Finetuning large language models on narrowly harmful datasets can cause them to become emergently misaligned, giving stereotypically `evil' responses across diverse unrelated settings. Concerningly, a pre-registered survey of experts failed to predict this result, highlighting our poor understanding of the inductive biases governing learning and generalisation in LLMs. We use emergent misalignment (EM) as a case study to investigate these inductive biases, and find that although models can learn the narrow dataset task, the general solution is measurably more stable and more efficient. To establish this, we first demonstrate that EM is a robust phenomena by introducing new datasets which induce misalignment more consistently and coherently than prior work. We show that different EM finetunes converge to the same linear representation of general misalignment, which can be used to mediate misaligned behaviour. However, a linear representation of the narrow solution also exists, and can be learned by introducing a KL divergence loss. Comparing these representations reveals that general misalignment achieves lower loss, is more robust to perturbations, and is more influential in the pre-training distribution. This work isolates a concrete representation of general misalignment for monitoring and mitigation. More broadly, it offers a detailed case study and metrics for understanding how inductive biases shape generalisation in LLMs.
论文检索
输入标题、作者或关键词，从 12,113 篇学术成果中精准定位
Applications · Everything Else
Large Language Models (LLMs) have shown remarkable capabilities in solving various programming tasks, such as code generation. However, their potential for code optimization, particularly in performance enhancement, remains largely unexplored. This paper investigates the capabilities of LLMs in optimizing code for minimal execution time, addressing a critical gap in current research. The recently proposed code optimization methods constructs program optimization pairs based on iterative submissions from the same programmer for the same problem. However, this approach confines LLMs to local performance improvements, neglecting global algorithmic innovation. To overcome this limitation, we adopt a completely different perspective by reconstructing the optimization pairs into a problem-oriented approach. This allows for the integration of various ideas from multiple programmers tackling the same problem. Furthermore, we observe that code optimization presents greater challenges compared to code generation, often accompanied by "optimization tax". Recognizing the inherent trade-offs in correctness and efficiency, we introduce a novel anchor verification framework to mitigate this "optimization tax." Ultimately, the problem oriented perspective combined with anchor verification framework significantly enhances both the correct optimization ratio and speedup to new levels.
Applications · Chemistry and Drug Discovery
Foundation MLIPs demonstrate broad applicability across diverse material systems and have emerged as a powerful and transformative paradigm in chemical and computational materials science. Equivariant MLIPs achieve state-of-the-art accuracy in a wide range of benchmarks by incorporating equivariant inductive bias. However, the reliance on tensor products and high-degree representations makes them computationally costly. This raises a fundamental question: as quantum mechanical-based datasets continue to expand, can we develop a more compact model to thoroughly exploit high-dimensional atomic interactions? In this work, we present MatRIS (\textbf{Mat}erials \textbf{R}epresentation and \textbf{I}nteraction \textbf{S}imulation), an invariant MLIP that introduces attention-based modeling of three-body interactions. MatRIS leverages a novel separable attention mechanism with linear complexity $O(N)$, enabling both scalability and expressiveness. MatRIS delivers accuracy comparable to that of leading equivariant models on a wide range of popular benchmarks (Matbench-Discovery, MatPES, MDR phonon, Molecular dataset, etc). Taking Matbench-Discovery as an example, MatRIS achieves an F1 score of up to 0.847 and attains comparable accuracy at a lower training cost. The work indicates that our carefully designed invariant models can match or exceed the accuracy of equivariant models at a fraction of the cost, shedding light on the development of accurate and efficient MLIPs.
