Singapore

Speculative decoding accelerates language model inference by separating generation into fast drafting and parallel verification. Its main limitation is drafter–verifier misalignment, which limits token acceptance and reduces overall effectiveness. While small drafting heads trained from scratch compensate with speed, they struggle when verification dominates latency or when inputs are out of distribution. In contrast, pretrained drafters, though slower, achieve higher acceptance rates thanks to stronger standalone generation capabilities, making them competitive when drafting latency is negligible relative to verification or communication overhead. In this work, we aim to improve the acceptance rates of pretrained drafters by introducing a lightweight dynamic alignment mechanism: a \emph{steering vector} computed from the verifier’s hidden states and injected into the pretrained drafter. Compared to existing offline alignment methods such as distillation, our approach boosts the number of accepted tokens by up to 35\% under standard sampling and 22\% under greedy sampling, all while incurring negligible computational overhead. Importantly, our approach can be retrofitted to existing architectures and pretrained models, enabling rapid adoption.

AAAI 2026

Steering Pretrained Drafters During Speculative Decoding

nlp: (large) language models

nlp: applications

nlp: generation

## Welcome!
"I am excited to welcome you to this year’s edition of the Conference on Empirical Methods in Natural Language Processing! Importantly, it marks the 30th edition of EMNLP. With over 8,000 submissions, more than 3,000 accepted papers, and thousands of attendees, we have come a long way from that first
workshop, which had 14 accepted papers. As the field looks ahead, Suzhou is the fitting location for celebrating this milestone: rooted in a long literary tradition, yet modern and forward-looking, and home to a large share of the NLP community." 

*Message from the General Chair, Dirk Hovy*

EMNLP 2025

Celebrate 30 Years of EMNLP! 
EMNLP 2025 will be held in Suzhou, China from November 5th to November 9th, 2025.

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 

Registration Required

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.

In this talk, we discuss efficient model specialization
algorithm to adapt the pretrained model towards downstream
tasks while improving its efficiency, efficiently
generalizing to multiple tasks via dynamic architectures,
and improving inference-time efficiency utilizing the
diversity within model block functionalities. These
research directions serve as the foundation towards
co-designing models, tasks, systems, and hardware for a
reconfigurable efficient intelligence future.

Efficient Model Specialization via Training-time and Test-time Adaptation

We investigate whether large language models can discover and analyze U.S. tax-minimization strategies. This real-world domain challenges even seasoned human experts, and progress can reduce tax revenue lost from well-advised, wealthy taxpayers. We evaluate the most advanced LLMs on their ability to (1) interpret and verify tax strategies, (2) fill in gaps in partially specified strategies, and (3) generate complete, end-to-end strategies from scratch. This domain should be of particular interest to the LLM reasoning community: unlike synthetic challenge problems or scientific reasoning tasks, U.S. tax law involves navigating hundreds of thousands of pages of statutes, case law, and administrative guidance, all of which is updated regularly. Notably, LLM-based reasoning identified an entirely novel tax strategy, highlighting these models' potential for revolutionizing tax agencies' fight against tax abuse.

Can LLMs Identify Tax Abuse?

Existing LiDAR point cloud (LPC) data coding methods primarily focus on balancing compression efficiency and reconstruction quality according to the human vision system (HVS). However, these methods rarely consider the requirements of downstream scene understanding tasks from the perspective of the machine vision system (MVS). To address this challenge, we explore the maximum degree of LPC compression that has negligible impact on perception accuracy, called LPC-based just recognizable compression distortion (lpcJRCD). Specifically, we introduce a novel point-wise quantization approach for constructing a MVS-based LiDAR dataset and present a new lpcJRCD-guided intelligent compression framework tailored for MVS applications. To enhance MVS-based LPC compression efficiency, we develop a dual-feature interaction (DFI) module that fuses point and voxel features. Additionally, we propose a mask-based loss function to ensure accurate point-wise quality level prediction. 
Experimental results demonstrate the effectiveness of our proposed model in reducing the average bit rate by up to 94.98\% while preserving perception accuracy in autonomous

