstruction. However, the reliance on pairwise prediction and the limited generalization capability inherently restrict the global geometric consistency. In this work, we introduce Test3R, a surprisingly simple test-time learning technique that significantly boosts geometric accuracy. Using image triplets ($I_1,I_2,I_3$), Test3R generates reconstructions from pairs ($I_1,I_2$) and ($I_1,I_3$). The core idea is to optimize the network at test time via a self-supervised objective: maximizing the geometric consistency between these two reconstructions relative to the common image $I_1$. This ensures the model produces cross-pair consistent outputs, regardless of the inputs. Extensive experiments demonstrate that our technique significantly outperforms previous state-of-the-art methods on the 3D reconstruction and multi-view depth estimation tasks. Moreover, it is universally applicable and nearly cost-free, making it easily applied to other models and implemented with minimal test-time training overhead and parameter footprint. Code is available at https://github.com/nopQAQ/Test3R.
Large Language Models (LLMs) are central to many contemporary AI applications, yet their extensive parameter counts pose significant challenges for deployment in memory- and compute-constrained environments. Recent works in eXplainable AI (XAI), particularly on attribution methods, suggest that interpretability can also enable model compression by identifying and removing components irrelevant to inference. In this paper, we leverage Layer-wise Relevance Propagation (LRP) to perform attribution-guided pruning of LLMs. While LRP has shown promise in structured pruning for vision models, we extend it to unstructured pruning in LLMs and demonstrate that it can substantially reduce model size with minimal performance loss. Our method is especially effective in extracting task-relevant subgraphs -- so-called ``circuits'' -- which can represent core functions (e.g., indirect object identification). Building on this, we introduce a technique for model correction, by selectively removing circuits responsible for spurious behaviors (e.g., toxic outputs). All in all, we gather these techniques as a uniform holistic framework and showcase its effectiveness and limitations through extensive experiments for compression, circuit discovery and model correction on Llama and OPT models, highlighting its potential for improving both model efficiency and safety. Our code is publicly available at https://github.com/erfanhatefi/SparC3.
In this work, we develop new optimization algorithms that use approximate second-order information combined with the gradient regularization technique to achieve fast global convergence rates for both convex and non-convex objectives. The key innovation of our analysis is a novel notion called Gradient-Normalized Smoothness, which characterizes the maximum radius of a ball around the current point that yields a good relative approximation of the gradient field. Our theory establishes a natural intrinsic connection between Hessian approximation and the linearization of the gradient. Importantly, Gradient-Normalized Smoothness does not depend on the specific problem class of the objective functions, while effectively translating local information about the gradient field and Hessian approximation into the global behavior of the method. This new concept equips approximate second-order algorithms with universal global convergence guarantees, recovering state-of-the-art rates for functions with H\"older-continuous Hessians and third derivatives, quasi-self-concordant functions, as well as smooth classes in first-order optimization. These rates are achieved automatically and extend to broader classes, such as generalized self-concordant functions. We demonstrate direct applications of our results for global linear rates in logistic regression and softmax problems with approximate Hessians, as well as in non-convex optimization using Fisher and Gauss-Newton approximations.
Single-trajectory reinforcement learning (RL) methods aim to optimize policies from datasets consisting of (prompt, response, reward) triplets, where scalar rewards are directly available. This supervision format is highly practical, as it mirrors real-world human feedback, such as thumbs-up/down signals, and avoids the need for structured preference annotations. In contrast, pairwise preference-based methods like Direct Preference Optimization (DPO) rely on datasets with both preferred and dispreferred responses, which are harder to construct and less natural to collect. Among single-trajectory approaches, Direct Reward Optimization (DRO) has shown strong empirical performance due to its simplicity and stability. However, DRO requires approximating a value function, which introduces several limitations: high off-policy variance, coupling between policy and value learning, and a lack of absolute supervision on the policy itself. We introduce Reward Partitioning Optimization (RPO), a new method that resolves these limitations by removing the need to model the value function. Instead, RPO normalizes observed rewards using a partitioning approach estimated directly from data. This leads to a straightforward supervised learning objective on the policy, with no auxiliary models and no joint optimization. RPO provides direct and stable supervision on the policy, making it robust and easy to implement in practice. We validate RPO on scalar-feedback language modeling tasks using Flan-T5 encoder-decoder models. Our results demonstrate that RPO outperforms existing single-trajectory baselines such as DRO and Kahneman-Tversky Optimization (KTO). These findings confirm that RPO is a simple, effective, and theoretically grounded method for single-trajectory policy optimization.