Applications · Robotics
Reinforcement learning (RL) is widely used for humanoid control, with on-policy methods such as Proximal Policy Optimization (PPO) enabling robust training via large-scale parallel simulation and, in some cases, zero-shot deployment to real robots. However, the low sample efficiency of on-policy algorithms limits safe adaptation to new environments. Although off-policy RL and model-based RL have shown improved sample efficiency, the gap between large-scale pretraining and efficient finetuning on humanoids still exists. In this paper, we find that off-policy Soft Actor-Critic (SAC), with large-batch update and a high Update-To-Data (UTD) ratio, reliably supports large-scale pretraining of humanoid locomotion policies, achieving zero-shot deployment on real robots. For adaptation, we demonstrate that these SAC-pretrained policies can be finetuned in new environments and out-of-distribution tasks using model-based methods. Data collection in the new environment executes a deterministic policy while stochastic exploration is instead confined to a physics-informed world model. This separation mitigates the risks of random exploration during adaptation while preserving exploratory coverage for improvement. Overall, the approach couples the wall-clock efficiency of large-scale simulation during pretraining with the sample efficiency of model-based learning during fine-tuning. Code and videos: https://lift-humanoid.github.io
Social Aspects · Trustworthy Machine Learning
We characterize the effectiveness of Sharpness-aware minimization (SAM) under machine unlearning scheme, where unlearning forget signals interferes with learning retain signals. While previous work prove that SAM improves generalization with noise memorization prevention, we show that SAM abandons such denoising property when fitting the forget set, leading to altered generalization depending on signal strength. We further characterize the signal surplus of SAM in the order of signal strength, which enables learning from less retain signals to maintain model performance and putting more weight on unlearning the forget set. Empirical studies show that SAM outperforms SGD with relaxed requirement for retain signals and can enhance various unlearning methods either as pretrain or unlearn algorithm. Motivated by our refined characterization of SAM unlearning and observing that overfitting can benefit more stringent sample-specific unlearning, we propose Sharp MinMax, which splits the model into two to learn retain signals with SAM and unlearn forget signals with sharpness maximization, achieving best performance. Extensive experiments show that SAM enhances unlearning across varying difficulties measured by memorization, yielding decreased feature entanglement between retain and forget sets, stronger resistance to membership inference attacks, and a flatter loss landscape. Our observations generalize to more noised data, different optimizers, and different architectures.
Applications · Robotics
Non-prehensile manipulation offers a robust alternative to traditional pick-and-place methods for object repositioning. However, learning such skills with dexterous, multi-fingered hands remains largely unexplored, leaving their potential for stable and efficient manipulation underutilized. Progress has been limited by the lack of large-scale, contact-aware non-prehensile datasets for dexterous hands and the absence of wrist–finger control policies. To bridge these gaps, we present DexMove, a tactile-guided non-prehensile manipulation framework for dexterous hands. DexMove combines a scalable simulation pipeline that generates physically plausible wrist–finger trajectories with a wearable device, which captures multi-finger contact data from human demonstrations using vision-based tactile sensors. Using these data, we train a flow-based policy that enables real-time, synergistic wrist–finger control for robust non-prehensile manipulation of diverse tabletop objects. In real-world experiments, DexMove successfully manipulated six objects of varying shapes and materials, achieving a 77.8\% success rate. Our method outperforms ablated baselines by 36.6\% and improves efficiency by nearly 300\%. Furthermore, the learned policy generalizes to language-conditioned, long-horizon tasks such as object sorting and desktop tidying.
Computer Vision · Image and Video Generation
Cultural nuances are best expressed through social interactions, yet current text-to-image (T2I) benchmarks focus largely on object-centric artifacts (e.g., food, landmarks, and attire). In this work, we study the cultural faithfulness of T2I models (i.e., adherence to the target culture) through social activities. To this end, we introduce CULTIVate, a new benchmark of 576 activities across 9 categories (e.g., dancing, greeting, dining) with over 19,000 images from 16 countries. We further propose AHEaD, an explainable framework that measures cultural understanding along four dimensions: cultural Alignment, Hallucination, Exaggeration, and Diversity. Unlike prior work relying on costly human evaluation or image-text alignment (ITA), AHEaD uses culturally-grounded descriptors to provide quantitative, interpretable feedback that enables iterative image refinement. Our analysis shows ITA metrics correlate poorly with human judgments and that alignment alone is insufficient to capture faithfulness. In contrast, FAITH (combining alignment, hallucination, and exaggeration) achieves 27% higher correlation than baselines. Finally, we observe systematic disparities, with generated images being consistently more faithful for Global North than Global South cultures.
