论文检索

Deep Learning · Large Language Models

Wenbo Pan, Zhichao Liu, Qiguang Chen, Xiangyang Zhou, Yu Haining, Xiaohua Jia

Large Language Models' safety-aligned behaviors, such as refusing harmful queries, can be represented by linear directions in activation space. Previous research modeled safety behavior with a single direction, limiting mechanistic understanding to an isolated safety feature. In this work, we discover that safety-aligned behavior is jointly controlled by multi-dimensional directions. Namely, we study the vector space of representation shifts during safety fine-tuning on Llama 3 8B for refusing jailbreaks. By studying orthogonal directions in the space, we first find that a dominant direction governs the model's refusal behavior, while multiple smaller directions represent distinct and interpretable features like hypothetical narrative and role-playing. We then measure how different directions promote or suppress the dominant direction, showing the important role of secondary directions in shaping the model's refusal representation. Finally, we demonstrate that removing certain trigger tokens in harmful queries can mitigate these directions to bypass the learned safety capability, providing new insights on understanding safety alignment vulnerability from a multi-dimensional perspective.

Deep Learning · Attention Mechanisms

Matteo Saponati, Pascal J. Sager, Pau Vilimelis Aceituno, Thilo Stadelmann, Benjamin F. Grewe

Self-attention is essential to Transformer architectures, yet how information is embedded in the self-attention matrices and how different objective functions impact this process remains unclear. We present a mathematical framework to analyze self-attention matrices by deriving the structures governing their weight updates. Using this framework, we demonstrate that bidirectional training induces symmetry in the weight matrices, while autoregressive training results in directionality and column dominance. Our theoretical findings are validated across multiple Transformer models — including ModernBERT, GPT, LLaMA3, and Mistral — and input modalities like text, vision, and audio. Finally, we apply these insights by showing that symmetric initialization improves the performance of encoder-only models on language tasks. This mathematical analysis offers a novel theoretical perspective on how information is embedded through self-attention, thereby improving the interpretability of Transformer models.

Theory · Probabilistic Methods

Lorenzo Lucchese, Mikko S. Pakkanen, Almut E. D. Veraart

The expected signature maps a collection of data streams to a lower dimensional representation, with a remarkable property: the resulting feature tensor can fully characterize the data generating distribution. This "model-free"' embedding has been successfully leveraged to build multiple domain-agnostic machine learning (ML) algorithms for time series and sequential data. The convergence results proved in this paper bridge the gap between the expected signature's empirical discrete-time estimator and its theoretical continuous-time value, allowing for a more complete probabilistic interpretation of expected signature-based ML methods. Moreover, when the data generating process is a martingale, we suggest a simple modification of the expected signature estimator with significantly lower mean squared error and empirically demonstrate how it can be effectively applied to improve predictive performance.

Applications · Everything Else

Yinyan Bu, Jiajie Yu, Kai Zheng, Xinyu Zhang, Piya Pal

We address the challenge of achieving angular super-resolution in multi-antenna radar systems that are widely used for localization, navigation, and automotive perception. A multi-antenna radar achieves very high resolution by computationally creating a large virtual sensing system using very few physical antennas. However, practical constraints imposed by hardware, noise, and a limited number of antennas can impede its performance. Conventional supervised learning models that rely on extensive pre-training with large datasets, often exhibit poor generalization in unseen environments. To overcome these limitations, we propose NEAR, an untrained implicit neural representation (INR) framework that predicts radar responses at unseen locations from sparse measurements, by leveraging latent harmonic structures inherent in radar wave propagation. We establish new theoretical results linking antenna array response to expressive power of INR architectures, and develop a novel physics-informed and latent geometry-aware regularizer. Our approach integrates classical signal representation with modern implicit neural learning, enabling high-resolution radar sensing that is both interpretable and generalizable. Extensive simulations and real-world experiments using radar platforms demonstrate NEAR's effectiveness and its ability to adapt to unseen environments.

