ICLR 2025 论文列表

In-context learning, the ability to adapt based on a few examples in the input prompt, is a ubiquitous feature of large language models (LLMs). However, as LLMs' in-context learning abilities continue to improve, understanding this phenomenon mechanistically becomes increasingly important. In particular, it is not well-understood how LLMs learn to solve specific classes of problems, such as reinforcement learning (RL) problems, in-context. Through three different tasks, we first show that Llama $3$ $70$B can solve simple RL problems in-context. We then analyze the residual stream of Llama using Sparse Autoencoders (SAEs) and find representations that closely match temporal difference (TD) errors. Notably, these representations emerge despite the model only being trained to predict the next token. We verify that these representations are indeed causally involved in the computation of TD errors and $Q$-values by performing carefully designed interventions on them. Taken together, our work establishes a methodology for studying and manipulating in-context learning with SAEs, paving the way for a more mechanistic understanding.

Decentralized learning offers a promising approach to crowdsource data consumptions and computational workloads across geographically distributed compute interconnected through peer-to-peer networks, accommodating the exponentially increasing demands. However, proper incentives are still in absence, considerably discouraging participation. Our vision is that a fair incentive mechanism relies on fair attribution of contributions to participating nodes, which faces non-trivial challenges arising from the localized connections making influence ``cascade'' in a decentralized network. To overcome this, we design the first method to estimate Data Influence CascadE (DICE) in a decentralized environment. Theoretically, the framework derives tractable approximations of influence cascade over arbitrary neighbor hops, suggesting the influence cascade is determined by an interplay of data, communication topology, and the curvature of loss landscape.DICE also lays the foundations for applications including selecting suitable collaborators and identifying malicious behaviors. Project page is available at https://raiden-zhu.github.io/blog/2025/DICE.

The hypothesis of \textit{Universality} in interpretability suggests that different neural networks may converge to implement similar algorithms on similar tasks. In this work, we investigate two mainstream architectures for language modeling, namely Transformers and Mambas, to explore the extent of their mechanistic similarity. We propose to use Sparse Autoencoders (SAEs) to isolate interpretable features from these models and show that most features are similar in these two models. We also validate the correlation between feature similarity and~\univ. We then delve into the circuit-level analysis of Mamba models and find that the induction circuits in Mamba are structurally analogous to those in Transformers. We also identify a nuanced difference we call \emph{Off-by-One motif}: The information of one token is written into the SSM state in its next position. Whilst interaction between tokens in Transformers does not exhibit such trend.

Semantic text embedding is essential to many tasks in Natural Language Processing (NLP). While black-box models are capable of generating high-quality embeddings, their lack of interpretability limits their use in tasks that demand transparency. Recent approaches have improved interpretability by leveraging domain-expert-crafted or LLM-generated questions, but these methods rely heavily on expert input or well-prompt design, which restricts their generalizability and ability to generate discriminative questions across a wide range of tasks. To address these challenges, we introduce \algo{CQG-MBQA} (Contrastive Question Generation - Multi-task Binary Question Answering), a general framework for producing interpretable semantic text embeddings across diverse tasks. Our framework systematically generates highly discriminative, low cognitive load yes/no questions through the \algo{CQG} method and answers them efficiently with the \algo{MBQA} model, resulting in interpretable embeddings in a cost-effective manner. We validate the effectiveness and interpretability of \algo{CQG-MBQA} through extensive experiments and ablation studies, demonstrating that it delivers embedding quality comparable to many advanced black-box models while maintaining inherently interpretability. Additionally, \algo{CQG-MBQA} outperforms other interpretable text embedding methods across various downstream tasks. The source code is available at \url{https://github.com/dukesun99/CQG-MBQA}.

Understanding the internal computations of large language models (LLMs) is crucial for aligning them with human values and preventing undesirable behaviors like toxic content generation. However, mechanistic interpretability is hindered by *polysemanticity*—where individual neurons respond to multiple, unrelated concepts. While Sparse Autoencoders (SAEs) have attempted to disentangle these features through sparse dictionary learning, they have compromised LLM performance due to reliance on post-hoc reconstruction loss. To address this issue, we introduce **Mixture of Monosemantic Experts for Transformers (Monet)** architecture, which incorporates sparse dictionary learning directly into end-to-end Mixture-of-Experts pretraining. Our novel expert decomposition method enables scaling the expert count to 262,144 per layer while total parameters scale proportionally to the square root of the number of experts. Our analyses demonstrate mutual exclusivity of knowledge across experts and showcase the parametric knowledge encapsulated within individual experts. Moreover, **Monet** allows knowledge manipulation over domains, languages, and toxicity mitigation without degrading general performance. Our pursuit of transparent LLMs highlights the potential of scaling expert counts to enhance mechanistic interpretability and directly resect the internal knowledge to fundamentally adjust model behavior.

