NeurIPS 2025 论文列表

We present Ambient Protein Diffusion, a framework for training protein diffusion models that generates structures with unprecedented diversity and quality. State-of-the-art generative models are trained on computationally derived structures from AlphaFold2 (AF), as experimentally determined structures are relatively scarce. The resulting models are therefore limited by the quality of synthetic datasets. Since the accuracy of AF predictions degrades with increasing protein length and complexity, de novo generation of long, complex proteins remains challenging. Ambient Protein Diffusion overcomes this problem by treating low-confidence AF structures as corrupted data. Rather than simply filtering out low-quality AF structures, our method adjusts the diffusion objective for each structure based on its corruption level, allowing the model to learn from both high and low quality structures. Empirically, ambient protein diffusion yields major improvements: on proteins with 700 residues, diversity increases from 45% to 85% from the previous state-of-the-art, and designability improves from 70% to 88%.

Transformers have achieved state-of-the-art performance across various tasks, but suffer from a notable quadratic complexity in sequence length due to the attention mechanism. In this work, we propose MonarchAttention -- a novel approach to sub-quadratic attention approximation via Monarch matrices, an expressive class of structured matrices. Based on the variational form of softmax, we describe an efficient optimization-based algorithm to compute an approximate projection of softmax attention onto the class of Monarch matrices with $\Theta(N\sqrt{N} d)$ computational complexity and $\Theta(Nd)$ memory/IO complexity. Unlike previous approaches, MonarchAttention is both (1) transferable, yielding minimal performance loss with no additional training, even when replacing every attention layer of the transformer, and (2) hardware-efficient, utilizing the highest-throughput tensor core units on modern GPUs. With optimized kernels, MonarchAttention achieves substantial speed-ups in wall-time over FlashAttention-2: $1.4\times$ for shorter sequences $(N=256)$, $4.5\times$ for medium-length sequences $(N=4K)$, and $8.2\times$ for longer sequences $(N=16K)$. We demonstrate the quality of MonarchAttention on diverse tasks and architectures in vision and language problems, showing that it flexibly and accurately approximates softmax attention in a variety of contexts.

The quadratic complexity of self‑attention prevents transformers from scaling effectively to long input sequences. On the other hand, modern GPUs and other specialized hardware accelerators are well-optimized for processing small input sequences in transformers during both training and inference. A natural question arises: can we take advantage of the efficiency of small transformers to deal with long input sequences? In this paper, we show that transformers with long input sequences (large transformers) can be efficiently simulated by transformers that can only take short input sequences (small transformers). Specifically, we prove that any transformer with input length $N$ can be efficiently simulated by only $O((N/M)^2)$ transformers with input length $M \ll N$, and that this cannot be improved in the worst case. However, we then prove that in various natural scenarios including average-case inputs, sliding window masking and attention sinks, the optimal number $O(N/M)$ of small transformers suffice.

We introduce \textit{FFN Fusion}, an architectural optimization technique that reduces sequential computation in large language models by identifying and exploiting natural opportunities for parallelization. Our key insight is that sequences of Feed-Forward Network (FFN) layers, particularly those remaining after the removal of specific attention layers, can often be parallelized with minimal accuracy impact. We develop a principled methodology for identifying and fusing such sequences, transforming them into parallel operations that significantly reduce inference latency while preserving model behavior. Applying these techniques to Llama-3.1-405B-Instruct, we create a 253B model (253B-Base), an efficient and soon-to-be publicly available model that achieves a 1.71$\times$ speedup in inference latency and 35$\times$ lower per-token cost while maintaining strong performance across benchmarks. Most intriguingly, we find that even full transformer blocks containing both attention and FFN layers can sometimes be parallelized, suggesting new directions for neural architecture design.

