ICLR 2025 Papers

Fine-tuning pre-trained models has become the standard approach to endow them with specialized knowledge, but it poses fundamental challenges. In particular, (i) fine-tuning often leads to catastrophic forgetting, where improvements on a target domain degrade generalization on other tasks, and (ii) merging fine-tuned checkpoints from disparate tasks can lead to significant performance loss. To address these challenges, we introduce LiNeS, Layer-increasing Network Scaling, a post-training editing technique designed to preserve pre-trained generalization while enhancing fine-tuned task performance. LiNeS scales parameter updates linearly based on their layer depth within the network, maintaining shallow layers close to their pre-trained values to preserve general features while allowing deeper layers to retain task-specific representations. In multi-task model merging scenarios, layer-wise scaling of merged parameters reduces negative task interference. LiNeS demonstrates significant improvements in both single-task and multi-task settings across various benchmarks in vision and natural language processing. It mitigates forgetting, enhances out-of-distribution generalization, integrates seamlessly with existing multi-task model merging baselines improving their performance across benchmarks and model sizes, and can boost generalization when merging LLM policies aligned with different rewards via RLHF. Our method is simple to implement, computationally efficient and complementary to many existing techniques. Our source code is available at github.com/wang-kee/LiNeS.

As LLM-as-a-Judge emerges as a new paradigm for assessing large language models (LLMs), concerns have been raised regarding the alignment, bias, and stability of LLM evaluators. While substantial work has focused on alignment and bias, little research has concentrated on the stability of LLM evaluators. In this paper, we conduct extensive experiments involving 9 widely used LLM evaluators across 2 different evaluation settings to investigate the uncertainty in model-based LLM evaluations. We pinpoint that LLM evaluators exhibit varying uncertainty based on model families and sizes. With careful comparative analyses, we find that employing special prompting strategies, whether during inference or post-training, can alleviate evaluation uncertainty to some extent. By utilizing uncertainty to enhance LLM's reliability and detection capability in Out-Of-Distribution (OOD) data, we further fine-tune an uncertainty-aware LLM evaluator named ConfiLM using a human-annotated fine-tuning set and assess ConfiLM's OOD evaluation ability on a manually designed test set sourced from the 2024 Olympics. Experimental results demonstrate that incorporating uncertainty as additional information during the fine-tuning phase can largely improve the model's evaluation performance in OOD scenarios. The code and data are released at: https://github.com/hasakiXie123/LLM-Evaluator-Uncertainty.

Deep neural networks trained using gradient descent with a fixed learning rate $\eta$ often operate in the regime of ``edge of stability'' (EOS), where the largest eigenvalue of the Hessian equilibrates about the stability threshold $2/\eta$. In this work, we present a fine-grained analysis of the learning dynamics of (deep) linear networks (DLNs) within the deep matrix factorization loss beyond EOS. For DLNs, loss oscillations beyond EOS follow a period-doubling route to chaos. We theoretically analyze the regime of the 2-period orbit and show that the loss oscillations occur within a small subspace, with the dimension of the subspace precisely characterized by the learning rate. The crux of our analysis lies in showing that the symmetry-induced conservation law for gradient flow, defined as the balancing gap among the singular values across layers, breaks at EOS and decays monotonically to zero. Overall, our results contribute to explaining two key phenomena in deep networks: (i) shallow models and simple tasks do not always exhibit EOS; and (ii) oscillations occur within top features}. We present experiments to support our theory, along with examples demonstrating how these phenomena occur in nonlinear networks and how they differ from those which have benign landscape such as in DLNs.

Incremental learning aims to adapt to new sets of categories over time with minimal computational overhead. Prior work often addresses this task by training efficient task-specific adaptors that modify frozen layer weights or features to capture relevant information without affecting predictions on any previously learned categories. While these adaptors are generally more efficient than finetuning the entire network, they still can require tens to hundreds of thousands task-specific trainable parameters even for relatively small networks. This makes it challenging to operate in resource-constrained environments with high communication costs, such as edge devices or mobile phones. Thus, we propose Reparameterized, Compact weight Adaptation for Sequential Tasks (RECAST), a novel method that dramatically reduces the number of task-specific trainable parameters to fewer than 50 – several orders of magnitude less than competing methods like LoRA. RECAST accomplishes this efficiency by learning to decompose layer weights into a soft parameter-sharing framework consisting of a set of shared weight templates and very few module-specific scaling factors or coefficients. This soft parameter-sharing framework allows for effective task-wise reparameterization by tuning only these coefficients while keeping templates frozen. A key innovation of RECAST is the novel weight reconstruction pipeline called Neural Mimicry, which eliminates the need for pretraining from scratch. This allows for high-fidelity emulation of existing pretrained weights within our framework and provides quick adaptability to any model scale and architecture. Extensive experiments across six diverse datasets demonstrate RECAST outperforms the state-of-the-art by up to ∼ 1.5% and improves baselines by > 3% across various scales, architectures, and parameter spaces. Moreover, we show that RECAST’s architecture-agnostic nature allows for seamless integration with existing methods, further boosting performance.

