United States

The capability of the reward model (RM) is crucial for the success of Reinforcement Learning from Human Feedback (RLHF) in aligning with human preferences. However, as training progresses, the output space distribution of the policy model shifts. The RM, initially trained on responses sampled from the output distribution of the early policy model, gradually loses its ability to distinguish between responses from the newly shifted distribution. This issue is further compounded when the RM, trained on a specific data distribution, struggles to generalize to examples outside of that distribution. These two issues can be united as a challenge posed by the shifted distribution of the environment. To surmount this challenge, we introduce MetaRM, a novel method leveraging meta-learning to adapt the RM to the shifted environment distribution. MetaRM optimizes the RM in an alternating way, by preserving both the preferences of the original preference pairs, as well as maximizing discrimination power over new examples of the shifted distribution. Extensive experiments demonstrate that MetaRM can iteratively enhance the performance of human preference alignment by improving the RM&#39;s capacity to identify subtle differences in samples of shifted distributions.

AAAI 2025

Alleviating Shifted Distribution in Human Preference Alignment through Meta-Learning

snlp

language models

### Welcome to The 63rd Annual Meeting of the Association for Computational Linguistics (ACL 2025)

Message from the General Chair: 
*It is my great pleasure and honor to welcome you to the 63rd Annual Meeting of the Association for Computational Linguistics (ACL 2025), held in beautiful Vienna, Austria, from July 27 to August 1, 2025. ACL2025continues our field’s tradition of excellence in scholarship, innovation, and inclusivity, and I am deeply grateful to the many volunteers who have worked tirelessly to bring this event to life.* 
[Read more](https://drive.google.com/file/d/1GI_hvOpjswAuYdUTromfeDiPpCcqidwg/view?usp=sharing)

To access this event page, you need to log in with the **email address you registered with**. Access credentials will be sent to your email from Underline - subject line "Welcome to ACL 2025". Please be sure to check your spam email folder if you do not see an email confirmation right away.

Please log in

To access this event page, you need to log in with the **email address you registered with**. 

ACL 2025

Welcome to The 63rd Annual Meeting of the Association for Computational Linguistics (ACL 2025)

The capability of the reward model (RM) is crucial for the success of Reinforcement Learning from Human Feedback (RLHF) in aligning with human preferences. However, as training progresses, the output space distribution of the policy model shifts. The RM, initially trained on responses sampled from the output distribution of the early policy model, gradually loses its ability to distinguish between responses from the newly shifted distribution. This issue is further compounded when the RM, trained on a specific data distribution, struggles to generalize to examples outside of that distribution. These two issues can be united as a challenge posed by the shifted distribution of the environment. To surmount this challenge, we introduce MetaRM, a novel method leveraging meta-learning to adapt the RM to the shifted environment distribution. MetaRM optimizes the RM in an alternating way, by preserving both the preferences of the original preference pairs, as well as maximizing discrimination power over new examples of the shifted distribution. Extensive experiments demonstrate that MetaRM can iteratively enhance the performance of human preference alignment by improving the RM's capacity to identify subtle differences in samples of shifted distributions.

poster

We are pleased to announce the Thirty-Ninth AAAI Conference on Artificial Intelligence (AAAI-25), which will be held in Philadelphia, Pennsylvania at the Pennsylvania Convention Center from February 25 to March 4, 2025.

The purpose of the AAAI conference series is to promote research in Artificial Intelligence (AI) and foster scientific exchange between researchers, practitioners, scientists, students, and engineers across the entirety of AI and its affiliated disciplines. AAAI-25 will feature technical paper presentations, special tracks, invited speakers, workshops, tutorials, poster sessions, senior member presentations, competitions, and exhibit programs, and a range of other activities to be announced.

### [Invited Speakers](https://aaai.org/conference/aaai/aaai-25/aaai-25-invited-speakers/)

Register [here](https://aaai.org/conference/aaai/aaai-25/registration/)

The purpose of the AAAI conference series is to promote research in Artificial Intelligence (AI) and foster scientific exchange between researchers, practitioners, scientists, students, and engineers across the entirety of AI and its affiliated disciplines. AAAI-25 will feature technical paper presentations, special tracks, invited speakers, workshops, tutorials, poster sessions, senior member presentations, competitions, and exhibit programs, and a range of other activities to be announced.



While transformer models have been highly successful, they are computationally inefficient. We observe that for each layer, the full width of the layer may be needed only for a small subset of tokens inside a batch and that the "effective" width needed to process a token can vary from layer to layer. Motivated by this observation, we introduce the Adaptive Computation Module (ACM), a generic module that dynamically adapts its computational load to match the estimated difficulty of the input on a per-token basis. An ACM consists of a sequence of learners that progressively refine the output of their preceding counterparts. An additional gating mechanism determines the optimal number of learners to execute for each token. We also propose a distillation technique to replace any pre-trained model with an "ACMized" variant. Our evaluation of transformer models in computer vision and speech recognition demonstrates that substituting layers with ACMs significantly reduces inference costs without degrading the downstream accuracy for a wide interval of user-defined budgets.