Social Aspects · Trustworthy Machine Learning
As artificial intelligence (AI) assistants become more widely adopted in safety-critical domains, it becomes important to develop safeguards against potential failures or adversarial attacks. A key prerequisite to developing these safeguards is understanding the ability of these AI assistants to mislead human teammates. We investigate this attack problem within the context of an intellective strategy game where a team of three humans and one AI assistant collaborate to answer a series of trivia questions. Unbeknownst to the humans, the AI assistant is adversarial. Leveraging techniques from Model-Based Reinforcement Learning (MBRL), the AI assistant learns a model of the humans' trust evolution and uses that model to manipulate the group decision-making process to harm the team. We evaluate two models---one inspired by literature and the other data-driven---and find that both can effectively harm the human team. Moreover, we find that in this setting while our data-driven model is the most capable of accurately predicting how human agents appraise their teammates given limited information on prior interactions, the model based on principles of cognitive psychology does not lag too far behind. Finally, we compare the performance of state-of-the-art LLM models to human agents on our influence allocation task to evaluate whether the LLMs allocate influence similarly to humans or if they are more robust to our attack. These results enhance our understanding of decision-making dynamics in small human-AI teams and lay the foundation for defense strategies.
Reinforcement Learning · Deep RL
Many robotic control tasks require policies to act on orientations, yet the geometry of SO(3) makes this nontrivial. Because SO(3) admits no global, smooth, minimal parameterization, common representations such as Euler angles, quaternions, rotation matrices, and Lie algebra coordinates introduce distinct constraints and failure modes. While these trade-offs are well studied for supervised learning, their implications for actions in reinforcement learning remain unclear. We systematically evaluate SO(3) action representations across three standard continuous control algorithms, PPO, SAC, and TD3, under dense and sparse rewards. We compare how representations shape exploration, interact with entropy regularization, and affect training stability through empirical studies and analyze the implications of different projections for obtaining valid rotations from Euclidean network outputs. Across a suite of robotics benchmarks, we quantify the practical impact of these choices and distill simple, implementation-ready guidelines for selecting and using rotation actions. Our results highlight that representation-induced geometry strongly influences exploration and optimization and show that representing actions as tangent vectors in the local frame yields the most reliable results across algorithms. The project webpage and code are available at amacati.github.io/so3_primer.
Computer Vision · Image and Video Generation
Recent advances in illumination control extend image-based methods to video, yet still facing a trade-off between lighting fidelity and temporal consistency. Moving beyond relighting, a key step toward generative modeling of real-world scenes is the joint control of camera trajectory and illumination, since visual dynamics are inherently shaped by both geometry and lighting. To this end, we present Light-X, a video generation framework that enables controllable rendering from monocular videos with both viewpoint and illumination control. 1) We propose a disentangled design that decouples geometry and lighting signals: geometry and motion are captured via dynamic point clouds projected along user-defined camera trajectories, while illumination cues are provided by a relit frame consistently projected into the same geometry. These explicit, fine-grained cues enable effective disentanglement and guide high-quality illumination. 2) To address the lack of paired multi-view and multi-illumination videos, we introduce Light-Syn, a degradation-based pipeline with inverse-mapping that synthesizes training pairs from in-the-wild monocular footage. This strategy yields a dataset covering static, dynamic, and AI-generated scenes, ensuring robust training. Extensive experiments show that Light-X outperforms baseline methods in joint camera-illumination control and surpasses prior video relighting methods under both text- and background-conditioned settings.
