Singapore

Large language models (LLMs) suffer from a lack of decision-making transparency, limiting their deployment in high-stakes domains such as healthcare. We propose a mechanistic interpretability framework that introduces two novel paradigms: Medical Fine-Tuning with Frozen Attention Layers (FTFA) and Posterior Adaptation Transcoders (PAT). FTFA freezes attention layers while fine-tuning only feed-forward network (FFN) parameters, enabling PAT to efficiently adapt pre-trained transcoders on the same data. This approach achieves over 1000× efficiency improvement compared to training transcoders from scratch. We theoretically justify this methodology and demonstrate its cost-effectiveness for cross-domain transfer. Transcoders are sparse autoencoders that replace MLP layers to provide interpretable feature representations. By substituting MLP layers of both base Gemma2-2b and its medical fine-tuned variant with per-layer transcoders, we enable feature-level attribution analysis. Through systematic pruning and node merging of resulting attribution graphs, we construct human-interpretable decision pathways. Our analysis reveals that LLMs employ two parallel mechanisms for medical diagnosis: pattern matching and multi-hop reasoning, with fine-tuned models demonstrating enhanced correct reasoning patterns. This work provides a practical framework for training transcoders on fine-tuned models at minimal cost, enabling broader application of mechanistic interpretability across domains and potentially guiding model training through transcoder-based analysis.

AAAI 2026

Efficient Transcoder Adaptation for Fine-Tuned Models: Revealing Medical Reasoning Mechanisms in Large Language Models

mechanistic interpretability

evidence-based medicine

feature attribution

poster

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.<br><br>

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

Dataset diversity plays a pivotal role for the successful training of many machine learning models, particularly in the supervised fine-tuning (SFT) stage of large language model (LLM) development. 
Despite increasing recognition of its importance, systematic analyses of dataset diversity still remain underexplored. 
To address this gap, this work presents a systematic taxonomy of existing diversity-control strategies, which primarily focus on the $\textit{instruction}$ component, operating at either $\underline{macroscopic}$ (entire instruction semantics) or $\underline{mesoscopic}$ levels (instruction units), and furthermore introduces a novel analysis of $\underline{microscopic}$ diversity within the $\textit{response}$ component, specifically analyzing the statistical distribution of **tokens** in SFT training samples.
In the experimental evaluation, we construct fixed-size datasets (e.g., 10,000 samples each) from a corpus of 117,000 open-source SFT samples, incorporating six distinct diversity-control strategies spanning macro-, meso-, and microscopic levels applied to both instructions and responses. 
We then fine-tune LLMs on these datasets to assess the six diversity-control strategies.
Results reveal that while macroscopic and mesoscopic strategies lead to higher performance with increasing diversity, the microscopic strategy in responses exhibits both a stronger correlation between model performance and the degree of diversity and superior performance with maximum diversity across all strategies.
These findings offer actionable insights for constructing high-performance SFT datasets.

From Macro to Micro: Probing Dataset Diversity in Language Model Fine-Tuning

Large Language Models (LLMs) have demonstrated remarkable performance across a wide range of natural language processing tasks. However, temporal reasoning, particularly under complex temporal constraints, remains a major challenge. To this end, existing approaches have explored symbolic methods, which encode temporal structure explicitly, and reflective mechanisms, which revise reasoning errors through multi-step inference. Nonetheless, symbolic approaches often underutilize the reasoning capabilities of LLMs, while reflective methods typically lack structured temporal representations, which can result in inconsistent or hallucinated reasoning. As a result, even when the correct temporal context is available, LLMs may still misinterpret or misapply time-related information, leading to incomplete or inaccurate answers. To address these limitations, in this work, we propose Neuro-Symbolic Temporal Reasoning (NeSTR), a novel framework that integrates structured symbolic representations with hybrid reflective reasoning to enhance the temporal sensitivity of LLM inference. NeSTR preserves explicit temporal relations through symbolic encoding, enforces logical consistency via verification, and corrects flawed inferences using abductive reflection. Extensive experiments on diverse temporal question answering benchmarks demonstrate that NeSTR achieves superior zero-shot performance and consistently improves temporal reasoning without any fine-tuning, showcasing the advantage of neuro-symbolic integration in enhancing temporal understanding in large language models.

