Singapore

Large language models (LLMs) increasingly rely on reinforcement learning (RL) to enhance their reasoning capabilities through feedback. A critical challenge is verifying the consistency of model-generated responses and reference answers, since these responses are often lengthy, diverse, and nuanced. Rule-based verifiers struggle with complexity, prompting the use of model-based verifiers. Existing research primarily focuses on building better verifiers, yet a systematic evaluation of different types of verifiers&#39; performance across domains remains lacking, severely constraining the reliable development of Reinforcement Learning with Verifiable Reward (RLVR). To address this, we propose VerifyBench--a cross-domain comprehensive benchmark for systematically evaluating verifiers. We construct about 4,000 expert-level questions covering mathematics, physics, chemistry, and biology. Questions are equipped with reference answers and diverse responses. The reliability of the evaluation is ensured through a rigorous collection and annotation process conducted by a multidisciplinary expert team. We design a four-dimensional experimental framework to comprehensively compare the performance boundaries of specialized verifiers and general LLMs under combined conditions of extracted answers vs. complete responses, and short vs. long outputs. Our evaluation uncovers fundamental trade-offs in verifiers: while specialized verifiers achieve leading accuracy (the best model reaching 96.48\% in chemistry), they exhibit deficiencies in recall; general models show stronger inclusivity but unstable accuracy. More importantly, we discover verifiers&#39; high sensitivity to input structure and inherent limitations in cross-domain generalization, providing critical insights into the bottlenecks of current verifier technology.

AAAI 2026

VerifyBench: A Systematic Benchmark for Evaluating Reasoning Verifiers Across Domains

reinforcement learning with verifiable reward

verifier evaluation

large language models

It is my great pleasure and honor to welcome you to the 64th Annual Meeting of the Association for Computational Linguistics (ACL 2026), held in San Diego, California, USA, July 2–7, 2026. ACL 2026 continues our fields 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.

It is exciting to see the many developments in the field reflected at our conference, such as the advancement of large language models (born out of our community) into many application directions. This is not without challenges, partly due to unsolved problems, and partly due to misuse of the technology. All this is reflected well at ACL 2026. The conference also reflects that the field is not only about advancing technology, but also the science of fundamentally understanding language and how it is use by many different speaker communities. 
[Full Message from General Chair Philipp Koehn](https://drive.google.com/file/d/1QiJFOmMjvDgsDo3i1XzlQnImxe3RU_Cy/view?usp=sharing) 
<html><button style="display: inline-flex; align-items: center; justify-content: center; white-space: nowrap; border-radius: 9999px; font-weight: bold; background: #7c3aed; color: white; font-family: 'Space Grotesk', sans-serif; height: 40px; font-size: 16px; padding: 0 20px; border: none; cursor: pointer" onclick="window.open('https://drive.google.com/file/d/1QiJFOmMjvDgsDo3i1XzlQnImxe3RU_Cy/view?usp=sharing')">Conference Handbook</button></html>

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 2026". Please be sure to check your spam email folder if you do not see an email confirmation right away.

To access this event page, you are required to register.
Please complete your registration to continue.

We recommend reading [**the registration information**](https://2026.aclweb.org/registration/) first.

**Online Registration Form**: https://acl.swoogo.com/acl2026

Registration Required

ACL 2026

The 64th Annual Meeting of the Association for Computational Linguistics (ACL 2026) will take place in San Diego, California, from July 2 - 7, 2026.

Large language models (LLMs) increasingly rely on reinforcement learning (RL) to enhance their reasoning capabilities through feedback. A critical challenge is verifying the consistency of model-generated responses and reference answers, since these responses are often lengthy, diverse, and nuanced. Rule-based verifiers struggle with complexity, prompting the use of model-based verifiers. Existing research primarily focuses on building better verifiers, yet a systematic evaluation of different types of verifiers' performance across domains remains lacking, severely constraining the reliable development of Reinforcement Learning with Verifiable Reward (RLVR). To address this, we propose VerifyBench--a cross-domain comprehensive benchmark for systematically evaluating verifiers. We construct about 4,000 expert-level questions covering mathematics, physics, chemistry, and biology. Questions are equipped with reference answers and diverse responses. The reliability of the evaluation is ensured through a rigorous collection and annotation process conducted by a multidisciplinary expert team. We design a four-dimensional experimental framework to comprehensively compare the performance boundaries of specialized verifiers and general LLMs under combined conditions of extracted answers vs. complete responses, and short vs. long outputs. Our evaluation uncovers fundamental trade-offs in verifiers: while specialized verifiers achieve leading accuracy (the best model reaching 96.48\% in chemistry), they exhibit deficiencies in recall; general models show stronger inclusivity but unstable accuracy. More importantly, we discover verifiers' high sensitivity to input structure and inherent limitations in cross-domain generalization, providing critical insights into the bottlenecks of current verifier technology.

technical paper

We are pleased to announce the Fortieth AAAI Conference on Artificial Intelligence (AAAI-26), which will be held in Singapore EXPO from January 20 to January 27, 2026.