The advent of Large Language Models (LLMs) has paved the way for AI search engines, e.g., SearchGPT, showcasing a new paradigm in human-internet interaction. However, most current AI search engines are limited to text-only settings, neglecting the multimodal user queries and the text-image interleaved nature of website information. Recently, Large Multimodal Models (LMMs) have made impressive strides. Yet, whether they can function as AI search engines remains under-explored, leaving the potential of LMMs in multimodal search an open question. To this end, we first design a delicate pipeline, MMSearch-Engine, to empower any LMMs with multimodal search capabilities. On top of this, we introduce MMSearch, a comprehensive evaluation benchmark to assess the multimodal search performance of LMMs. The curated dataset contains 300 manually collected instances spanning 14 subfields, which involves no overlap with the current LMMs' training data, ensuring the correct answer can only be obtained within searching. By using MMSearch-Engine, the LMMs are evaluated by performing three individual tasks (requery, rerank, and summarization), and one challenging end-to-end task with a complete searching process. We conduct extensive experiments on closed-source and open-source LMMs. Among all tested models, GPT-4o with MMSearch-Engine achieves the best results, which surpasses the commercial product, Perplexity Pro, in the end-to-end task, demonstrating the effectiveness of our proposed pipeline. We further present error analysis to unveil current LMMs still struggle to fully grasp the multimodal search tasks, and conduct ablation study to indicate the potential of scaling test-time computation for AI search engine. We hope MMSearch may provide unique insights to guide the future development of multimodal AI search engine.

Tool-Augmented Language Models (TALMs) leverage external APIs to answer user queries across various domains. However, existing benchmark datasets for TALM research often feature simplistic dialogues that do not reflect real-world scenarios, such as the need for models to ask clarifying questions or proactively call additional APIs when essential information is missing. To address these limitations, we construct and release ToolDial, a dataset comprising 11,111 multi-turn dialogues, with an average of 8.95 turns per dialogue, based on APIs from RapidAPI. ToolDial has two key characteristics. First, the dialogues incorporate 16 user and system actions (e.g., request, clarify, fail inform) to capture the rich dynamics of real-world interactions. Second, we simulate dialogues where the system requests necessary information from the user based on API documentation and seeks additional APIs if the user fails to provide the required information. To facilitate this process, we introduce a method for generating an API graph that represents input and output compatibility between APIs. Using ToolDial, we evaluate a suite of language models on their ability to predict correct actions and extract input parameter values for API calls from the dialogue history. Modern language models achieve accuracy scores below 70\%, indicating substantial room for improvement. We provide a detailed analysis of the areas where these models fall short.

Large Language Models (LLMs) exhibit impressive capabilities but require careful alignment with human preferences. Traditional training-time methods finetune LLMs using human preference datasets but incur significant training costs and require repeated training to handle diverse user preferences. Test-time alignment methods address this by using reward models (RMs) to guide frozen LLMs without retraining. However, existing test-time approaches rely on trajectory-level RMs which are designed to evaluate complete responses, making them unsuitable for autoregressive text generation that requires computing next-token rewards from partial responses. To address this, we introduce GenARM, a test-time alignment approach that leverages the Autoregressive Reward Model—a novel reward parametrization designed to predict next-token rewards for efficient and effective autoregressive generation. Theoretically, we demonstrate that this parametrization can provably guide frozen LLMs toward any distribution achievable by traditional RMs within the KL-regularized reinforcement learning framework. Experimental results show that GenARM significantly outperforms prior test-time alignment baselines and matches the performance of training-time methods. Additionally, GenARM enables efficient weak-to-strong guidance, aligning larger LLMs with smaller RMs without the high costs of training larger models. Furthermore, GenARM supports multi-objective alignment, allowing real-time trade-offs between preference dimensions and catering to diverse user preferences without retraining. Our project page is available at: https://genarm.github.io.