Perceive More with Less: LiDAR Point Cloud Compression at Just Recognizable Distortion for 3D Scene Understanding

Recently, large reasoning models have demonstrated strong mathematical and coding abilities, and deep search leverages their reasoning capabilities in challenging information retrieval tasks. Existing deep search works are generally limited to a single knowledge source, either local or the Web. However, enterprises often require private deep search systems that can leverage search tools over both local and the Web corpus. Simply training an agent equipped with multiple search tools using flat reinforcement learning (RL) is a straightforward idea, but it has problems such as low training data efficiency and poor mastery of complex tools. To address the above issue, we propose a hierarchical agentic deep search framework, HierSearch, trained with hierarchical RL. At the low level, a local deep search agent and a Web deep search agent are trained to retrieve evidence from their corresponding domains. At the high level, a planner agent coordinates low-level agents and provides the final answer. Moreover, to prevent direct answer copying and error propagation, we design a knowledge refiner that filters out hallucinations and irrelevant evidence returned by low-level agents. Experiments show that HierSearch achieves better performance compared to f l RL, and outperforms various deep search and multi-source retrieval-augmented generation baselines in six benchmarks across general, finance, and medical domains.

HierSearch: A Hierarchical Enterprise Deep Search Framework Integrating Local and Web Searches

Vision-Language Models (VLMs), exemplified by CLIP, have emerged as foundational for multimodal intelligence. However, their capacity for logical understanding remains significantly underexplored, resulting in critical **''logical blindspots''** that limit their reliability in practical applications. To systematically diagnose this, we introduce **LogicBench**, a comprehensive benchmark with over 50,000 vision-language pairs across 9 logical categories and 4 diverse scenarios: images, videos, anomaly detection, and medical diagnostics.
Our evaluation reveals that existing VLMs, even the state-of-the-art ones, fall at over 40 accuracy points below human performance, particularly in challenging tasks like Causality and Conditionality, highlighting their reliance on surface semantics over critical logical structures. To bridge this gap, we propose **LogicCLIP**, a novel training framework designed to boost VLMs' logical sensitivity through advancements in both data generation and optimization objectives. LogicCLIP utilizes logic-aware data generation and a contrastive learning strategy that combines coarse-grained alignment, a fine-grained multiple-choice objective, and a novel logical structure-aware objective. Extensive experiments demonstrate LogicCLIP's substantial improvements in logical comprehension across all LogicBench domains, significantly outperforming baselines. Moreover, LogicCLIP retains, and often surpasses, competitive performance on general vision-language benchmarks, demonstrating that the enhanced logical understanding does not come at the expense of general alignment. We believe that LogicBench and LogicCLIP will be important resources for advancing VLM logical capabilities. Relevant data, code, and models will be publicly released.

Logic Unseen: Revealing the Logical Blindspots of Vision-Language Models

Conventional document layout analysis (DLA) traditionally depends on empirical priors or a fixed set of learnable queries executed in a single forward pass. While sufficient for early-generation documents with a small, predetermined number of regions, this paradigm struggles with contemporary documents, which exhibit diverse element counts and increasingly complex layouts. To address challenges posed by modern documents, we present HybriDLA, a novel generative framework that unifies diffusion and autoregressive decoding within a single layer. The diffusion component iteratively refines bounding-box hypotheses, whereas the autoregressive component injects semantic and contextual awareness, enabling precise region prediction even in highly varied layouts. To further enhance detection quality, we design a multi-scale feature-fusion encoder that captures both fine-grained and high-level visual cues. This architecture elevates performance to 83.5% mean Average Precision (mAP). Extensive experiments on the DocLayNet and M$^{6}$Doc benchmarks demonstrate that HybriDLA sets a state-of-the-art performance, outperforming previous approaches. All data and models will be made publicly available.