Inference and prediction under partial knowledge of a physical system is challenging, particularly when multiple confounding sources influence the measured response. Explicitly accounting for these influences in physics-based models is often infeasible due to epistemic uncertainty, cost, or time constraints, resulting in models that fail to accurately describe the behavior of the system. On the other hand, data-driven machine learning models such as variational autoencoders are not guaranteed to identify a parsimonious representation. As a result, they can suffer from poor generalization performance and reconstruction accuracy in the regime of limited and noisy data. We propose a physics-informed variational autoencoder architecture that combines the interpretability of physics-based models with the flexibility of data-driven models. To promote disentanglement of the known physics and confounding influences, the latent space is partitioned into physically meaningful variables that parametrize a physics-based model, and data-driven variables that capture variability in the domain and class of the physical system. The encoder is coupled with a decoder that integrates physics-based and data-driven components, and constrained by an adversarial training objective that prevents the data-driven components from overriding the known physics, ensuring that the physics-grounded latent variables remain interpretable. We demonstrate that the model is able to disentangle features of the input signal and separate the known physics from confounding influences using supervision in the form of class and domain observables. The model is evaluated on a series of synthetic case studies relevant to engineering structures, demonstrating the feasibility of the proposed approach.
Model merging aims to integrate the strengths of multiple fine-tuned models into a unified model while preserving task-specific capabilities. Existing methods, represented by task arithmetic, are typically classified into global- and local-aware methods. However, global-aware methods inevitably cause parameter interference, while local-aware methods struggle to maintain the effectiveness of task-specific details in the merged model. To address these limitations, we propose a Consensus-Aware Localized Merging (CALM) method which incorporates localized information aligned with global task consensus, ensuring its effectiveness post-merging. CALM consists of three key components: (1) class-balanced entropy minimization sampling, providing a more flexible and reliable way to leverage unsupervised data; (2) an efficient-aware framework, selecting a small set of tasks for sequential merging with high scalability; (3) a consensus-aware mask optimization, aligning localized binary masks with global task consensus and merging them conflict-free. Experiments demonstrate the superiority and robustness of our CALM, significantly outperforming existing methods and achieving performance close to traditional MTL.
Large language models (LLMs) have shown strong performance across natural language tasks, but remain vulnerable to backdoor attacks. Recent model editing-based approaches enable efficient backdoor injection by directly modifying parameters to map specific triggers to attacker-desired responses. However, these methods often suffer from safety fallback, where the model initially responds affirmatively but later reverts to refusals due to safety alignment. In this work, we propose DualEdit, a dual-objective model editing framework that jointly promotes affirmative outputs and suppresses refusal responses. To address two key challenges -- balancing the trade-off between affirmative promotion and refusal suppression, and handling the diversity of refusal expressions -- DualEdit introduces two complementary techniques. (1) Dynamic loss weighting calibrates the objective scale based on the pre-edited model to stabilize optimization. (2) Refusal value anchoring compresses the suppression target space by clustering representative refusal value vectors, reducing optimization conflict from overly diverse token sets. Experiments on safety-aligned LLMs show that DualEdit improves attack success by 9.98\% and reduces safety fallback rate by 10.88\% over baselines.
World models are critical for autonomous driving to simulate environmental dynamics and generate synthetic data. Existing methods struggle to disentangle ego-vehicle motion (perspective shifts) from scene evolvement (agent interactions), leading to suboptimal predictions. Instead, we propose to separate environmental changes from ego-motion by leveraging the scene-centric coordinate systems. In this paper, we introduce COME: a framework that integrates scene-centric forecasting Control into the Occupancy world ModEl. Specifically, COME first generates ego-irrelevant, spatially consistent future features through a scene-centric prediction branch, which are then converted into scene condition using a tailored ControlNet. These condition features are subsequently injected into the occupancy world model, enabling more accurate and controllable future occupancy predictions. Experimental results on the nuScenes-Occ3D dataset show that COME achieves consistent and significant improvements over state-of-the-art (SOTA) methods across diverse configurations, including different input sources (ground-truth, camera-based, fusion-based occupancy) and prediction horizons (3s and 8s). For example, under the same settings, COME achieves 26.3% better mIoU metric than DOME and 23.7% better mIoU metric than UniScene. These results highlight the efficacy of disentangled representation learning in enhancing spatio-temporal prediction fidelity for world models. Code and videos will be available at https://github.com/synsin0/COME.
Cross-domain time series imputation is an underexplored data-centric research task that presents significant challenges, particularly when the target domain suffers from high missing rates and domain shifts in temporal dynamics. Existing time series imputation approaches primarily focus on the single-domain setting, which cannot effectively adapt to a new domain with domain shifts. Meanwhile, conventional domain adaptation techniques struggle with data incompleteness, as they typically assume the data from both source and target domains are fully observed to enable adaptation. For the problem of cross-domain time series imputation, missing values introduce high uncertainty that hinders distribution alignment, making existing adaptation strategies ineffective. Specifically, our proposed solution tackles this problem from three perspectives: (i) Data: We introduce a frequency-based time series interpolation strategy that integrates shared spectral components from both domains while retaining domain-specific temporal structures, constructing informative priors for imputation. (ii) Model: We design a diffusion-based imputation model that effectively learns domain-shared representations and captures domain-specific temporal dependencies with dedicated denoising networks. (iii) Algorithm: We further propose a cross-domain consistency alignment strategy that selectively regularizes output-level domain discrepancies, enabling effective knowledge transfer while preserving domain-specific characteristics. Extensive experiments on three real-world datasets demonstrate the superiority of our proposed approach. Our code implementation is available here.