NeSTR: A Neuro-Symbolic Abductive Framework for Temporal Reasoning in Large Language Models

Multi-agent reinforcement learning (MARL) is a powerful paradigm for solving cooperative and competitive decision-making problems. While many MARL benchmarks have been proposed, few combine continuous state and action spaces with challenging coordination and planning tasks. We introduce CAMAR, a new MARL benchmark designed explicitly for multi-agent pathfinding in environments with continuous actions. CAMAR supports cooperative and competitive interactions between agents and runs efficiently at up to 100,000 environment steps per second. We also propose a three-tier evaluation protocol to better track algorithmic progress and enable deeper analysis of performance. In addition, CAMAR allows the integration of classical planning methods such as RRT and RRT* into MARL pipelines. We use them as standalone baselines and combine RRT* with popular MARL algorithms to create hybrid approaches. We provide a suite of test scenarios and benchmarking tools to ensure reproducibility and fair comparison. Experiments show that CAMAR presents a challenging and realistic testbed for the MARL community.

CAMAR: Continuous Actions Multi-Agent Routing

Although lip-to-speech synthesis (L2S) has achieved significant progress in recent years, current state-of-the-art methods typically rely on intermediate representations such as mel-spectrograms or discrete self-supervised learning (SSL) tokens. The potential of latent diffusion models (LDMs) in this task remains largely unexplored. In this paper, we introduce SLD-L2S, a novel L2S framework built upon a hierarchical subspace latent diffusion model. Our method aims to directly map visual lip movements to the continuous latent space of a pre-trained neural audio codec, thereby avoiding the information loss inherent in traditional intermediate representations. The core of our method is a hierarchical architecture that processes visual representations through multiple parallel subspaces, initiated by a subspace decomposition module. To efficiently enhance interactions within and between these subspaces, we design the diffusion convolution block (DiCB) as our network backbone. Furthermore, we employ a reparameterized flow matching technique to directly generate the target latent vectors. This enables a principled inclusion of speech language model (SLM) and semantic losses during training, moving beyond conventional flow matching objectives and improving synthesized speech quality. Our experiments show that SLD-L2S achieves state-of-the-art generation quality on multiple benchmark datasets, surpassing existing methods in both objective and subjective evaluations.

SLD-L2S: Hierarchical Subspace Latent Diffusion for High-Fidelity Lip to Speech Synthesis

Deploying language models often requires navigating accuracy vs. performance trade-offs to meet latency constraints while preserving utility. Traditional model distillation reduces size but incurs substantial costs through training separate models. We introduce ModularStarEncoder (MoSE), a 1-billion-parameter multi-exit encoder for code retrieval and classification that employs a novel Self-Distillation mechanism. This approach significantly enhances lower-layer representations, enabling flexible deployment of different model portions with favorable performance trade-offs. Our architecture improves text-to-code and code-to-code search by targeting specific encoder layers as exit heads, where higher layers guide earlier ones during training—improving intermediate representations at minimal additional cost. We further enhance MoSE with a repository-level contextual loss that maximizes training context window utilization. Additionally, we release a new dataset created through code translation that extends text-to-code benchmarks with cross-language code-to-code pairs. Evaluations demonstrate the effectiveness of Self-Distillation as a principled approach to trading inference cost for accuracy across various code understanding tasks.

MoSE: Hierarchical Self-Distillation Enhances Early Layer Embeddings

Multimodal Emotion Recognition (MER) aims to accurately identify human emotional states by integrating heterogeneous modalities such as visual, auditory, and textual data. Existing approaches predominantly rely on unified emotion labels to supervise model training, often overlooking a critical challenge: inter-modal emotion conflicts, wherein different modalities within the same sample may express divergent emotional tendencies. In this work, we address this overlooked issue by proposing a novel framework, Typicality-based Consistent-aware Multimodal Emotion Recognition (TiCAL), inspired by the stage-wise nature of human emotion perception. TiCAL dynamically assesses the consistency of each training sample by leveraging pseudo unimodal emotion labels alongside a typicality estimation. To further enhance emotion representation, we embed features in a hyperbolic space, enabling the capture of fine-grained distinctions among emotional categories. By incorporating consistency estimates into the learning process, our method improves model performance, particularly on samples exhibiting high modality inconsistency. Extensive experiments on benchmark datasets, e.g, CMU-MOSEI and MER2023, validate the effectiveness of TiCAL in mitigating inter-modal emotional conflicts and enhancing overall recognition accuracy, e.g., with about 2.6% improvements over the state-of-the-art DMD. The anonymous code is available in the supplementary materials.