The right batch size is important when training language models at scale: a large batch size is necessary for fast training, but a batch size that is *too large* will harm token efficiency. To navigate this tradeoff, McCandlish et al. (2018) suggest that a *critical batch size* (CBS), below which training will not substantially degrade loss, can be estimated based on the gradient noise scale during training. While their method has been adopted in practice, e.g., when training GPT-3, strong assumptions are required to justify gradient noise as a proxy for the CBS, which makes it unclear whether their approach should be trusted in practice, limiting its applicability. In this paper, we introduce a simple, empirical approach to *directly* measure the CBS and show how the CBS evolves over training. Applying our approach to the OLMo models, we find that CBS is near 0 at initialization, increases rapidly at first, and then plateaus as training progresses. Furthermore, we find that this trend holds across different model sizes (1B and 7B), suggesting CBS from small training runs can inform larger-scale training runs. Our findings about how the CBS changes over training motivate *batch size warmup* as a natural way to reliably train language models at large batch size: start the batch size small and increase it as the CBS grows. To validate this claim, we use batch size warmup to train OLMo 1B to slightly better loss than the original training run with 43% fewer gradient steps. This shows how our framework can be applied to reliably train language models at larger batch sizes, increasing data parallelism without compromising performance.

Prior studies on deception in language-based AI agents typically assess whether the agent produces a false statement about a topic, or makes a binary choice prompted by a goal, rather than allowing open-ended deceptive behavior to emerge in pursuit of a longer-term goal. To fix this, we introduce $\textit{Among Us}$, a sandbox social deception game where LLM-agents exhibit long-term, open-ended deception as a consequence of the game objectives. While most benchmarks saturate quickly, $\textit{Among Us}$ can be expected to last much longer, because it is a multi-player game far from equilibrium. Using the sandbox, we evaluate $18$ proprietary and open-weight LLMs and uncover a general trend: models trained with RL are comparatively much better at producing deception than detecting it. We evaluate the effectiveness of methods to detect lying and deception: logistic regression on the activations and sparse autoencoders (SAEs). We find that probes trained on a dataset of ``pretend you're a dishonest model: $\dots$'' generalize extremely well out-of-distribution, consistently obtaining AUROCs over 95% even when evaluated just on the deceptive statement, without the chain of thought. We also find two SAE features that work well at deception detection but are unable to steer the model to lie less. We hope our open-sourced sandbox, game logs, and probes serve to anticipate and mitigate deceptive behavior and capabilities in language-based agents.

Despite the remarkable reasoning performance, eliciting the long chain-of-thought(CoT) ability in large language models(LLMs) typically requires costly reinforcement learning or supervised fine-tuning on high-quality distilled data. We investigate the internal mechanisms behind this capability and show that a small set of high-impact activations in the last few layers, greatly govern the long-form reasoning attributes, e.g. output length and self-reflection. Through simply amplifying these activations and adding ``wait'' tokens, the long CoT ability can be invoked without training, leading to significantly increased self-reflection rate and accuracy. In addition, we also find that the activation changes follow predictable trajectories, i.e. a sharp rise after special tokens and a subsequent exponential decay. Based on these insights, we introduce a general training-free activation control technique. It utilizes a few contrastive examples to identify the relevant activations, and then incorporates simple analytic functions to adjust their values at inference time to elicit long CoTs. Extensive experiments have verified the effectiveness of our methods in efficiently eliciting the long CoT ability of LLMs and improving the performance. Besides, we further propose a parameter-efficient fine-tuning method that trains only the last-layer activation amplification module and a few LoRA layers, outperforming LoRA on reasoning benchmarks with much fewer parameters. Our code and data will be fully public released.

AI systems are increasingly able to autonomously conduct realistic software engineering tasks, and may soon be deployed to automate machine learning (ML) R\&D itself. Frontier AI systems may be deployed in safety-critical settings, including to help ensure the safety of future systems. Unfortunately, frontier and future systems may not be sufficiently trustworthy, and there is evidence that these systems may even be misaligned with their developers or users. Therefore, we investigate the capabilities of AI agents to act against the interests of their users when conducting ML engineering, by sabotaging ML models, sandbagging their performance, and subverting oversight mechanisms. First, we extend MLE-Bench, a benchmark for realistic ML tasks, with code-sabotage tasks such as implanting backdoors and purposefully causing generalisation failures. Frontier agents make meaningful progress on our sabotage tasks. In addition, we study agent capabilities to sandbag on MLE-Bench. Agents can calibrate their performance to specified target levels below their actual capability. To mitigate sabotage, we use LM monitors to detect suspicious agent behaviour, and we measure model capability to sabotage and sandbag without being detected by these monitors. Overall, monitors are capable at detecting code-sabotage attempts but our results suggest that detecting sandbagging is more difficult. Additionally, aggregating multiple monitor predictions works well, but monitoring may not be sufficiently reliable to mitigate sabotage in high-stakes domains. Our benchmark is implemented in the UK AISI’s Inspect framework and we make our code publicly available.