HybriDLA: Hybrid Generation for Document Layout Analysis

Incomplete multi-view clustering (IMVC) aims to discover shared cluster structures from multi-view data with partial observations. The core challenges lie in accurately imputing missing views without introducing bias, while maintaining semantic consistency across views and compactness within clusters. 
To address these challenges, we propose DIMVC-HIA, a novel deep IMVC framework that integrates hierarchical imputation and alignment with four key components: 
(1) view-specific autoencoders for latent feature extraction, coupled with a view-shared clustering predictor to produce soft cluster assignments; 
(2) a hierarchical imputation module that first estimates missing cluster assignments based on cross-view contrastive similarity, and then reconstructs missing features using intra-view, intra-cluster statistics; 
(3) an energy-based semantic alignment module, which promotes intra-cluster compactness by minimizing energy variance around low-energy cluster anchors;
and (4) a contrastive assignment alignment module, which enhances cross-view consistency and encourages confident, well-separated cluster predictions. 
Experiments on benchmarks demonstrate that our framework achieves superior performance under varying levels of missingness.

Deep Incomplete Multi-View Clustering via Hierarchical Imputation and Alignment

The rapid advancement of GPU technology has unlocked powerful parallel processing capabilities, creating new opportunities to enhance classic search algorithms. This hardware has been exploited in best-first search algorithms with neural network-based heuristics by creating batched versions of A* and Weighted A* that delay heuristic evaluation until sufficiently many states can be evaluated in parallel on the GPU. But, research has not addressed how depth-first algorithms like IDA* or Budgeted Tree Search (BTS) can have their heuristic computations batched. This is more complicated in a tree search, because progress in the search tree is blocked until heuristic evaluations are complete. In this paper we show that GPU parallelization of heuristics can be effectively performed when the tree search is parallelized on the CPU while heuristic evaluations are parallelized on the GPU. We develop a parallelized cost-bounded depth-first search (CB-DFS) framework that can be applied to both IDA* and BTS, significantly improving their performance. We demonstrate the strength of the approach on the 3x3 Rubik's Cube and the 4x4 sliding tile puzzle (STP) with both classifier-based and regression-based heuristics.

A Parallel CPU-GPU Framework for Batching Heuristic Operations in Depth-First Heuristic Search

In many planning problems there are non-deterministic actions for which the outcome cannot be fully controlled by the planning agent. For critical tasks, we need to ensure that the strategy achieves the goal within a predictable time-frame and/or cost. Thus, we consider an adversarial planning setting and compute optimal policies that optimize the worst-case cost to reach the goal.

In this work, we introduce domain-independent optimal heuristic search algorithms for this adversarial setting. To guide the search, we show how to leverage classical planning heuristics by applying single-outcome determinization. We also generalize dominance techniques, that analyse when a state is as good as another, to the non-deterministic setting and apply them to prune the search space. Our experimental analysis shows that both methods greatly helps to compute optimal policies across multiple domains.

Dominance Pruning and Heuristics in Optimal Adversarial Non-Deterministic Planning

Numerous social movements (SMs) around the world help support the UN's Sustainable Development Goals (SDGs). Understanding how key events shape SMs is key to the achievement of the SDGs. We have developed SMART (Social Media Analysis \& Reasoning Tool) to track social movements related to the SDGs. SMART was designed by a multidisciplinary team of AI researchers, journalists, communications scholars and legal experts. This paper describes SMART's transformer-based multivariate time series Discourse Evolution Engine for Predictions about Social Movements (DEEP) to predict the volume of future articles/posts and the emotions expressed. DEEP outputs probabilistic forecasts with uncertainty estimates, providing critical support for editorial planning and strategic decision-making. We evaluate DEEP with a case study of the \emph{\#MeToo} movement by creating a novel longitudinal dataset (433K Reddit posts and 121K news articles) from September 2024 to June 2025 that will be publicly released for research purposes upon publication of this paper.

Downloads

Next from AAAI 2026

Efficient Model Specialization via Training-time and Test-time Adaptation

Similar lecture

Mamba Drafters for Speculative Decoding