TiCAL:Typicality-Based Consistency-Aware Learning for Multimodal Emotion Recognition

Randomized Response (RR) is a protocol designed to collect and analyze categorical data with local differential privacy guarantees. It has been used as a building block of mechanisms deployed by Big tech companies to collect app or web users' data. Each user reports an automatic random alteration of their true value to the analytics server, which then estimates the histogram of the true unseen values of all users using a debiasing rule to compensate for the added randomness. A known issue is that the standard debiasing rule can yield a vector with negative values (which can not be interpreted as a histogram), and there is no consensus on the best fix. An elegant but slow solution is the Iterative Bayesian Update algorithm (IBU), which converges to the Maximum Likelihood Estimate (MLE) as the number of iterations goes to infinity. This paper bypasses IBU by providing a direct mathematical formula for the MLE for RR and compares it with other estimation methods experimentally to help practitioners decide which one to use.

Estimating the True Distribution of Data Collected with Randomized Response

Chart editing reduces manual effort in visualization design. Typical benchmarks assume access to complete chart code, which is unrealistic for real-world applications. In this paper, we present ChartEditVista, a comprehensive benchmark consisting of 7,964 samples spanning 31 chart categories. It encompasses diverse editing instruction types and covers nearly all editable chart elements. The inputs in ChartEditVista include only the original chart image and natural language editing instructions, without the original chart codes. ChartEditVista is generated through a fully automated pipeline that produces, edits, and verifies charts, ensuring high-quality data. Besides, we introduce two novel fine-grained, rule-based evaluation metrics: the layout metric, which evaluates the position, size; and color of graphical components, and the text metric, which jointly assesses textual content and font styling. Building on top of ChartEditVista, we present ChartEditor, a model trained using a reinforcement learning framework that incorporates a novel rendering reward to simultaneously enforce code executability and visual fidelity. Through extensive experiments and human evaluations, we demonstrate that ChartEditVista provides a robust evaluation, while ChartEditor consistently outperforms models with similar-scale and larger-scale on chart editing tasks.

ChartEditor: A Reinforcement Learning Framework for Robust Chart Editing

Transformers are the cornerstone of modern large language models, but their quadratic computational complexity limits efficiency in long-sequence processing. Recent advancements in Mamba, a state space model (SSM) with linear complexity, offer promising efficiency gains but suffer from unstable contextual learning and multitask generalization. Some works conduct layer-level hybrid structures that combine Transformer and Mamba layers, aiming to make full use of both advantages. This paper proposes TransMamba, a novel sequence-level hybrid framework that unifies Transformer and Mamba through shared parameter matrices (QKV and CBx), and thus could dynamically switch between attention and SSM mechanisms at different token lengths and layers. We design the Memory converter to bridge Transformer and Mamba by converting attention outputs into SSM-compatible states, ensuring seamless information flow at TransPoints where the transformation happens. The TransPoint scheduling is also thoroughly explored for balancing effectiveness and efficiency. We conducted extensive experiments demonstrating that TransMamba achieves superior training efficiency and performance compared to single and hybrid baselines, and validated the deeper consistency between Transformer and Mamba paradigms at sequence level, offering a scalable solution for next-generation language modeling.

TransMamba: A Sequence-Level Hybrid Transformer-Mamba Language Model

Large Language Models (LLMs) have demonstrated remarkable performance across a wide range of tasks, yet they generally lack self-awareness—often displaying overconfidence when confronted with questions beyond their knowledge boundaries. This limitation severely hinders their trustworthiness in high-stakes scenarios. Existing calibration methods typically rely on sampling accuracy—derived from multiple outputs—as a proxy for model confidence. However, this coarse-grained metric fails to capture the model’s internal cognitive states, such as confusion, hallucination, or persistent belief in false knowledge.
To address this, we propose \texttt{CogConf} (Cognitive Confidence), a cognitively grounded uncertainty signal that extends sampling accuracy by incorporating the semantic diversity of incorrect answers and the model’s abstention behaviors. By shifting the focus from sampling-based to cognition-oriented uncertainty modeling, \texttt{CogConf} offers a more faithful reflection of the model's internal beliefs. Building on this signal, we introduce \textsc{CogAlign}, a simple yet effective alignment framework that explicitly aligns the model’s verbalized confidence with \texttt{CogConf}, thereby producing uncertainty estimates that better reflect the model’s internal cognition.
Experimental results on six knowledge-intensive in-domain and out-of-domain QA datasets demonstrate that \texttt{CogConf} robustly characterizes the model's internal uncertainty. Building on this foundation, \textsc{CogAlign} guides the model's expression to significantly enhance the trustworthiness and utility of its uncertainty calibration without compromising its underlying QA capabilities, while also demonstrating strong cross-task generalization and output stability. Offering a new pathway toward building more trustworthy LLMs.

Content not yet available

Next from AAAI 2026

From Macro to Micro: Probing Dataset Diversity in Language Model Fine-Tuning

Stay up to date with the latest Underline news!

PRESENTATIONS

CONFERENCES

COMPANY

RESOURCES