Singapore

Recent LLMs like DeepSeek-R1 have demonstrated state-of-the-art performance by integrating deep thinking and complex reasoning during generation. However, the internal mechanisms behind these reasoning processes remain unexplored. We observe *reasoning* LLMs consistently use vocabulary associated with human reasoning processes. We hypothesize these words correspond to specific reasoning moments within the models&#39; internal mechanisms. To test this hypothesis, we employ Sparse Autoencoders (SAEs), a technique for sparse decomposition of neural network activations into human-interpretable features. We introduce *ReasonScore*, an automatic metric to identify active SAE features during these reasoning moments. We perform manual and automatic interpretation of the features detected by our metric, and find those with activation patterns matching uncertainty, exploratory thinking, and reflection. Through steering experiments, we demonstrate that amplifying these features increases performance on reasoning-intensive benchmarks ($+2.2$%) while producing longer reasoning traces ($+20.5$%). Using model diffing technique, we provide evidence that these features are present only in models with reasoning capabilities. Our work provides the first step towards a mechanistic understanding of reasoning in LLMs.

AAAI 2026

I Have Covered All the Bases Here: Interpreting Reasoning Features in Large Language Models via Sparse Autoencoders

and evaluation of nlp models

(large) language models

representation learning

interpretability

analysis

Recent LLMs like DeepSeek-R1 have demonstrated state-of-the-art performance by integrating deep thinking and complex reasoning during generation. However, the internal mechanisms behind these reasoning processes remain unexplored. We observe *reasoning* LLMs consistently use vocabulary associated with human reasoning processes. We hypothesize these words correspond to specific reasoning moments within the models' internal mechanisms. To test this hypothesis, we employ Sparse Autoencoders (SAEs), a technique for sparse decomposition of neural network activations into human-interpretable features. We introduce *ReasonScore*, an automatic metric to identify active SAE features during these reasoning moments. We perform manual and automatic interpretation of the features detected by our metric, and find those with activation patterns matching uncertainty, exploratory thinking, and reflection. Through steering experiments, we demonstrate that amplifying these features increases performance on reasoning-intensive benchmarks ($+2.2$%) while producing longer reasoning traces ($+20.5$%). Using model diffing technique, we provide evidence that these features are present only in models with reasoning capabilities. Our work provides the first step towards a mechanistic understanding of reasoning in LLMs.

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.

Hash center-based deep hashing methods improve upon pairwise or triplet-based approaches by assigning fixed hash centers to each class as learning targets, thereby avoiding the inefficiency of local similarity optimization. However, random center initialization often disregards inter-class semantic relationships. While existing two-stage methods mitigate this by first refining hash centers with semantics and then training the hash function, they introduce additional complexity, computational overhead, and suboptimal performance due to stage-wise discrepancies.
To address these limitations, we propose $\textbf{Center-Reassigned Hashing (CRH)}$, an end-to-end framework that $\textbf{dynamically reassigns hash centers}$ from a preset codebook while jointly optimizing the hash function. Unlike previous methods, CRH adapts hash centers to the data distribution $\textbf{without explicit center optimization phases}$, enabling seamless integration of semantic relationships into the learning process. Furthermore, $\textbf{a multi-head mechanism}$ enhances the representational capacity of hash centers, capturing richer semantic structures. Extensive experiments on three benchmarks demonstrate that CRH learns semantically meaningful hash centers and outperforms state-of-the-art deep hashing methods in retrieval tasks.

Codebook-Centric Deep Hashing: End-to-End Joint Learning of Semantic Hash Centers and Neural Hash Function

Efficient visual backbone design remains crucial for resource-constrained computer vision applications. Inspired by the adaptive continuous-time dynamics observed in biological neurons, we propose FVNet, a novel lightweight architecture that integrates liquid neural dynamics for efficient and dynamic visual feature extraction. Central to FVNet is the Fluid Temporal Flow Unit (FTFU), which employs continuous-time equations with learnable time constants to capture spatio-temporal dependencies adaptively. By further stacking these units in a Multi-Phase Fluid Block (MPFB), our model processes features across parallel temporal scales, enabling context-aware feature encoding without incurring excessive computational overhead. Through a discrete closed-form solution, FVNet achieves the representational power of continuous-time models while avoiding the instability and overhead of iterative numerical solvers. Extensive experiments on various vision tasks demonstrate that FVNet achieves superior performance and efficiency over existing state-of-the-art lightweight networks.