Designing incentives for an adapting population is a ubiquitous problem in a wide array of economic applications and beyond. In this work, we study how to design additional rewards to steer multi-agent systems towards desired policies \emph{without} prior knowledge of the agents' underlying learning dynamics. Motivated by the limitation of existing works, we consider a new and general category of learning dynamics called \emph{Markovian agents}. We introduce a model-based non-episodic Reinforcement Learning (RL) formulation for our steering problem. Importantly, we focus on learning a \emph{history-dependent} steering strategy to handle the inherent model uncertainty about the agents' learning dynamics. We introduce a novel objective function to encode the desiderata of achieving a good steering outcome with reasonable cost. Theoretically, we identify conditions for the existence of steering strategies to guide agents to the desired policies. Complementing our theoretical contributions, we provide empirical algorithms to approximately solve our objective, which effectively tackles the challenge in learning history-dependent strategies. We demonstrate the efficacy of our algorithms through empirical evaluations.

A unique aspect of human visual understanding is the ability to flexibly interpret abstract concepts: acquiring lifted rules explaining what they symbolize, grounding them across familiar and unfamiliar contexts, and making predictions or reasoning about them. While off-the-shelf vision-language models excel at making literal interpretations of images (e.g., recognizing object categories such as tree branches), they still struggle to make sense of such visual abstractions (e.g., how an arrangement of tree branches may form the walls of a maze). To address this challenge, we introduce Deep Schema Grounding (DSG), a framework that leverages explicit structured representations of visual abstractions for grounding and reasoning. At the core of DSG are schemas—dependency graph descriptions of abstract concepts that decompose them into more primitive-level symbols. DSG uses large language models to extract schemas, then hierarchically grounds concrete to abstract components of the schema onto images with vision-language models. The grounded schema is used to augment visual abstraction understanding. We systematically evaluate DSG and different methods in reasoning on our new Visual Abstractions Benchmark, which consists of diverse, real-world images of abstract concepts and corresponding question-answer pairs labeled by humans. We show that DSG significantly improves the abstract visual reasoning performance of vision-language models, and is a step toward human-aligned understanding of visual abstractions.

LLMs are an integral component of retrieval-augmented generation (RAG) systems. While many studies focus on evaluating the overall quality of end-to-end RAG systems, there is a gap in understanding the appropriateness of LLMs for the RAG task. To address this, we introduce Trust-Score, a holistic metric that evaluates the trustworthiness of LLMs within the RAG framework. Our results show that various prompting methods, such as in-context learning, fail to effectively adapt LLMs to the RAG task as measured by Trust-Score. Consequently, we propose Trust-Align, a method to align LLMs for improved Trust-Score performance. 26 out of 27 models aligned using Trust-Align substantially outperform competitive baselines on ASQA, QAMPARI, and ELI5. Specifically, in LLaMA-3-8b, Trust-Align outperforms FRONT on ASQA (↑12.56), QAMPARI (↑36.04), and ELI5 (↑17.69). Trust-Align also significantly enhances models’ ability to correctly refuse and provide quality citations. We also demonstrate the effectiveness of Trust-Align across different open-weight models, including the LLaMA series (1b to 8b), Qwen-2.5 series (0.5b to 7b), and Phi3.5 (3.8b). We release our code at https://github.com/declare-lab/trust-align.

The proteins that exist today have been optimized over billions of years of natural evolution, during which nature creates random mutations and selects them. The discovery of functionally promising mutations is challenged by the limited evolutionary accessible regions, i.e., only a small region on the fitness landscape is beneficial. There have been numerous priors used to constrain protein evolution to regions of landscapes with high-fitness variants, among which the change in binding free energy ($\Delta\Delta G$) of protein complexes upon mutations is one of the most commonly used priors. However, the huge mutation space poses two challenges: (1) how to improve the efficiency of $\Delta\Delta G$ prediction for fast mutation screening; and (2) how to explain mutation preferences and efficiently explore accessible evolutionary regions. To address these challenges, we propose a lightweight $\Delta\Delta G$ predictor (Light-DDG), which adopts a structure-aware Transformer as the backbone and enhances it by knowledge distilled from existing powerful but computationally heavy $\Delta\Delta G$ predictors. Additionally, we augmented, annotated, and released a large-scale dataset containing millions of mutation data for pre-training Light-DDG. We find that such a simple yet effective Light-DDG can serve as a good unsupervised antibody optimizer and explainer. For the target antibody, we propose a novel Mutation Explainer to learn mutation preferences, which accounts for the marginal benefit of each mutation per residue. To further explore accessible evolutionary regions, we conduct preference-guided antibody optimization and evaluate antibody candidates quickly using Light-DDG to identify desirable mutations. Extensive experiments have demonstrated the effectiveness of Light-DDG in terms of test generalizability, noise robustness, and inference practicality, e.g., 89.7$\times$ inference acceleration and 15.45\% performance gains over previous state-of-the-art baselines. A case study of SARS-CoV-2 further demonstrates the crucial role of Light-DDG for mutation explanation and antibody optimization.