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-26 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.

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 AAAI 2026". 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 are required to register.
Please complete your registration to continue.

We recommend reading [**the registration information**](https://aaai.org/conference/aaai/aaai-26/registration/) first.

**Online Registration Form**: https://aaai.getregistered.net/conference-2026 

We are pleased to announce the Fortieth AAAI Conference on Artificial Intelligence (AAAI-26), which will be held in Singapore EXPO from January 20 to January 27, 2026.

Unsupervised representation learning on hypergraphs has recently drawn increasing attention due to its ability to capture high-order relationships without requiring labeled data. However, existing hypergraph contrastive learning methods predominantly follow spatial-based paradigms that rely on message-passing frameworks, which largely emphasize low-pass filtering. This restricts their ability to adapt to the diverse spectral characteristics of real-world hypergraphs. Motivated by the observation that different hypergraph datasets exhibit varied frequency energy distributions, we propose **HyperAim**, a novel contrastive learning framework that incorporates adaptive multi-frequency filtering into hypergraph representation learning. HyperAim integrates three complementary channels: a low-pass spatial channel, a high-pass spatial channel, and a spectral channel based on framelet transforms that jointly capture multi-frequency components. To fully exploit these diverse views, we introduce a frequency-aware contrastive learning strategy that constructs perturbed views via spectral and structural augmentations and enforces consistency across representations through inter- and intra-channel objectives. Extensive experiments on multiple benchmark datasets demonstrate that **HyperAim** consistently outperforms state-of-the-art baselines. Ablation studies further verify the effectiveness of adaptive frequency decomposition and frequency-aware contrastive learning in enhancing hypergraph representations.

HyperAim: Hypergraph Contrastive Learning with Adaptive Multi-frequency Filters

Pseudo-Boolean optimization (PBO) problem involves optimizing a linear objective function under linear inequality constraints defined over Boolean variables. PBO is widely used for modeling many combinational optimization problems, particularly in some real-world scenarios. 
In core-guided CDCL-based exact solvers, the way branching variables are assigned, known as phase selection, significantly affects the solving efficiency. 
This paper introduces two strategies to enhance solver performance by improving phase selection. 
Firstly, we design a new phase selection strategy that actively guides variables in the objective function toward assignments closer to the optimal solution. 
Secondly, to prevent the solver from becoming trapped in local solutions, we propose a reinforcement learning-based rephase mechanism that dynamically updates and resets variable phases, increasing search diversity and encouraging exploration of high-quality solution spaces.
We integrate two phase selection strategies into two state-of-the-art PBO solvers and compare them against top-performing solvers from the PB Competition 2024. The evaluation is conducted on benchmarks from the PB Competition 2016 and 2024. Experimental results show that our solvers outperform the PB Competition 2024 winning solver.

Improving Exact Algorithm for Pseudo Boolean Optimization with Two New Phase Selection Heuristics

Prompt tuning has shown promise for continual visual question answering (CVQA), facilitating modular and transferable knowledge across tasks. However, existing approaches often overlook the guiding role of prompts in the model’s implicit reasoning process. This oversight can lead to inconsistent reasoning paths and performance degradation across tasks. To address this issue, we propose the E Logic Prompt framework, which employs energy-based models (EBMs) to model the semantic compatibility between prompts and queries. In this framework, prompts function not only as adapters but also as reasoning guides that help maintain coherence throughout the inference process.
The framework enforces logical consistency at three levels. At the input level, it selects semantically aligned prompts by minimizing the energy between queries and prompts. Within the model, it aligns intermediate representations with prompts across layers to preserve step-by-step reasoning. Across tasks, it applies energy-based constraints to regulate prompt behavior, effectively suppressing semantic drift and enabling prompt reuse. These three levels of consistency together enhance the guiding capacity of prompts, allowing them to steer the model toward more stable and coherent reasoning. Extensive experiments show that E Logic Prompt outperforms existing methods in both accuracy and knowledge retention, while effectively maintaining balanced cross-modal reasoning throughout continual learning.