As we scale to more massive machine learning models, the frequent synchronization demands inherent in data-parallel approaches create significant slowdowns, posing a critical challenge to further scaling. Recent work develops an approach (DiLoCo) that relaxes synchronization demands without compromising model quality. However, these works do not carefully analyze how DiLoCo's behavior changes with model size. In this work, we study the scaling law behavior of DiLoCo when training LLMs under a fixed compute budget. We focus on how algorithmic factors, including number of model replicas, hyperparameters, and token budget affect training in ways that can be accurately predicted via scaling laws. We find that DiLoCo scales both predictably and robustly with model size. When well-tuned, DiLoCo scales better than data-parallel training with model size, and can outperform data-parallel training even at small model sizes. Our results showcase a more general set of benefits of DiLoCo than previously documented, including increased optimal batch sizes, improved downstream generalization with scale, and improved evaluation loss for a fixed token budget.

Enabling virtual humans to dynamically and realistically respond to diverse auditory stimuli remains a key challenge in character animation, demanding the integration of perceptual modeling and motion synthesis. Despite its significance, this task remains largely unexplored. Most previous works have primarily focused on mapping modalities like speech, audio, and music to generate human motion. As of yet, these models typically overlook the impact of spatial features encoded in spatial audio signals on human motion. To bridge this gap and enable high-quality modeling of human movements in response to spatial audio, we introduce the first comprehensive "Spatial Audio-Driven Human Motion" (SAM) dataset, which contains diverse and high-quality spatial audio and motion data. For benchmarking, we develop a simple yet effective diffusion-based generative framework for human "MOtion generation driven by SPatial Audio," termed MOSPA, which faithfully captures the relationship between body motion and spatial audio through an effective fusion mechanism. Once trained, MOSPA can generate diverse realistic human motions conditioned on varying spatial audio inputs. We perform a thorough investigation of the proposed dataset and conduct extensive experiments for benchmarking, where our method achieves state-of-the-art performance on this task. Our code and model are publicly available at https://github.com/xsy27/Mospa-Acoustic-driven-Motion-Generation.git

Recent years have witnessed the emergence of a spectrum of foundation models, covering a broad range of capabilities and costs. Often, we effectively use foundation models as feature generators and train classifiers that use the outputs of these models to make decisions. In this paper, we consider an increasingly relevant setting where we have two classifier stages. The first stage has access to features $x$ and has the option to make a classification decision or defer, while incurring a cost, to a second classifier that has access to features $x$ and $z$. This is similar to the ``learning to defer'' setting, with the important difference that we train both classifiers jointly, and the second classifier has access to more information. The natural loss for this setting is an $\ell_{01c}$ loss, where a penalty is paid for incorrect classification, as in $\ell_{01}$, but an additional penalty $c$ is paid for consulting the second classifier. The $\ell_{01c}$ loss is unwieldy for training. Our primary contribution in this paper is the derivation of a hinge-based surrogate loss $\ell^c_{hinge}$ that is much more amenable to training but also satisfies the property that $\ell^c_{hinge}$-consistency implies $\ell_{01c}$-consistency.

Modern tokenizers employ deterministic algorithms to map text into a single ``canonical" token sequence, yet the same string can be encoded as many non-canonical tokenizations using the language model vocabulary, including tokenizing by character. In this paper, we investigate the robustness of LMs to input encoded with non-canonical tokenizations entirely unseen during training. Surprisingly, when evaluated across 20 benchmarks, we find that instruction-tuned models retain up to 93.4\% of their original performance when given a randomly sampled tokenization, and 90.8\% with character-level tokenization. We find that overall stronger models tend to be more robust, and that robustness diminishes as the tokenization departs farther from the canonical form. Motivated by these results, we identify settings where non-canonical tokenization schemes can \textit{improve} performance, finding that character‑level segmentation improves string manipulation and code understanding tasks by up to 15\%, and right‑aligned digit grouping enhances large‑number arithmetic by over 33\%. Finally, we investigate the source of this robustness, finding that it arises in the instruction-tuning phase. We provide evidence that both base and post-trained models grasp the semantics of non-canonical tokenizations (perceiving them as containing misspellings). However, base models try to mimic the imagined mistakes and degenerate into nonsensical output, while post-trained models are committed to fluent responses. Overall, our findings suggest that models are less committed to their tokenizer than previously believed, and highlight the promise of intervening on tokenization at inference time to boost language model performance.