Social Aspects · Accountability, Transparency, and Interpretability

Jessica Dai, Paula Gradu, Inioluwa Raji, Benjamin Recht

When an individual reports a negative interaction with some system, how can their personal experience be contextualized within broader patterns of system behavior? We study the *reporting database* problem, where individual reports of adverse events arrive sequentially, and are aggregated over time. In this work, our goal is to identify whether there are subgroups—defined by any combination of relevantfeatures—that are disproportionately likely to experience harmful interactions with the system. We formalize this problem as a sequential hypothesis test, and identify conditions on reporting behavior that are sufficient for making inferences about disparities in true rates of harm across subgroups. We show that algorithms for sequential hypothesis tests can be applied to this problem with a standard multiple testingcorrection. We then demonstrate our method on real-world datasets, including mortgage decisions and vaccine side effects; on each, our method (re-)identifies subgroups known to experience disproportionate harm using only a fraction of the data that was initially used to discover them.

Deep Learning · Large Language Models

Wen Lai, Alexander Fraser, Ivan Titov

Parameter-efficient fine-tuning (PEFT) methods, such as LoRA, are commonly used to adapt LLMs. However, the effectiveness of standard PEFT methods is limited in low-resource scenarios with only a few hundred examples. Recent advances in interpretability research have inspired the emergence of activation editing (or steering) techniques, which modify the activations of specific model components. Due to their extremely small parameter counts, these methods show promise for small datasets. However, their performance is highly dependent on identifying the correct modules to edit and often lacks stability across different datasets. In this paper, we propose Joint Localization and Activation Editing (JoLA), a method that jointly learns (1) which heads in the Transformer to edit (2) whether the intervention should be additive, multiplicative, or both and (3) the intervention parameters themselves - the vectors applied as additive offsets or multiplicative scalings to the head output. Through evaluations on three benchmarks spanning commonsense reasoning, natural language understanding, and natural language generation, we demonstrate that JoLA consistently outperforms existing methods.

Position · Methodology

Patrik Reizinger, Randall Balestriero, David Klindt, Wieland Brendel

Self-Supervised Learning (SSL) powers many current AI systems. As research interest and investment grow, the SSL design space continues to expand. The Platonic view of SSL, following the Platonic Representation Hypothesis (PRH), suggests that despite different methods and engineering approaches, all representations converge to the same Platonic ideal. However, this phenomenon lacks precise theoretical explanation. By synthesizing evidence from Identifiability Theory (IT), we show that the PRH can emerge in SSL. There is a gap between SSL theory and practice: Current IT cannot explain SSL's empirical success, though it has practically relevant insights. Our work formulates a blueprint for SSL research to bridge this gap: we propose expanding IT into what we term Singular Identifiability Theory (SITh), a broader theoretical framework encompassing the entire SSL pipeline. SITh would allow deeper insights into the implicit data assumptions in SSL and advance the field towards learning more interpretable and generalizable representations. We highlight three critical directions for future research: 1) training dynamics and convergence properties of SSL; 2) the impact of finite samples, batch size, and data diversity; and 3) the role of inductive biases in architecture, augmentations, initialization schemes, and optimizers.

Reinforcement Learning · Multi-agent

Ziyan Wang, Zhicheng Zhang, Fei Fang, Yali Du

Designing effective reward functions in multi-agent reinforcement learning (MARL) is a significant challenge, often leading to suboptimal or misaligned behaviors in complex, coordinated environments. We introduce Multi-agent Reinforcement Learning from Multi-phase Human Feedback of Mixed Quality ($\text{M}^3\text{HF}$), a novel framework that integrates multi-phase human feedback of mixed quality into the MARL training process. By involving humans with diverse expertise levels to provide iterative guidance, $\text{M}^3\text{HF}$ leverages both expert and non-expert feedback to continuously refine agents' policies. During training, we strategically pause agent learning for human evaluation, parse feedback using large language models to assign it appropriately and update reward functions through predefined templates and adaptive weights by using weight decay and performance-based adjustments. Our approach enables the integration of nuanced human insights across various levels of quality, enhancing the interpretability and robustness of multi-agent cooperation. Empirical results in challenging environments demonstrate that $\text{M}^3\text{HF}$ significantly outperforms state-of-the-art methods, effectively addressing the complexities of reward design in MARL and enabling broader human participation in the training process.

General Machine Learning · Representation Learning

Angel Villar-Corrales, Sven Behnke

Predicting future scene representations is a crucial task for enabling robots to understand andinteract with the environment. However, most existing methods rely on videos and simulationswith precise action annotations, limiting their ability to leverage the large amount of avail-able unlabeled video data. To address this challenge, we propose PlaySlot, an object-centricvideo prediction model that infers object representations and latent actions from unlabeledvideo sequences. It then uses these representations to forecast future object states and videoframes. PlaySlot allows the generation of multiple possible futures conditioned on latent actions,which can be inferred from video dynamics, provided by a user, or generated by a learned actionpolicy, thus enabling versatile and interpretable world modeling. Our results show that PlaySlotoutperforms both stochastic and object-centric baselines for video prediction across different environments. Furthermore, we show that our inferred latent actions can be used to learn robot behaviors sample-efficiently from unlabeled videodemonstrations. Videos and code are available on our project website.