This paper introduces MedTrinity-25M, a comprehensive, large-scale multimodal dataset for medicine, covering over 25 million images across 10 modalities with multigranular annotations for more than 65 diseases. These multigranular annotations encompass both global information, such as modality and organ detection, and local information like ROI analysis, lesion texture, and region-wise correlations. Unlike the existing multimodal datasets, which are limited by the availability of image-text pairs, we have developed the first automated pipeline that scales up multimodal data by generating multigranular visual and textual annotations in the form of image-ROI-description triplets without the need for any paired text descriptions. Specifically, data from over 30 different sources have been collected, preprocessed, and grounded using domain-specific expert models to identify ROIs related to abnormal regions. We then build a comprehensive knowledge base and prompt multimodal large language models to perform retrieval-augmented generation with the identified ROIs as guidance, resulting in multigranular textual descriptions. Compared to existing datasets, MedTrinity-25M provides the most enriched annotations, supporting a comprehensive range of multimodal tasks such as captioning and report generation, as well as vision-centric tasks like classification and segmentation. We propose LLaVA-Tri by pretraining LLaVA on MedTrinity-25M, achieving state-of-the-art performance on VQA-RAD, SLAKE, and PathVQA, surpassing representative SOTA multimodal large language models. Furthermore, MedTrinity-25M can also be utilized to support large-scale pre-training of multimodal medical AI models, contributing to the development of future foundation models in the medical domain. We will make our dataset available. The dataset is publicly available at https://yunfeixie233.github.io/MedTrinity-25M/.

Label differential privacy (DP) is a framework that protects the privacy of labels in training datasets, while the feature vectors are public. Existing approaches protect the privacy of labels by flipping them randomly, and then train a model to make the output approximate the privatized label. However, as the number of classes K increases, stronger randomization is needed, thus the performances of these methods become significantly worse. In this paper, we propose a vector approximation approach for learning with label local differential privacy, which is easy to implement and introduces little additional computational overhead. Instead of flipping each label into a single scalar, our method converts each label into a random vector with K components, whose expectations reflect class conditional probabilities. Intuitively, vector approximation retains more information than scalar labels. A brief theoretical analysis shows that the performance of our method only decays slightly with K. Finally, we conduct experiments on both synthesized and real datasets, which validate our theoretical analysis as well as the practical performance of our method.

Recent work has shown that scale is important for improved brain decoding, with more data leading to greater decoding accuracy. However, large-scale decoding across many different datasets is challenging because neural circuits are heterogeneous---each brain region contains a unique mix of cellular sub-types, and the responses to different stimuli are diverse across regions and sub-types. It is unknown whether it is possible to pre-train and transfer brain decoding models between distinct tasks, cellular sub-types, and brain regions. To address these questions, we developed a multi-task transformer architecture and trained it on the entirety of the Allen Institute's Brain Observatory dataset. This dataset contains responses from over 100,000 neurons in 6 areas of the brains of mice, observed with two-photon calcium imaging, recorded while the mice observed different types of visual stimuli. Our results demonstrate that transfer is indeed possible -combining data from different sources is beneficial for a number of downstream decoding tasks. As well, we can transfer the model between regions and sub-types, demonstrating that there is in fact common information in diverse circuits that can be extracted by an appropriately designed model. Interestingly, we found that the model's latent representations showed clear distinctions between different brain regions and cellular sub-types, even though it was never given any information about these distinctions. Altogether, our work demonstrates that training a large-scale neural decoding model on diverse data is possible, and this provides a means of studying the differences and similarities between heterogeneous neural circuits.

Reinforcement Learning from Human Feedback aligns the outputs of Large Language Models with human values and preferences. Central to this process is the reward model (RM), which translates human feedback into training signals for optimising LLM behaviour. However, RMs can develop biases by exploiting spurious correlations in their training data, such as favouring outputs based on length or style rather than true quality. These biases can lead to incorrect output rankings, sub-optimal model evaluations, and the amplification of undesirable behaviours in LLMs alignment. This paper addresses the challenge of correcting such biases without additional data and training, introducing the concept of Post-hoc Reward Calibration. We first propose to use local average reward to estimate the bias term and, thus, remove it to approximate the underlying true reward. We then extend the approach to a more general and robust form with the Locally Weighted Regression. Focusing on the prevalent length bias, we validate our proposed approaches across three experimental settings, demonstrating consistent improvements: (1) a 3.11 average performance gain across 33 reward models on the RewardBench dataset; (2) improved agreement of RM produced rankings with GPT-4 evaluations and human preferences based on the AlpacaEval benchmark; and (3) improved Length-Controlled win rate (Dubois et al., 2024) of the RLHF process in multiple LLM–RM combinations. According to our experiments, our method is computationally efficient and generalisable to other types of bias and RMs, offering a scalable and robust solution for mitigating biases in LLM alignment and evaluation.