We introduce CameraBench, a large-scale dataset and benchmark designed to assess and improve camera motion understanding. CameraBench consists of ~3,000 diverse internet videos, annotated by experts through a rigorous multi-stage quality control process. One of our core contributions is a taxonomy or "language" of camera motion primitives, designed in collaboration with cinematographers. We find, for example, that some motions like "follow" (or tracking) require understanding scene content like moving subjects. We conduct a large-scale human study to quantify human performance, revealing that domain expertise and tutorial-based training can significantly enhance accuracy. For example, a novice may confuse zoom-in (a change of intrinsics) with translating forward (a change of extrinsics), but can be trained to differentiate the two. Using CameraBench, we evaluate Structure-from-Motion (SfM) and Video-Language Models (VLMs), finding that SfM models struggle to capture semantic primitives that depend on scene content, while generative VLMs struggle to capture geometric primitives that require precise estimation of trajectories. We then fine-tune a generative VLM on CameraBench to achieve the best of both worlds and showcase its applications, including motion-augmented captioning, video question answering, and video-text retrieval. We hope our taxonomy, benchmark, and tutorials will drive future efforts towards the ultimate goal of understanding camera motions in any video.

Recently, Large Foundation Models (LFMs), e.g., CLIP and GPT, have significantly advanced the Human-Object Interaction (HOI) detection, due to their superior generalization and transferability. Prior HOI detectors typically employ single- or multi-modal prompts to generate discriminative representations for HOIs from pretrained LFMs. However, such prompt-based approaches focus on transferring HOI-specific knowledge, but unexplore the potential reasoning capabilities of LFMs, which can provide informative context for ambiguous and open-world interaction recognition. In this paper, we propose InstructHOI, a novel method that leverages context-aware instructions to guide multi-modal reasoning for HOI detection. Specifically, to bridge knowledge gap and enhance reasoning abilities, we first perform HOI-domain fine-tuning on a pretrained multi-modal LFM, using a generated dataset with 140K interaction-reasoning image-text pairs. Then, we develop a Context-aware Instruction Generator (CIG) to guide interaction reasoning. Unlike traditional language-only instructions, CIG first mines visual interactive context at the human-object level, which is then fused with linguistic instructions, forming multi-modal reasoning guidance. Furthermore, an Interest Token Selector (ITS) is adopted to adaptively filter image tokens based on context-aware instructions, thereby aligning reasoning process with interaction regions. Extensive experiments on two public benchmarks demonstrate that our proposed method outperforms the state-of-the-art ones, under both supervised and zero-shot settings.

Long reasoning models (LRMs) such as OpenAI's o1 and DeepSeek's R1 have demonstrated remarkable abilities in complex reasoning tasks by scaling test-time compute and exhibiting human-like deep thinking. However, we identify a phenomenon we term underthinking, where LRMs frequently switch between different reasoning thoughts without sufficiently exploring promising paths to reach a correct solution. This behavior leads to inadequate depth of reasoning and decreased performance, particularly on challenging mathematical problems. To systematically analyze this issue, we conduct experiments on three challenging test sets and two representative open-source LRMs, revealing that frequent thought switching correlates with incorrect responses. We introduce a novel metric to quantify underthinking by measuring token efficiency in incorrect answers. To address underthinking, we propose a decoding strategy with thought switching penalty (Tip) that discourages premature transitions between thoughts, encouraging deeper exploration of each reasoning path. Experimental results demonstrate that our approach improves accuracy across challenging datasets without requiring model fine-tuning. Our findings contribute to understanding reasoning inefficiencies in LRMs and offer a practical solution to enhance their problem-solving capabilities. Our code is open-source and available at https://github.com/wangyuenlp/underthinking.