FVNet: Harnessing Liquid Neural Dynamics for Lightweight Visual Representation

Generative diffusion models show promise for data augmentation. However, applying them to fine-grained tasks presents a significant challenge: ensuring synthetic images accurately capture the subtle, category-defining features critical for high fidelity. Standard approaches, such as text-based Classifier-Free Guidance (CFG), often lack the required specificity, potentially generating misleading examples that degrade fine-grained classifier performance. To address this, we propose Hierarchically Guided Fine-grained Augmentation (HiGFA). HiGFA leverages the temporal dynamics of the diffusion sampling process. It employs strong text and transformed contour guidance with fixed strengths in the early-to-mid sampling stages to establish overall scene, style, and structure. In the final sampling stages, HiGFA activates a specialized fine-grained classifier guidance and dynamically modulates the strength of all guidance signals based on prediction confidence. This hierarchical, confidence-driven orchestration enables HiGFA to generate diverse yet faithful synthetic images by intelligently balancing global structure formation with precise detail refinement. Experiments on several FGVC datasets demonstrate the effectiveness of HiGFA.

HiGFA: Hierarchical Guidance for Fine-grained Data Augmentation with Diffusion Models

Recent advances in text-to-image (T2I) diffusion models have significantly improved semantic image editing, yet most methods fall short in performing 3D-aware object manipulation. In this work, we present ***FFSE***, a 3D-aware autoregressive framework designed to enable intuitive, physically-consistent object editing directly on real-world images. Unlike previous approaches that either operate in image space or require slow and error-prone 3D reconstruction, ***FFSE*** models editing as a sequence of learned 3D transformations, allowing users to perform arbitrary manipulations, such as translation, scaling, and rotation, while preserving realistic background effects (e.g. shadows, reflections) and maintaining global scene consistency across multiple editing rounds. To support learning of multi-round 3D-aware object manipulation, we introduce ***3DObjectEditor***, a hybrid dataset constructed from simulated editing sequences across diverse objects and scenes, enabling effective training under multi-round and dynamic conditions. Extensive experiments show that the proposed ***FFSE*** significantly outperforms existing methods in both single-round and multi-round 3D-aware editing scenarios.

Free-Form Scene Editor: Enabling Multi-Round Object Manipulation Like in a 3D Engine

We propose Federated Preconditioned Mixing (FedPM), a novel Federated Learning (FL) method that leverages second-order optimization. Prior methods - such as LocalNewton, LTDA, and FedSophia - have incorporated second-order optimization in FL by performing iterative local updates on clients and applying simple mixing of local parameters on the server. However, these methods often suffer from drift in local preconditioners, which significantly disrupts the convergence of parameter training, particularly in heterogeneous data settings. To overcome this issue, we refine the update rules by decomposing the ideal second-order update - computed using globally preconditioned global gradients - into parameter mixing on the server and local parameter updates on clients. As a result, our FedPM introduces preconditioned mixing of local parameters on the server, effectively mitigating drift in local preconditioners. We provide a theoretical convergence analysis demonstrating a superlinear rate for strongly convex objectives in scenarios involving a single local update. To demonstrate the practical benefits of FedPM, we conducted extensive experiments. The results showed significant improvements with FedPM in the test accuracy compared to conventional methods incorporating simple mixing, fully leveraging the potential of second-order optimization.

FedPM: Federated Learning Using Second-order Optimization with Preconditioned Mixing of Local Parameters

Recent advances in large language models (LLMs) and dense retrievers have driven significant progress in retrieval-augmented generation (RAG). However, existing approaches face significant challenges in complex reasoning-oriented multi-hop retrieval tasks: 1) Ineffective reasoning-oriented planning: Prior methods struggle to generate robust multi-step plans for complex queries, as rule-based decomposers perform poorly on out-of-template questions. 2) Suboptimal reasoning-driven retrieval: Related methods employ limited query reformulation, leading to iterative retrieval loops that often fail to locate golden documents. 3) Insufficient reasoning-guided filtering: Prevailing methods lack the fine-grained reasoning to effectively filter salient information from noisy results, hindering utilization of retrieved knowledge. Fundamentally, these limitations all stem from the weak coupling between retrieval and reasoning in current RAG architectures. We introduce the Orchestrated Planner-Executor Reasoning Architecture (OPERA), a novel reasoning-driven retrieval framework. OPERA's Goal Planning Module (GPM) decomposes questions into sub-goals, which are executed by a Reason-Execute Module (REM) with specialized components for precise reasoning and effective retrieval. To train OPERA, we propose Multi-Agents Progressive Group Relative Policy Optimization (MAPGRPO), a novel variant of GRPO. Experiments on complex multi-hop benchmarks show OPERA's superior performance, validating both the MAPGRPO method and OPERA's design.