Disentangling model activations into human-interpretable features is a central problem in interpretability. Sparse autoencoders (SAEs) have recently attracted much attention as a scalable unsupervised approach to this problem. However, our imprecise understanding of ground-truth features in realistic scenarios makes it difficult to measure the success of SAEs. To address this challenge, we propose to evaluate SAEs on specific tasks by comparing them to supervised feature dictionaries computed with knowledge of the concepts relevant to the task. Specifically, we suggest that it is possible to (1) compute supervised sparse feature dictionaries that disentangle model computations for a specific task; (2) use them to evaluate and contextualize the degree of disentanglement and control offered by SAE latents on this task. Importantly, we can do this in a way that is agnostic to whether the SAEs have learned the exact ground-truth features or a different but similarly useful representation. As a case study, we apply this framework to the indirect object identification (IOI) task using GPT-2 Small, with SAEs trained on either the IOI or OpenWebText datasets. We find that SAEs capture interpretable features for the IOI task, and that more recent SAE variants such as Gated SAEs and Top-K SAEs are competitive with supervised features in terms of disentanglement and control over the model. We also exhibit, through this setup and toy models, some qualitative phenomena in SAE training illustrating feature splitting and the role of feature magnitudes in solutions preferred by SAEs.

Understanding the loss landscapes of neural networks (NNs) is critical for optimizing model performance. Previous research has identified the phenomenon of mode connectivity on curves, where two well-trained NNs can be connected by a continuous path in parameter space where the path maintains nearly constant loss. In this work, we extend the concept of mode connectivity to explore connectivity on surfaces, significantly broadening its applicability and unlocking new opportunities. While initial attempts to connect models via linear surfaces in parameter space were unsuccessful, we propose a novel optimization technique that consistently discovers Bézier surfaces with low-loss and high-accuracy connecting multiple NNs in a nonlinear manner. We further demonstrate that even without optimization, mode connectivity exists in certain cases of Bézier surfaces, where the models are carefully selected and combined linearly. This approach provides a deeper and more comprehensive understanding of the loss landscape and offers a novel way to identify models with enhanced performance for model averaging and output ensembling. We demonstrate the effectiveness of our method on CIFAR-10, CIFAR-100, and Tiny-ImageNet datasets using VGG16, ResNet18, and ViT architectures.

Comprehensible neural network explanations are foundations for a better understanding of decisions, especially when the input data are infused with malicious perturbations. Existing solutions generally mitigate the impact of perturbations through adversarial training, yet they fail to generate comprehensible explanations under unknown perturbations. To address this challenge, we propose AGAIN, a factor graph-based interpretable neural network, which is capable of generating comprehensible explanations under unknown perturbations. Instead of retraining like previous solutions, the proposed AGAIN directly integrates logical rules by which logical errors in explanations are identified and rectified during inference. Specifically, we construct the factor graph to express logical rules between explanations and categories. By treating logical rules as exogenous knowledge, AGAIN can identify incomprehensible explanations that violate real-world logic. Furthermore, we propose an interactive intervention switch strategy rectifying explanations based on the logical guidance from the factor graph without learning perturbations, which overcomes the inherent limitation of adversarial training-based methods in defending only against known perturbations. Additionally, we theoretically demonstrate the effectiveness of employing factor graph by proving that the comprehensibility of explanations is strongly correlated with factor graph. Extensive experiments are conducted on three datasets and experimental results illustrate the superior performance of AGAIN compared to state-of-the-art baselines.

Language models (LMs) are susceptible to bias, sycophancy, backdoors, and other tendencies that lead to unfaithful responses to the input context. Interpreting internal states of LMs could help monitor and correct unfaithful behavior. We hypothesize that LMs faithfully represent their input contexts in a latent world model, and we seek to extract these latent world states as logical propositions. For example, given the input context ``Greg is a nurse. Laura is a physicist.'', we aim to decode the propositions WorksAs(Greg, nurse) and WorksAs(Laura, physicist) from the model's internal activations. To do so we introduce _propositional probes_, which compositionally extract lexical concepts from token activations and bind them into propositions. Key to this is identifying a _binding subspace_ in which bound tokens have high similarity (Greg $\leftrightarrow$ nurse) but unbound ones do not (Greg $\not\leftrightarrow$ physicist). Despite only being trained on linguistically simple English templates, we find that propositional probes generalize to inputs written as short stories and translated to Spanish. Moreover, in three settings where LMs respond unfaithfully to the input context---prompt injections, backdoor attacks, and gender bias--- the decoded propositions remain faithful. This suggests that LMs often encode a faithful world model but decode it unfaithfully, which motivates the search for better interpretability tools for monitoring LMs.

Machine learning is increasingly used to select which individuals receive limited-resource interventions in domains such as human services, education, development, and more. However, it is often not apparent what the right quantity is for models to predict. In particular, policymakers rarely have access to data from a randomized controlled trial (RCT) that would enable accurate estimates of treatment effects -- which individuals would benefit more from the intervention. Observational data is more likely to be available, creating a substantial risk of bias in treatment effect estimates. Practitioners instead commonly use a technique termed "risk-based targeting" where the model is just used to predict each individual's status quo outcome (an easier, non-causal task). Those with higher predicted risk are offered treatment. There is currently almost no empirical evidence to inform which choices lead to the most effect machine learning-informed targeting strategies in social domains. In this work, we use data from 5 real-world RCTs in a variety of domains to empirically assess such choices. We find that when treatment effects can be estimated reliably (which we simulate by using direct outcome observations), treatment effect based targeting substantially outperforms risk-based targeting, even when treatment effect estimates are biased. Moreover, these results hold even when the policymaker has strong normative preferences for assisting higher-risk individuals. However, when treatment effects must be predicted from features alone (as is always the case in practice), performance can degrade significantly due to limited data making it difficult to learn accurate mappings from features to treatment effects. Our results suggest treatment effect targeting has significant potential benefits, but realizing these benefits requires careful attention to model training and validation.