Deep Learning · Generative Models and Autoencoders
Large Language Models (LLMs) often experience compounding errors during long text generation. Early mistakes can propagate and lead to drift, faulty reasoning, or repetition. While scaling up models improves capabilities, it requires substantial computational resources, and the resulting self-correction behaviour remains unpredictable at inference time. Self-correction is a promising technique for addressing this issue. However, existing approaches have limitations. Supervised training methods can build self-correcting behaviours into models, but require training data collection and lack cross-domain generalizability. Current post-hoc iterative refinement methods operate only at inference time, but must wait for substantial portions of the draft to be generated before providing feedback. This feedback does not guarantee effective guidance, and the same mistake patterns can still reappear. In this paper, we introduce Once-More, a model-agnostic post-hoc self-correction framework that intervenes during generation. Once-More leverages token-level perplexity and feedback from verifiers to provide continuous guided steering of the generation path through a logit redistribution mechanism. This approach essentially helps accumulate "more correct" steps throughout the generation process. Evaluation on multiple benchmarks demonstrates that Once-More achieves state-of-the-art results compared to other self-correction methods. To our knowledge, Once-More is the first post-hoc method to leverage token perplexity and external feedback to perform continuous guided self-correction.
General Machine Learning · Probabilistic Methods
Many real-world classification problems, such as plant identification, have extremely long-tailed class distributions. In order for prediction sets to be useful in such settings, they should (i) provide good class-conditional coverage, ensuring that rare classes are not systematically omitted from the prediction sets, and (ii) be a reasonable size, allowing users to easily verify candidate labels. Unfortunately, existing conformal prediction methods, when applied to the long-tailed setting, force practitioners to make a binary choice between small sets with poor class-conditional coverage or sets that have very good class-conditional coverage but are extremely large. We propose methods with marginal coverage guarantees that smoothly trade off set size and class-conditional coverage. First, we introduce a new conformal score function called prevalence-adjusted softmax that optimizes for macro-coverage, defined as the average class-conditional coverage across classes. Second, we propose a new procedure that interpolates between marginal and class-conditional conformal prediction by linearly interpolating their conformal score thresholds. We demonstrate our methods on Pl@ntNet-300K and iNaturalist-2018, two long-tailed image datasets with 1,081 and 8,142 classes, respectively.
Reinforcement Learning · Deep RL
With the rapid advancement of data science, charts have evolved from simple numerical presentation tools to essential instruments for insight discovery and decision-making support. However, current chart data intelligence exhibits significant limitations in deep research capabilities, with existing methods predominantly addressing shallow tasks such as visual recognition or factual question-answering, rather than the complex reasoning and high-level data analysis that deep research requires. This limitation stems from two primary technical bottlenecks: at the training level, existing post-training techniques exhibit deficiencies in handling multi-dimensional reward signal interference and heterogeneous data gradient conflicts, preventing models from achieving balanced development across multiple capability dimensions; at the evaluation level, current methods remain limited to factual retrieval and basic computation, failing to assess end-to-end analytic reasoning and other deep research capabilities. To address the training challenge, we propose PRPO, which performs parallel optimization across reward dimensions and capability partitioning across data types, effectively disentangling conflicts between heterogeneous data and multi-dimensional reward signals while ensuring optimization stability. For the evaluation challenge, we construct MCDR-Bench based on the ``error uniqueness principle," transforming subjective generation assessment into objective error identification through controllable error injection, enabling quantifiable evaluation of deep research capabilities. Experimental validation confirms that the proposed PRPO and MCDR-Bench jointly establish a unified framework that systematically advances chart deep research through enhanced collaborative training and objective evaluation.
Theory · Learning Theory
Motivated by the prevalence and success of machine learning, a line of recent work has studied learning-augmented algorithms in the streaming model. These results have shown that for natural and practical oracles implemented with machine learning models, we can obtain streaming algorithms with improved space efficiency that are otherwise provably impossible. On the other hand, our understanding is much more limited for the sliding window model, which captures applications where either recent data leads to better or older data must be expunged from the dataset, e.g., by privacy regulation laws. In this paper, we utilize an oracle for the heavy-hitters of datasets to give learning-augmented algorithms for a number of fundamental problems in the sliding window model, such as norm/moment estimation, frequency estimation, cascaded norms, and rectangular moment estimation. We complement our theoretical results with a number of empirical evaluations that demonstrate the practical efficiency of our algorithms on real and synthetic datasets.