OPERA: A Reinforcement Learning--Enhanced Orchestrated Planner-Executor Architecture for Reasoning-Oriented Multi-Hop Retrieval

Large Language Models (LLMs) have demonstrated impressive capabilities in code generation. Like human programmers, LLMs call high-level APIs/libraries to program efficiently. However, this shortcut may hinder LLMs from learning the essential algorithm reasoning, leading instead to rote memorization of API usage. As a result, LLMs often struggle to generalize to new or domain-specific algorithms that lack ready-made library support. In this work, we propose **ARBench**, a novel benchmark for evaluating LLMs’ ability to generate machine learning algorithms from scratch, beyond merely invoking high-level APIs. It emphasizes algorithmic reasoning and implementation, distinguishing genuine understanding from superficial API usage. It covers fundamental and advanced machine learning tasks, rigorously assessing current LLMs’ capacity to implement these algorithms from scratch. Our evaluation reveals the strengths and weaknesses of state-of-the-art LLMs in algorithmic reasoning and generalization, offering valuable insights to guide future research and development. The code and data will be released after the paper is accepted.

ARBench: Algorithmic Reasoner or API Alchemist? Evaluating LLMs Beyond API Calls

The ASP(Q) language extends Answer Set Programming (ASP) with Quantifiers that operate over answer sets.
In this way, ASP(Q) allows for natural modeling of problems of complexity beyond NP with ASP. 
In this paper we focus on ASP(Q) programs with two quantifiers, i.e., 2-ASP(Q) programs, which can be used to model problems in the second level of the PH. 
In particular, we propose an approach for evaluating 2-ASP(Q) programs that is inspired by Counterexample Guided Abstraction Refinement (CEGAR).
Unlike existing state-of-the-art ASP(Q) solvers, which are typically based on QBF solvers, our new approach leverages ASP solvers, and suffers no overhead due to the effects of translating ASP(Q) in QBF.
Experimental results demonstrate that our technique consistently outperforms both state-of-the-art ASP(Q) solvers and ASP solvers that rely on disjunctive encodings, across benchmark problems located at the second level of the polynomial hierarchy.

2-ASP(Q) Solving Based on CEGAR

Decentralized cooperative multi-agent multi-armed bandits (DeCMA2B) consider how multiple agents collaborate in a decentralized multi-armed bandit setting. Though this problem has been extensively studied in previous works, most existing methods remain susceptible to various adversarial attacks. In this paper, we first study DeCMA2B with adversarial corruption, where an adversary can corrupt reward observations of all agents with a limited corruption budget. We propose a robust algorithm, called DeMABAR, which ensures that each agent’s individual regret only suffers an additive term proportional to the corruption budget. Then we consider a more realistic scenario where the adversary can only attack a small number of agents. Our theoretical analysis shows that the DeMABAR algorithm can also almost completely eliminate the influence of adversarial attacks and is inherently robust in the Byzantine setting, where an unknown fraction of the agents can be Byzantine, i.e., arbitrarily select arms and communicate wrong information. We also conduct numerical experiments to illustrate the robustness and effectiveness of proposed methods.

Robust Decentralized Multi-armed Bandits: From Corruption-Resilience to Byzantine-Resilience

Semantic understanding of large-scale aerial scenes represents a critical challenge in 3D computer vision, hindered by the prohibitive cost of dense annotation. This paper introduces EvoPropGS, a novel approach for the semantic segmentation of 3D Gaussian Splatting models that requires only minimal supervision. Our core insight is to leverage the inherent structural repetitions within aerial environments to propagate semantic information from a sparse set of annotations across the entire 3D scene. Our approach constructs a prompt library by pairing SAM-generated mask candidates with DINOv2 feature embeddings from annotated views. For unannotated regions, we generate pseudo-labels by matching region proposals with these featured prompts via cosine similarity. We then formulate optimal prompt selection as a discrete optimization problem solved via evolutionary search, guided by our novel fitness function that evaluates both 3D consistency and 2D semantic coherence. Extensive experiments demonstrate that EvoPropGS achieves accurate segmentation with only 2\% annotated pixels.

Downloads

Next from AAAI 2026

Codebook-Centric Deep Hashing: End-to-End Joint Learning of Semantic Hash Centers and Neural Hash Function

Stay up to date with the latest Underline news!

PRESENTATIONS

CONFERENCES

COMPANY

RESOURCES