Applications · Robotics

Yuqi Fan, Zhiyong Cui, Zhenning Li, Yilong Ren, Haiyang Yu

Reliable planning is crucial for achieving autonomous driving. Rule-based planners are efficient but lack generalization, while learning-based planners excel in generalization yet have limitations in real-time performance and interpretability. In long-tail scenarios, these challenges make planning particularly difficult. To leverage the strengths of both rule-based and learning-based planners, we proposed the **Scenario-Aware Hybrid Planner** (SAH-Drive) for closed-loop vehicle trajectory planning. Inspired by human driving behavior, SAH-Drive combines a lightweight rule-based planner and a comprehensive learning-based planner, utilizing a dual-timescale decision neuron to determine the final trajectory. To enhance the computational efficiency and robustness of the hybrid planner, we also employed a diffusion proposal number regulator and a trajectory fusion module. The experimental results show that the proposed method significantly improves the generalization capability of the planning system, achieving state-of-the-art performance in interPlan, while maintaining computational efficiency without incurring substantial additional runtime.

Applications · Language, Speech and Dialog

Tom Labiausse, Laurent Mazaré, Edouard Grave, Alexandre Défossez, Neil Zeghidour

We introduce Hibiki, a decoder-only model for simultaneous speech translation. Hibiki leverages a multistream language model to synchronously process source and target speech, and jointly produces text and audio tokens to perform speech-to-text and speech-to-speech translation. We furthermore address the fundamental challenge of simultaneous interpretation, which unlike its consecutive counterpart --where one waits for the end of the source utterance to start translating-- adapts its flow to accumulate just enough context to produce a correct translation in real-time, chunk by chunk. To do so, we introduce a weakly-supervised method that leverages the perplexity of an off-the-shelf text translation system to identify optimal delays on a per-word basis and create aligned synthetic data. After supervised training, Hibiki performs adaptive, simultaneous speech translation with vanilla temperature sampling. On a French-English simultaneous speech translation task, Hibiki demonstrates state-of-the-art performance in translation quality, speaker fidelity and naturalness. Moreover, the simplicity of its inference process makes it compatible with batched translation and even real-time on-device deployment. We provide examples on *huggingface.co/spaces/kyutai/hibiki-samples* as well as models and inference code at *github.com/kyutai-labs/hibiki*.

Applications · Health / Medicine

Adibvafa Fallahpour, Jun Ma, Alif Munim, Hongwei Lyu, BO WANG

Chest X-rays (CXRs) play an integral role in driving critical decisions in disease management and patient care. While recent innovations have led to specialized models for various CXR interpretation tasks, these solutions often operate in isolation, limiting their practical utility in clinical practice. We present MedRAX, the first versatile AI agent that seamlessly integrates state-of-the-art CXR analysis tools and multimodal large language models into a unified framework. MedRAX dynamically leverages these models to address complex medical queries without requiring additional training. To rigorously evaluate its capabilities, we introduce ChestAgentBench, a comprehensive benchmark containing 2,500 complex medical queries across 7 diverse categories. Our experiments demonstrate that MedRAX achieves state-of-the-art performance compared to both open-source and proprietary models, representing a significant step toward the practical deployment of automated CXR interpretation systems. Data and code have been publicly available at https://github.com/bowang-lab/MedRAX