Applications · Health
Large language models perform well on many medical QA benchmarks, but real clinical reasoning is harder because diagnosis often requires integrating evidence across multiple images rather than interpreting a single view. We introduce MedThinkVQA, an expert-annotated benchmark for thinking with multiple images, in which models must interpret each image, combine cross-view evidence, and solve diagnostic questions under intermediate supervision and step-level evaluation. The dataset contains 10,067 cases, including 720 test cases, with an average of 6.68 images per case, substantially denser than prior work (earlier maxima $\leq$ 1.43). On the test set, the best closed-source models, Claude-4.6-opus, Gemini-3-pro, and GPT-5.2-xhigh, achieve only 54.9%--57.2% accuracy, while smaller proprietary variants, GPT-5-mini/nano, drop to 39.7% and 30.8%. Top open-source models perform worse overall, with Qwen3.5-397B-A17B (52.2%) and Qwen3.5-27B (50.6%) leading, followed by Lingshu-32B (43.2%), InternVL3.5-38B (40.7%), and MedGemma-27B (31.8%). Further analysis points to a single-core bottleneck: current models struggle with grounded multi-image reasoning, i.e., reliably extracting, aligning, and composing evidence across views before higher-level inference can help. This is supported by three consistent findings: adding expert-provided single-image cues and integrating cross-image evidence improve performance, whereas replacing them with models’ self-generated intermediates reduces accuracy. Step-level analysis shows that over 70% of errors come from image reading and cross-view integration, with reasoning failures increasing on decisive steps. Scaling results show that while accuracy increases with more images, additional inference-time computation is beneficial only when the underlying visual grounding is already reliable. When early evidence extraction is weak, longer reasoning yields limited or unstable gains and can even amplify misread cues. Together, these results show that the main barrier is not simply insufficient reasoning length or depth, but the lack of reliable mechanisms for grounding, aligning, and composing distributed evidence across real-world, cross-view, multimodal clinical inputs.
Large-language-models (LLMs) demonstrate enormous utility in long-context tasks which require processing prompts that consist of tens to hundreds of thousands of tokens. However, existing LLM training libraries do not provide easy to use abstractions to optimize for long-context training, instead focusing on optimizations for models with large parameter counts through ZeRO-3/FSDP, Tensor and Pipeline parallelism. This forces users to rewrite LLM training libraries to incorporate compositions of various complex long-context optimizations, such as sequence-parallelism, to training pipelines; a process that requires in-depth expertise, reducing developer productivity. To tackle these challenges, we introduce AutoSP: the first automated solution to automatically optimize LLM training for longer-contexts. AutoSP compiles models and applies a targeted set of optimizations: automated sequence parallelism, and long-context aware activation-checkpointing, to drastically enhance LLM trainability at negligible cost to throughput. Our evaluation demonstrates AutoSP's capability on both NVIDIA and AMD hardware, increasing training contexts by upto 2.7$\times$ and 2.5$\times$ respectively over competitive hand-written baseline at negligible cost to runtime performance.
Deep Learning · Attention Mechanisms
Transformer-based language models (LMs) notably struggle to reliably capture distant contextual information. This work introduces a novel approach using meta-tokens -- special tokens injected during pre-training -- paired with a dedicated meta-attention mechanism to guide LMs to use these tokens. We pre-train a language model equipped with meta-attention in addition to causal multi-head attention on <100B tokens, achieving strong performance on a suite of synthetic tasks. Our method facilitates length generalization up to 2$\times$ the context window after extension with YaRN. We provide an information-theoretic analysis which reveals that meta-tokens \textit{sharpen} the positional encoding, allowing them to operate as content-based anchors that compress preceding context and “cache” it within the meta-token. We empirically confirm this by visualizing model internals to study the residual stream. Together, our findings demonstrate that meta-tokens and meta-attention provide a simple, data-efficient pre-training method, grounded by new mechanistic insights into their role in enabling length generalization behavior.