Incremental Objection Detection (IOD) facilitates the expansion of the usage scope of object detectors without forgetting previously acquired knowledge. Current approaches mostly adopt response-level knowledge distillation to overcome forgetting issues, by conducting implicit memory replay from the teacher model on new training data. However, this indirect learning paradigm does not fully leverage the knowledge generated by the teacher model. In this paper, we dive deeper into the mechanism of pseudo-labeling in incremental object detection by investigating three critical problems: (a) the upper bound quality of the pseudo labels is greatly limited by the previous model, (b) fixed score thresholds for label filtering, without considering the distribution across categories, and (c) the confidence score generated by the model does not well reflect the quality of the localization. Based on these observations, we propose a simple yet effective pseudo-labeling continual object detection framework, namely PseDet. Specifically, we introduce the spatio-temporal enhancement module to alleviate the negative effects when learning noisy data from the previous model. Considering the score distribution divergence across different classes, we propose the Categorical Adaptive Label Selector with a simple mathematical prior and fast K-Means pre-computation to dynamically determine the class-wise filtering threshold. In order to align the label score with the localization quality of the pseudo labels, we project the score through non-linear mapping to calibrate the distribution and integrate it into the new-step supervision. Extensive experiments on the competitive COCO benchmarks demonstrate the effectiveness and generalization of PseDet. Notably, it achieves 43.5+/41.2+ mAP under the 1/4-step incremental settings, achieving new state-of-the-art performance.

Recent advancements in large language models (LLMs) have resulted in increasingly anthropomorphic language concerning the ability of LLMs to reason. Whether \textit{reasoning} in LLMs should be understood to be inherently different is, however, widely debated. We propose utilizing a representation engineering approach wherein model activations are read from the residual stream of an LLM when processing a reasoning task. The activations are used to derive a control vector that is applied to the model as an inference-time intervention, modulating the representational space of the model, to improve performance on the specified task. We publish the code for deriving control vectors and analyzing model representations. The method allows us to improve performance on reasoning benchmarks and assess how control vectors influence the final logit distribution of a model via metrics such as KL divergence and entropy. We apply control vectors to Mistral-7B-Instruct and a range of Pythia models on an inductive, a deductive and mathematical reasoning task. We show that an LLM can, to a certain degree, be controlled to improve its perceived reasoning ability by modulating activations. The intervention is dependent upon the ability to reliably extract the model's typical state when correctly solving a task. Our results suggest that reasoning performance can be modulated in the same manner as other information-processing tasks performed by LLMs and demonstrate that we are capable of improving performance on specific tasks via a simple intervention on the residual stream with no additional training.

The ventral, dorsal, and lateral streams in high-level human visual cortex are implicated in distinct functional processes. Yet, deep neural networks (DNNs) trained on a single task model the entire visual system surprisingly well, hinting at common computational principles across these pathways. To explore this inconsistency, we applied a novel sparse decomposition approach to identify the dominant components of visual representations within each stream. Consistent with traditional neuroscience research, we find a clear difference in component response profiles across the three visual streams—identifying components selective for faces, places, bodies, text, and food in the ventral stream; social interactions, implied motion, and hand actions in the lateral stream; and some less interpretable components in the dorsal stream. Building on this, we introduce Sparse Component Alignment (SCA), a new method for measuring representational alignment between brains and machines that better captures the latent neural tuning of these two visual systems. We find that standard visual DNNs are more aligned with ventral than either dorsal or lateral representations. SCA reveals these distinctions with greater resolution than conventional population-level geometry, offering a measure of representational alignment that is sensitive to a system’s underlying axes of neural tuning.

In this paper, we address conditional testing problems through the conformal inference framework. We define the localized conformal $p$-values by inverting prediction intervals and prove their theoretical properties. These defined $p$-values are then applied to several conditional testing problems to illustrate their practicality. Firstly, we propose a conditional outlier detection procedure to test for outliers in the conditional distribution with finite-sample false discovery rate (FDR) control. We also introduce a novel conditional label screening problem with the goal of screening multivariate response variables and propose a screening procedure to control the family-wise error rate (FWER). Finally, we consider the two-sample conditional distribution test and define a weighted U-statistic through the aggregation of localized $p$-values. Numerical simulations and real-data examples validate the superior performance of our proposed strategies.