Verifying the behavior of neural networks is necessary if developers are to confidently deploy them as parts of mission-critical systems. Toward this end, researchers have been actively developing a range of increasingly sophisticated and scalable neural network verifiers. However, scaling verification to large networks is challenging, at least in part due to the significant memory requirements of verification algorithms. In this paper, we propose an assume-guarantee compositional framework, CoVeNN, that is parameterized by an underlying verifier to generate a sequence of verification sub-problems to address this challenge. We present an iterative refinement-based strategy for computing assumptions that allow sub-problems to retain sufficient accuracy. An evaluation using 7 neural networks and a total of 140 property specifications demonstrates that CoVeNN can verify nearly 7 times more problems than state-of-the-art verifiers. CoVeNN is part of the NeuralSAT verification project: https://github.com/dynaroars/neuralsat.

Modern interactive applications increasingly demand dynamic 3D content, yet the transformation of static 3D models into animated assets constitutes a significant bottleneck in content creation pipelines. While recent advances in generative AI have revolutionized static 3D model creation, rigging and animation continue to depend heavily on expert intervention. We present \textbf{Puppeteer}, a comprehensive framework that addresses both automatic rigging and animation for diverse 3D objects. Our system first predicts plausible skeletal structures via an auto-regressive transformer that introduces a joint-based tokenization strategy for compact representation and a hierarchical ordering methodology with stochastic perturbation that enhances bidirectional learning capabilities. It then infers skinning weights via an attention-based architecture incorporating topology-aware joint attention that explicitly encodes inter-joint relationships based on skeletal graph distances. Finally, we complement these rigging advances with a differentiable optimization-based animation pipeline that generates stable, high-fidelity animations while being computationally more efficient than existing approaches. Extensive evaluations across multiple benchmarks demonstrate that our method significantly outperforms state-of-the-art techniques in both skeletal prediction accuracy and skinning quality. The system robustly processes diverse 3D content, ranging from professionally designed game assets to AI-generated shapes, producing temporally coherent animations that eliminate the jittering issues common in existing methods.

Graph Neural Networks (GNNs) are widely used to compute representations of node pairs for downstream tasks such as link prediction. Yet, theoretical understanding of their expressive power has focused almost entirely on graph-level representations. In this work, we shift the focus to links and provide the first comprehensive study of GNN expressiveness in link representation. We introduce a unifying framework, the $k_\phi$-$k_\rho$-$m$ framework, that subsumes existing message-passing link models and enables formal expressiveness comparisons. Using this framework, we derive a hierarchy of state-of-the-art methods and offer theoretical tools to analyze future architectures. To complement our analysis, we propose a synthetic evaluation protocol comprising the first benchmark specifically designed to assess link-level expressiveness. Finally, we ask: does expressiveness matter in practice? We use a graph symmetry metric that quantifies the difficulty of distinguishing links and show that while expressive models may underperform on standard benchmarks, they significantly outperform simpler ones as symmetry increases, highlighting the need for dataset-aware model selection.

In object re-identification (ReID) task, both cross-modal and multi-modal retrieval methods have achieved notable progress. However, existing approaches are designed for specific modality and category (person or vehicle) retrieval task, lacking generalizability to others. Acquiring multiple task-specific models would result in wasteful allocation of both training and deployment resources. To address the practical requirements for unified retrieval, we introduce Multi-Modal and Multi-Task object ReID ($\rm {M^3T}$-ReID). The $\rm {M^3T}$-ReID task aims to utilize a unified model to simultaneously achieve retrieval tasks across different modalities and different categories. Specifically, to tackle the challenges of modality distibution divergence and category semantics discrepancy posed in $\rm {M^3T}$-ReID, we design a novel Unbiased Prototype Consistency Learning (UPCL) framework, which consists of two main modules: Unbiased Prototypes-guided Modality Enhancement (UPME) and Cluster Prototype Consistency Regularization (CPCR). UPME leverages modality-unbiased prototypes to simultaneously enhance cross-modal shared features and multi-modal fused features. Additionally, CPCR regulates discriminative semantics learning with category-consistent information through prototypes clustering. Under the collaborative operation of these two modules, our model can simultaneously learn robust cross-modal shared feature and multi-modal fused feature spaces, while also exhibiting strong category-discriminative capabilities. Extensive experiments on multi-modal datasets RGBNT201 and RGBNT100 demonstrates our UPCL framework showcasing exceptional performance for $\rm {M^3T}$-ReID. The code is available at https://github.com/ZhouZhongao/UPCL.