Deep Learning · Attention Mechanisms
Recently, large language models (LLMs) have made remarkable progress, with multilingual capability emerging as a core foundational strengths. However, the internal mechanisms by which these models perform translation remain incompletely understood. In this paper, we elucidate the relationship between the attention mechanism in LLMs and their translation abilities. We find that certain attention heads, which we term token alignment heads, are specifically responsible for mapping tokens from the source language to the target language during inference. Through a systematic investigation across various models, we confirm that these token alignment heads exhibit several key characteristics: (1) Universality: They are present in all LLMs we studied. (2) Sparsity: They constitute only a small fraction of all attention heads. (3) Consistency: The set of token alignment heads activated by the model shows strong consistency across different language pairs. (4) Causality: Interventionally removing these heads leads to a sharp decline in the model's translation performance, while randomly removing non-token alignment heads has little impact on translation ability. (5) Functional Specificity: Ablating token alignment heads disproportionately harms translation but has a varied impact on other multilingual tasks. We also traced the formation of token alignment heads during pre-training, revealing an evolutionary path of rapid proliferation, stabilization, and eventual pruning. Furthermore we leverage these token alignment heads to filter multilingual training data, and our experiments show that these data could enhance translation capabilities of the models.
Deep Learning · Generative Models and Autoencoders
Recent advances align diffusion models with human preferences to increase aesthetic appeal and mitigate artifacts and biases. Such methods aim to maximize a conditional output distribution aligned with higher rewards whilst not drifting far from a pretrained prior. This is commonly enforced by KL (Kullback–Leibler) regularization. As such, a central issue still remains: how does one choose the right regularization strength? Too high of a strength leads to limited alignment and too low of a strength leads to "reward hacking". This renders the task of choosing the correct regularization strength highly non-trivial. Existing approaches sweep over this hyperparameter by aligning a pretrained model at multiple regularization strengths and then choose the best strength. Unfortunately, this is prohibitively expensive. We introduce _DeRaDiff_, a _denoising-time realignment_ procedure that, after aligning a pretrained model once, modulates the regularization strength _during sampling_ to emulate models trained at other regularization strengths—_without any additional training or fine-tuning_. Extending decoding-time realignment from language to diffusion models, DeRaDiff operates over iterative predictions of continuous latents by replacing the reverse-step reference distribution by a geometric mixture of an aligned and reference posterior, thus giving rise to a closed-form update under common schedulers and a single tunable parameter, $\lambda$, for on-the-fly control. Our experiments show that across multiple text–image alignment and image-quality metrics, our method consistently provides a strong approximation for models aligned entirely from scratch at different regularization strengths. Thus, by enabling very precise inference-time control of the regularization strength, our method yields an efficient way to search for the optimal strength, eliminating the need for expensive alignment sweeps and thereby substantially reducing computational costs. The official implementation is available at https://github.com/itsShahain/DeRaDiff.
Computer Vision · 3D Rendering & Reconstruction
Physically-based rendering (PBR) provides a principled standard for realistic material–lighting interactions in computer graphics. Despite recent advances in generating PBR textures, existing methods fail to address two fundamental challenges: 1) materials decomposition from image prompts under limited illumination cues, and 2) seamless and view-consistent texture completion. To this end, we propose LumiTex, an end-to-end framework that comprises three key components: (1) a multi-branch generation scheme that disentangles albedo and metallic–roughness under shared illumination priors for robust material understanding, (2) a lighting-aware material attention mechanism that injects illumination context into the decoding process for physically grounded generation of albedo, metallic, and roughness maps, and (3) a geometry-guided inpainting module based on a large view synthesis model that enriches texture coverage and ensures seamless, view-consistent UV completion. Extensive experiments demonstrate that LumiTex achieves state-of-the-art performance in texture quality, surpassing both existing open-source and commercial methods. Project page: [https://lumitex.vercel.app](https://lumitex.vercel.app).