Applications · Computer Vision

Zhuoying Li, Zhu Xu, Yuxin Peng, Yang Liu

Instruction-based image editing, which aims to modify the image faithfully towards instruction while preserving irrelevant content unchanged, has made advanced progresses. However, there still lacks a comprehensive metric for assessing the editing quality. Existing metrics either require high costs concerning human evaluation, which hinders large-scale evaluation, or adapt from other tasks and lose specified concerns, failing to comprehensively evaluate the modification of instruction and the preservation of irrelevant regions, resulting in biased evaluation. To tackle it, we introduce a new metric Balancing Preservation Modification (BPM), that tailored for instruction-based image editing by explicitly disentangling the image into editing-relevant and irrelevant regions for specific consideration. We first identify and locate editing-relevant regions, followed by a two-tier process to assess editing quality: Region-Aware Judge evaluates whether the position and size of the edited region align with instruction, and Semantic-Aware Judge further assesses the instruction content compliance within editing-relevant regions as well as content preservation within irrelevant regions, yielding comprehensive and interpretable quality assessment. Moreover, the editing-relevant region localization in BPM can be integrated into image editing approaches to improve the editing quality, manifesting its wild application. We verify the effectiveness of BPM metric on comprehensive instruction-editing data, and the re- sults show that we yield the highest alignment with human evaluation compared to existing metrics, indicating efficacy. The code is available at https://joyli-x.github.io/BPM/.

Social Aspects · Accountability, Transparency, and Interpretability

Weiqiu You, Helen Qu, Marco Gatti, Bhuvnesh Jain, Eric Wong

Self-attributing neural networks (SANNs) present a potential path towards interpretable models for high-dimensional problems, but often face significant trade-offs in performance. In this work, we formally prove a lower bound on errors of per-feature SANNs, whereas group-based SANNs can achieve zero error and thus high performance. Motivated by these insights, we propose Sum-of-Parts (SOP), a framework that transforms any differentiable model into a group-based SANN, where feature groups are learned end-to-end without group supervision. SOP achieves state-of-the-art performance for SANNs on vision and language tasks, and we validate that the groups are interpretable on a range of quantitative and semantic metrics. We further validate the utility of SOP explanations in model debugging and cosmological scientific discovery.

Deep Learning · Large Language Models

Xinyu Zhao, Fangcong Yin, Greg Durrett

Long-context LLMs are increasingly in demand for applications such as retrieval-augmented generation. To defray the cost of pretraining LLMs over long contexts, recent work takes an approach of synthetic context extension: fine-tuning LLMs with synthetically-generated long-context data. However, it remains unclear how and why this synthetic context extension imparts abilities for downstream long-context tasks. In this paper, we investigate fine-tuning on synthetic data for three long-context tasks that require retrieval and reasoning. We vary the realism of "needle'' concepts to be retrieved and diversity of the surrounding "haystack'' context, from using LLMs to construct synthetic documents to using templated relations and creating symbolic datasets. Although models trained on synthetic data underperform models trained on the real data, the impacts of both training settings can be understood via a shared feature of the attention computation, retrieval heads (Wu et al., 2024). The retrieval heads learned from synthetic data have high overlap with retrieval heads learned on real data. Furthermore, there is a strong correlation between the recall of heads learned and the downstream performance of a model, allowing us to interpret and predict the performance of models trained in different settings. Our results shed light on how to interpret synthetic data fine-tuning performance and how to approach creating better data for learning real-world LLM capabilities over long contexts.

Deep Learning · Generative Models and Autoencoders

Kelsey Allen, Carl Doersch, Guangyao Zhou, Mohammed Suhail, Danny Driess, Ignacio Rocco, Yulia Rubanova, Thomas Kipf, Mehdi S. M. Sajjadi, Kevin Murphy 等

A current limitation of video generative video models is that they generate plausible looking frames, but poor motion --- an issue that is not well captured by FVD and other popular methods for evaluating generated videos. Here we go beyond FVD by developing a metric which better measures plausible object interactions and motion. Our novel approach is based on auto-encoding point tracks and yields motion features that can be used to not only compare distributions of videos (as few as one generated and one ground truth, or as many as two datasets), but also for evaluating motion of single videos. We show that using point tracks instead of pixel reconstruction or action recognition features results in a metric which is markedly more sensitive to temporal distortions in synthetic data, and can predict human evaluations of temporal consistency and realism in generated videos obtained from open-source models better than a wide range of alternatives. We also show that by using a point track representation, we can spatiotemporally localize generative video inconsistencies, providing extra interpretability of generated video errors relative to prior work. An overview of the results and link to the code can be found on the project page: trajan-paper.github.io.