E-Logic Prompt: Unified Energy-Logic Framework for Continual Visual Question Answering

This paper tackles the challenging task of achieving storage-efficient yet high-fidelity motion representation in large-scale dynamic 3D Gaussian Splatting. Our motivation stems from the truth that existing urban-scale methods, which rely on massive and unstructured individual Gaussians for scene modeling, face a critical scalability bottleneck. Inspired by recent advances in the 3DGS-based compression beyond autonomous driving, we address this challenge by leveraging the compression capability of anchor-driven methods (Lu et al. 2024; Chen et al. 2024a). However, this is non-trivial as our exploratory experiments reveal that the direct application of this paradigm to dynamic, large-scale urban scenes results in performance degradation. We attribute this phenomenon to the hierarchical anchor design that severely loses dynamic information. To this end, we propose Hierarchical Dynamic Gaussian Splatting (HDGS), a novel framework designed to adapt the anchor-based Gaussian paradigm to 4D urban environments. We first establish a local support network to reinforce inter-anchor consistency, mitigating geometric and appearance fractures caused by supervision attenuation in deep hierarchies. Then, we handle heterogeneous object motion via coarse-to-fine decomposition, where high-level anchors model coarse dynamics and low-level anchors refine them with residual deformations. Third, we introduce a hybrid supervision scheme that synergistically fuses global geometric constraints and local pixel-level cues to alleviate geometrically inconsistent reconstruction under sparse LiDAR. Extensive experiments show that HDGS reduces storage by 69.0\% while maintaining or even improving rendering fidelity compared to state-of-the-art methods. Code will be released.

HDGS: Hierarchical Dynamic Gaussian Splatting for Urban Driving Scenes

Steering Vector (SV) is a powerful technique for controlling Large Language Models (LLMs) by manipulating their activations without altering model weights. However, when constructed from sensitive data, SV poses significant privacy risks, as it may leak private information. Existing differential privacy (DP) techniques for constructing SV cannot be directly applied to training-based SV construction paradigms, which offer higher task performance.
In this work, we present **PrivSV**, a general privacy-preserving approach for constructing SV with DP guarantees, compatible with arbitrary SV construction paradigms while maintaining high utility. In PrivSV, we propose three novel methods: a Layer-wise Noise-Resilient Reduction (LNR²) method to reduce the injected noise in high-dimensional SV; a Directional Prior Compensation (DPC) method to recover utility degraded by noise perturbation; and a Privacy-Aware Optimal Parameter Determination (POPD) method to adaptively maximize the performance of the final compensated SV. 
Extensive experiments on open-source LLMs of different families (i.e., LlaMa, Qwen, Mistral and Gemma) demonstrate that PrivSV outperforms several existing techniques across various privacy budgets.

PrivSV: Differentially Private Steering Vector for Large Language Models

Diffusion Large Language Models (dLLMs) have recently emerged as a competitive non-autoregressive paradigm due to their unique training and inference approach. However, there is currently a lack of safety study on this novel architecture. In this paper, we present the first analysis of dLLMs' safety performance and propose a novel safety alignment method tailored to their unique generation characteristics. Specifically, we identify a critical asymmetry between the defender and attacker in terms of security. For the defender, we reveal that the middle tokens of the response, rather than the initial ones, are more critical to the overall safety of dLLM outputs; this seems to suggest that aligning middle tokens can be more beneficial to the defender. The attacker, on the contrary, may have limited power to manipulate middle tokens, as we find dLLMs have a strong tendency towards a sequential generation order in practice, forcing the attack to meet this distribution and diverting it from influencing the critical middle tokens. Building on this asymmetry, we introduce Middle-tOken Safety Alignment (MOSA), a novel method that directly aligns the model's middle generation with safe refusals exploiting reinforcement learning. We implement MOSA and compare its security performance against eight attack methods on two benchmarks. We also test the utility of MOSA-aligned dLLM on coding, math, and general reasoning. The results strongly prove the superiority of MOSA.

Where to Start Alignment? Diffusion Large Language Model May Demand a Distinct Position

We revisit the setting of fair allocation of indivisible items among agents with heterogeneous, non-monotone valuations. We explore the existence and efficient computation of allocations that approximately satisfy either envy-freeness or equity constraints. Approximate envy-freeness ensures that each agent values her bundle at least as much as those given to the others, after some (or any) item removal, while approximate equity guarantees roughly equal valuations among agents, under similar adjustments.
As a key technical contribution of this work, by leveraging fixed-point theorems (such as Sperner's Lemma and its variants), we establish the existence of *envy-free-up-to-one-good-and-one-chore* (EF1$^c_g$) and *equitable-up-to-one-good-and-one-chore* (EQ1$^c_g$) allocations, for non-monotone valuations that are always either non-negative or non-positive. These notions represent slight relaxations of the well-studied *envy-free-up-to-one-item* (EF1) and *equitable-up-to-one-item* (EQ1) guarantees, respectively.
Our existential results hold even when items are arranged in a path and bundles must form connected sub-paths. The case of non-positive valuations, in particular, has been solved by proving a novel multi-colouring variant of Sperner's Lemma that constitutes a combinatorial result of independent interest. In addition, we also design a polynomial-time dynamic programming algorithm that computes an EQ1$^c_g$ allocation. For monotone non-increasing valuations and path-connected bundles, all the above results can be extended to EF1 and EQ1 guarantees as well. 
Finally, we focus on the problem of finding *equitable-up-to-any-good-or-any-chore* (EQX$^c_g$) allocations, which relax the notion of *equitable-up-to-any-item* (EQX) guarantee and strengthen that of EQ1. For objective valuations, where items can be partitioned into either goods or chores, we show that such allocations always exist and can be efficiently computed.