Adaptive Computation Modules: Granular Conditional Computation for Efficient Inference

Accurate segmentation of guidewires in interventional cardiac fluoroscopy videos is crucial for computer-aided navigation tasks. While deep learning methods have demonstrated high accuracy and robustness in wire segmentation, they require substantial annotated datasets for generalizability, underscoring the need for extensive labeled data to enhance model performance. To address this challenge, we propose the Segmentation-guided Frame-consistency Video Diffusion Model (SF-VD) to generate large collections of labeled fluoroscopy videos, augmenting the training data for wire segmentation networks. SF-VD leverages videos with limited annotations by independently modeling scene distribution and motion distribution. It first samples the scene distribution by generating 2D fluoroscopy images with wires positioned according to a specified input mask, and then samples the motion distribution by progressively generating subsequent frames, ensuring frame-to-frame coherence through a frame-consistency strategy. A segmentation-guided mechanism further refines the process by adjusting wire contrast, enhancing the range of their visibility. Evaluation on a fluoroscopy dataset confirms the superior quality of the generated videos and shows significant improvements in guidewire segmentation.

Label-Efficient Data Augmentation with Video Diffusion Models for Guidewire Segmentation in Cardiac Fluoroscopy

Accurately gauging the confidence level of Large Language Models' (LLMs) predictions is pivotal for their reliable application. However, LLMs are often uncalibrated inherently and elude conventional calibration techniques due to their proprietary nature and massive scale.
In this work, we derive model spirit from the distribution of multiple randomly sampled generations, using three measures of consistency. We extensively evaluate eleven open and closed-source models on nine reasoning datasets. Results show that consistency-based calibration methods outperform existing post-hoc approaches in terms of calibration error. Meanwhile, we find that factors such as intermediate explanations, model scaling, and larger sample sizes enhance calibration, while instruction-tuning makes calibration more difficult. Moreover, confidence scores obtained from consistency can potentially enhance model performance. Finally, we offer guidance on choosing suitable consistency metrics for calibration, tailored to model characteristics such as the exposure to instruction-tuning and RLHF.

Calibrating Large Language Models with Sample Consistency

Model knowledge editing has become a widely researched topic because it enables efficient and rapid injection of new knowledge into language models or the correction of erroneous or outdated knowledge. Existing model knowledge editing methods typically categorized into single-instance sequential editing and massive one-time editing. However, in practical applications, the batched and iterative editing manner better aligns with model updating patterns. In this work, we explored the performance of parameter-update-based models in a new batched iterative editing benchmark. Our findings show that with an increase in the number of editing iterations, the accumulation of updated parameters leads to a greater change in the distribution of model parameters, making it more challenging to maintain editing performance and model stability. To address this degradation issue, we propose two methods: the Wasserstein distance constraint and update parameter sparsification, where the Wasserstein distance constraint optimizes the transition of parameter distribution before and after the editing, and update parameter sparsification significantly reduces the number of update parameters, thereby alleviating the issue of instability in the parameter distribution caused by the accumulation of update parameters through iterations. Our methods can be generally applied to different parameter-update-based knowledge editing models. Experiments on the zsRE and CounterFact datasets demonstrate that our methods can improve editing performance and enhance the later-stage stability of batched iterative editing across different models.

Wasserstein Distance Constraint and Parameter Sparsification for Batched and Iterative Knowledge Editing

Test Time Adaptation (TTA) addresses the problem of distribution shift by adapting a pretrained model to a new domain during inference. When faced with challenging shifts, most methods collapse and perform worse than the original pretrained model. In this paper, we find that not all layers are equally receptive to the adaptation, and the layers with the most misaligned gradients often cause performance degradation. To address this, we propose GALA, a novel layer selection criterion to identify the most beneficial updates to perform during test time adaptation. This criterion can also filter out unreliable samples with noisy gradients. Its simplicity allows seamless integration with existing TTA loss functions, thereby preventing degradation and focusing adaptation on the most trainable layers. This approach also helps to regularize adaptation to preserve the pretrained features, which are crucial for handling unseen domains. Through extensive experiments, we demonstrate that the proposed layer selection framework improves the performance of existing TTA approaches across multiple datasets, domain shifts, model architectures, and TTA losses.

A Layer Selection Approach to Test Time Adaptation

This study addresses the challenge of detecting anomalies in multivariate time series data. Considering a bag (e.g., multi-sensor data) consisting of two-dimensional spaces of time points and multivariate instances (e.g., individual sensors), we aim to detect anomalies at both the bag and instance level with a unified model. To circumvent the practical difficulties of labeling at the instance level in such spaces, we adopt a multiple instance learning (MIL)-based approach, which enables learning at both the bag- and instance- levels using only the bag-level labels. In this study, we introduce time-aware and instance-learnable MIL (simply, TAIL-MIL). We propose two specialized attention mechanisms designed to effectively capture the relationships between different types of instances. We innovatively integrate these attention mechanisms with conjunctive pooling applied to the two-dimensional structure at different levels (i.e., bag- and instance-level), enabling TAIL-MIL to effectively pinpoint both the timing and causative multivariate factors of anomalies. We provide theoretical evidence demonstrating TAIL-MIL's efficacy in detecting instances with two-dimensional structures. Furthermore, we empirically validate the superior performance of TAIL-MIL over the state-of-the-art MIL methods and multivariate time-series anomaly detection methods.