Applications · Health / Medicine

Alexander Capstick, Rahul G. Krishnan, Payam Barnaghi

Large language models (LLMs) acquire a breadth of information across various domains. However, their computational complexity, cost, and lack of transparency often hinder their direct application for predictive tasks where privacy and interpretability are paramount. In fields such as healthcare, biology, and finance, specialised and interpretable linear models still hold considerable value. In such domains, labelled data may be scarce or expensive to obtain. Well-specified prior distributions over model parameters can reduce the sample complexity of learning through Bayesian inference; however, eliciting expert priors can be time-consuming. We therefore introduce AutoElicit to extract knowledge from LLMs and construct priors for predictive models. We show these priors are informative and can be refined using natural language. We perform a careful study contrasting AutoElicit with in-context learning and demonstrate how to perform model selection between the two methods. We find that AutoElicit yields priors that can substantially reduce error over uninformative priors, using fewer labels, and consistently outperform in-context learning. We show that AutoElicit saves over 6 months of labelling effort when building a new predictive model for urinary tract infections from sensor recordings of people living with dementia.

Reinforcement Learning · Inverse

Zhixian Xie, Haode Zhang, Yizhe Feng, Wanxin Jin

Reward design in reinforcement learning and optimal control is challenging. Preference-based alignment addresses this by enabling agents to learn rewards from ranked trajectory pairs provided by humans. However, existing methods often struggle from poor robustness to unknown false human preferences. In this work, we propose a robust and efficient reward alignment method based on a novel and geometrically interpretable perspective: hypothesis space batched cutting. Our method iteratively refines the reward hypothesis space through “cuts” based on batches of human preferences. Within each batch, human preferences, queried based on disagreement, are grouped using a voting function to determine the appropriate cut, ensuring a bounded human query complexity. To handle unknown erroneous preferences, we introduce a conservative cutting method within each batch, preventing erroneous human preferences from making overly aggressive cuts to the hypothesis space. This guarantees provable robustness against false preferences, while eliminating the need to explicitly identify them. We evaluate our method in a model predictive control setting across diverse tasks. The results demonstrate that our framework achieves comparable or superior performance to state-of-the-art methods in error-free settings while significantly outperforming existing methods when handling a high percentage of erroneous human preferences.

Deep Learning · Graph Neural Networks

Juwei Yue, Haikuo Li, Jiawei Sheng, Xiaodong Li, Taoyu Su, Tingwen Liu, Li Guo

Graph neural networks (GNNs) leverage message passing mechanisms to learn the topological features of graph data. Traditional GNNs learns node features in a spatial domain unrelated to the topology, which can hardly ensure topological features. In this paper, we formulates message passing as a system of hyperbolic partial differential equations (hyperbolic PDEs), constituting a dynamical system that explicitly maps node representations into a particular solution space. This solution space is spanned by a set of eigenvectors describing the topological structure of graphs. Within this system, for any moment in time, a node features can be decomposed into a superposition of the basis of eigenvectors. This not only enhances the interpretability of message passing but also enables the explicit extraction of fundamental characteristics about the topological structure. Furthermore, by solving this system of hyperbolic partial differential equations, we establish a connection with spectral graph neural networks (spectral GNNs), serving as a message passing enhancement paradigm for spectral GNNs.We further introduce polynomials to approximate arbitrary filter functions. Extensive experiments demonstrate that the paradigm of hyperbolic PDEs not only exhibits strong flexibility but also significantly enhances the performance of various spectral GNNs across diverse graph tasks.

Social Aspects · Accountability, Transparency, and Interpretability

Nicholas Goldowsky-Dill, Bilal Chughtai, Stefan Heimersheim, Marius Hobbhahn

AI models might use deceptive strategies as part of scheming or misaligned behaviour. Monitoring outputs alone is insufficient, since the AI might produce seemingly benign outputs while its internal reasoning is misaligned. We thus evaluate if linear probes can robustly detect deception by monitoring model activations. We test two probe-training datasets, one with contrasting instructions to be honest or deceptive (following Zou et al. (2023)) and one of responses to simple roleplaying scenarios. We test whether these probes generalize to realistic settings where Llama-3.3-70B-Instruct behaves deceptively, such as concealing insider trading Scheurer et al. (2023) and purposely underperforming on safety evaluations Benton et al. (2024). We find that our probe distinguishes honest and deceptive responses with AUROCs between 0.96 and 0.999 on our evaluation datasets. If we set the decision threshold to have a 1\% false positive rate on chat data not related to deception, our probe catches 95-99\% of the deceptive responses. Overall we think white-box probes are promising for future monitoring systems, but current performance is insufficient as a robust defence against deception. Our probes' outputs can be viewed at https://data.apolloresearch.ai/dd/ and our code at https://github.com/ApolloResearch/deception-detection.