Approximately Envy-free and Equitable Allocations of Indivisible Items for Non-monotone Valuations

Multi-modal image matching is a fundamental task in multi-view and multi-modal image processing. Its key challenge lies in extracting features that remain consistent despite drastic appearance variations across modalities. However, the learning of the feature is hindered by the scarcity and the inaccurate alignment of existing multi-modal datasets. To address this, we propose a knowledge distillation framework that transfers rich prior knowledge from large-scale unimodal tasks to enhance multi-modal representation learning. Specifically, semantic priors from a vision foundation model guide the feature extractor to identify shared semantic structures across modalities, enabling better generalization under large appearance gaps. In parallel, geometric priors derived from accurately aligned visible-light datasets improve detection precision on noisy aligned multi-modal pairs. Furthermore, we introduce a Heterogeneous Feature Aggregation (HFA) module to facilitate effective distillation and feature representation. Extensive experiments demonstrate that our method, SGPFeat, enhanced by Semantic and Geometric Priors, achieves state-of-the-art performance across diverse multi-modal image matching benchmarks.

SGPFeat: Semantic and Geometric Priors for Multi-modal Image Matching

Multi-view clustering aims to group data by integrating complementary information from multiple views. However, the inherent heterogeneity among views often leads to feature entanglement, severely limiting clustering performance. To address this challenge, we propose DC-SPAN—a Dual Contrastive Attention Network—grounded in a disentangle-then-fuse paradigm. DC-SPAN employs a dual-path variational architecture to explicitly decompose each view into shared and private latent subspaces. These representations are then robustly integrated via a Product-of-Experts (PoE) mechanism. At the heart of our model is a novel dual contrastive learning objective that simultaneously encourages alignment of shared components across views and enforces separation of private ones, enabling structured and disentangled representations. A cross-attention fusion module further adaptively aggregates these latent factors to yield a unified, discriminative embedding. The overall model is trained end-to-end using a composite loss function that incorporates reconstruction, orthogonality, and contrastive terms, along with a two-stage training scheme for improved stability. Extensive experiments on benchmark datasets demonstrate that DC-SPAN consistently outperforms existing state-of-the-art methods, highlighting its effectiveness and robustness in handling multi-view heterogeneity.

DC-SPAN: A Dual Contrastive Attention Network for Multi-View Clustering

Existing tool-augmented agentic systems are limited in the real world by (i) black-box reasoning steps that undermine trust of decision-making and pose safety risks, (ii) poor multimodal integration, which is inherently critical for healthcare tasks, and (iii) rigid and computationally inefficient agentic pipelines. We introduce **PASS** (**P**robabilistic **A**gentic **S**upernet **S**ampling), the *first* multimodal framework to address these challenges for Chest X-Ray (CXR) reasoning. PASS adaptively samples agentic workflows over a multi-tool graph, yielding decision paths annotated with interpretable probabilities. Given the complex CXR reasoning task with multimodal medical data, PASS leverages its learned task-conditioned distribution over the agentic supernet. Thus, it adaptively selects the most suitable tool at each supernet layer, offering probability-annotated trajectories for post-hoc audits and directly enhancing medical AI safety. PASS also continuously compresses salient findings into an evolving personalized memory, while dynamically deciding whether to deepen its reasoning path or invoke an early exit for efficiency. To optimize a Pareto frontier balancing performance and cost, we design a novel three-stage training procedure, including expert knowledge warm-up, contrastive path-ranking, and cost-aware reinforcement learning. To facilitate rigorous evaluation, we introduce `CAB-E`, a comprehensive benchmark for multi-step, safety-critical, free-form reasoning. Experiments across various benchmarks validate that PASS significantly outperforms strong baselines in multiple metrics (e.g., accuracy, AUC, LLM-J.) while balancing computational costs, pushing a new paradigm shift towards interpretable, adaptive, and multimodal medical agentic systems.

Downloads

Next from AAAI 2026

HyperAim: Hypergraph Contrastive Learning with Adaptive Multi-frequency Filters

Similar lecture

PRIME: A Process-Outcome Alignment Benchmark for Verifiable Reasoning in Mathematics and Engineering

.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 2026

HyperAim: Hypergraph Contrastive Learning with Adaptive Multi-frequency Filters

Similar lecture

PRIME: A Process-Outcome Alignment Benchmark for Verifiable Reasoning in Mathematics and Engineering

Downloads