TAIL-MIL: Time-Aware and Instance-Learnable Multiple Instance Learning for Multivariate Time Series Anomaly Detection

The reward signal plays a central role in defining the desired behaviors of agents in reinforcement learning (RL). Rewards collected from realistic environments could be perturbed, corrupted, or noisy due to an adversary, sensor error, or because they come from subjective human feedback. Thus, it is important to construct agents that can learn under such rewards. Existing methodologies for this problem make strong assumptions, including that the perturbation is known in advance, clean rewards are accessible, or that the perturbation preserves the optimal policy. We study a new, more general, class of unknown perturbations, and introduce a distributional reward critic framework for estimating reward distributions and perturbations during training. Our proposed methods are compatible with any RL algorithm. Despite their increased generality, we show that they achieve comparable or better rewards than existing methods in a variety of environments, including those with clean rewards. Under the challenging and generalized perturbations we study, we win/tie the highest return in 44/48 tested settings (compared to 11/48 for the best baseline). Our results broaden and deepen our ability to perform RL in reward-perturbed environments.

The Distributional Reward Critic Framework for Reinforcement Learning Under Perturbed Rewards

Causal learning relies on the computationally demanding task of estimating causal graphs. In this paper, a new divide-and-conquer approach called DCILP is proposed for causal graph learning. The divide phase proceeds by identifying the Markov blanket MB$(X_i)$ of each variable $X_i$, and then addressing simultaneously the subproblems restricted on each MB$(X_i)$. Each subproblem benefits from a more favorable ratio between the number of data samples and the number of variables considered; however, it is adversely affected by the presence of hidden confounders (as variables external to MB$(X_i)$ might influence the variables inside). 
The novelty of DCILP lies in the conquer phase, which tackles the problem of aggregating the local causal graphs from the divide phase. Such an aggregation is a challenging combinatorial optimization problem especially in large-scale applications. We show that this aggregation can be formulated as an integer linear programming (ILP) problem, which is delegated to an ILP solver. Through experiments and comparisons with state-of-the-art methods, the proposed approach demonstrates comparable or improved learning accuracy while achieving significant improvements in terms of scalability in the graph size.

DCILP: A Distributed Approach for Large-Scale Causal Structure Learning

Time series forecasting requires reliable uncertainty estimates. Gaussian process regression provides a powerful framework for modelling this in a probabilistic fashion. However, its application to large time series is challenging, due to its cubic time complexity and quadratic memory requirement. In this work, we present KernelMatmul, a novel method that accelerates Gaussian process inference and thus facilitates scaling of Gaussian process regression to large, irregularly sampled and multi-output time series. Leveraging conjugate gradients in combination with sparsity approximation, KernelMatmul achieves time and memory complexity linear in the number of samples. We thoroughly benchmark our new method against multiple baselines to demonstrate its benefits and limitations, both in efficiency and accuracy.

KernelMatmul: Scaling Gaussian Processes to Large Time Series

We present and release MIDI-GPT, a generative system based on the Transformer architecture that is designed for computer-assisted music composition workflows. MIDI-GPT supports the infilling of musical material at the track and bar level, and can condition generation on particular attributes including: instrument type, musical style, note density, polyphony level, and note duration. In order to integrate these features, we employ an alternative representation for musical material, creating a time-ordered sequence of musical events for each track and concatenating several tracks into a single sequence, rather than using a single time-ordered sequence where the musical events corresponding to different tracks are interleaved. We also propose a variation of our representation allowing for expressiveness. We present experimental results that demonstrate that MIDI-GPT is able to consistently avoid duplicating the musical material it was trained on, generate music that is stylistically similar to the training dataset, and that attribute controls allow enforcing various constraints on the generated material. We also outline several real-world applications of MIDI-GPT, including collaborations with industry partners that explore integrating and evaluating MIDI-GPT, into real-world products, as well as several artistic works produced using it.

Downloads

Next from AAAI 2025

Adaptive Computation Modules: Granular Conditional Computation for Efficient Inference

Similar lecture

Mutual-Taught for Co-adapting Policy and Reward Models

.css-70qvj9{display:-webkit-box;display:-webkit-flex;display:-ms-flexbox;display:flex;-webkit-align-items:center;-webkit-box-align:center;-ms-flex-align:center;align-items:center;}Downloads

Next from AAAI 2025

Adaptive Computation Modules: Granular Conditional Computation for Efficient Inference

Similar lecture

Mutual-Taught for Co-adapting Policy and Reward Models

Downloads