Singapore

Frédéric Berdoz, Luca A Lanzendörfer, Kaan Bayraktar, Roger Wattenhofer

AAAI 2026

Data Attribution in Large Language Models via Bidirectional Gradient Optimization

regulation

explainability

interpretability

### Welcome to 2025 Annual Conference of the Nations of the Americas Chapter of the Association for Computational Linguistics

Welcome to the 2025 meeting of the Nations of the Americas Chapter of the Association for Computational Linguistics! I am proud to help organize the first NAACL conference to carry the new name of our organization, one that emphasizes inclusion for all of the Americas. I am also pleased to welcome you to Albuquerque, New Mexico, a state whose unique blend of cultural influences will make for an excellent backdrop for NAACL 2025, especially with this year’s special theme on NLP in a Multicultural World. 
**[Continue reading...](https://drive.google.com/file/d/1jX-qGhqVSZZCIrAnJaz798pflu5Irrdn/view?usp=sharing)**

*- Colin Cherry, Google, NAACL 2025 General Chair* 

[![](https://assets.underline.io/markdown_image/1/image/b087f8a4dc5816d6e1a6514e59c59ac3.png)](https://drive.google.com/file/d/1T96GzPqObXrMln2BMByCSXTSizjTg69P/view?usp=sharing)

You need to log in with the email address you registered with. Access credentials have been sent to your email. 

Please be sure to check your spam and other email folders if you do not see an email confirmation right away.

Please log in to explore this event.

To access NAACL 2025 event page you are required to register. Please follow [**this link**](https://2025.naacl.org/registration/registration/) to register. Access will depend on your registration type.

Please register!

NAACL 2025

Welcome to 2025 Annual Conference of the Nations of the Americas Chapter of the Association for Computational Linguistics

workshop 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.

AI safety and alignment evaluations (benchmarks) have become de facto governance tools in the rapidly evolving landscape of artificial intelligence. However, these critical instruments themselves lack systematic governance oversight, raising the central question: "Who evaluates the evaluators?" Through a systematic mapping of 298 papers and in-depth thematic analysis of 50 core studies, this paper identifies fundamental governance challenges in current AI evaluation practices. The bibliometric analysis reveals explosive growth in this field since 2023, with persistent gaps in transparency, accountability, and institutional independence. This position paper synthesizes current knowledge on evaluation governance and proposes a future agenda grounded in three core principles: participatory design with adversarial scrutiny, radical transparency through "Evaluation Cards," and institutional pluralism to ensure independent oversight. The paper argues for a paradigm shift from "evaluating AI" to "governing evaluations," offering actionable recommendations for researchers, policymakers, and practitioners to build a trustworthy evaluation ecosystem.

Who Evaluates the Evaluators? Governance Challenges in AI Safety and Alignment Evaluations: A Review and Future Agenda

Offensive Yet Efficient: Semantic Summarization via Obscene Lexicon

Stop Harm Before It Starts: iCRAFT for Agentic LLM Governance

Gaurav Chawla, Caleb DeLeeuw, Vanessa Dietze, Aniket Sharma

The Secret Agenda: LLMs Strategically Lie and Our Current Safety Tools Are Blind

Panatchakorn Anantaprayoon, Nataliia Babina, Jad Tarifi, Nima Asgharbeygi

Dynamic Alignment for Collective Agency: Toward a Scalable Self-Improving Framework for Open-Ended LLM Alignment

Navigating Healthcare AI Governance: Regulatory Divergence and the Harmonization Challenge Across Medical Domains

We propose Poincaré Gradient Descent (PGD), a first-order optimization method on the Poincaré ball. PGD replaces the exponential map with a projection-based retraction that is first-order equivalent while preserving geodesic structure, thereby reducing per-iteration computational cost. We establish convergence guarantees for smooth geodesically convex and strongly convex objectives, showing that PGD matches the iteration complexity of Riemannian Gradient Descent (RGD). A Möbius isometric initialization and step-size equivalence analysis further ensure stable and efficient updates. Numerical experiments confirm that PGD achieves the same theoretical convergence rate as RGD but runs significantly faster in practice, offering a simple and effective framework for scalable optimization in hyperbolic space.

Poincaré Gradient Descent: A Projection-Based Riemannian Method on the Poincaré Ball

Spiking neural networks (SNNs) have been proposed as an (energy-)efficient alternative to conventional artificial neural networks. However, the aspired benefits have not yet been realized in practice. To gain a better understanding of why this gap persists, we theoretically study both discrete-time and continuous-time models of leaky integrate-and-fire neurons.
In the discrete-time model, which is a widely used framework due to its amenability to conventional deep learning software and hardware approaches, we analyze the impact of explicit recurrent connections on the network size required to approximate continuously differentiable functions. We contrast this view by investigating the computational efficiency of digital systems that simulate spike-based computations in the continuous-time model. It turns out that even in well-behaved settings, the computational complexity of this task may grow super-polynomially in the prescribed accuracy.
Thereby, we exemplarily highlight the intricacies of realizing potential strengths in the biological context, namely recurrent connections and computational efficiency, of spike-based computations on digital systems.

How to realize efficient Spiking Neural Networks?

Bayesian neural networks (BNNs) require scalable sampling algorithms to approximate posterior distributions over parameters. Existing stochastic gradient Markov Chain Monte Carlo (SGMCMC) methods are highly sensitive to the choice of stepsize and adaptive variants such as pSGLD typically fail to sample the correct invariant measure without addition of a costly divergence correction term. In this work, we build on the recently proposed `SamAdams' framework for timestep adaptation (Leimkuhler, Lohmann, and Whalley 2025), introducing an adaptive scheme: SA-SGLD, which employs time rescaling to modulate the stepsize according to a monitored quantity (typically the local gradient norm). SA-SGLD can automatically shrink stepsizes in regions of high curvature and expand them in flatter regions, improving both stability and mixing without introducing bias. We show that our method can achieve more accurate posterior sampling than SGLD on high-curvature 2D toy examples and in image classification with BNNs using sharp priors.

Adaptive Stepsizing for Stochastic Gradient Langevin Dynamics in Bayesian Neural Networks

We study the fundamental expressivity limits of transformer models. We formalize the notion of accessible sequences—those that a transformer can produce for some prompt—and characterize how accessibility depends on prompt length and model precision. By partitioning the embedding space via the decoder readout into next-token argmax regions and extending transformers to a mean-field map on probability measures, we derive theoretical upper bounds on the number and length of accessible output sequences. We prove that (i) the maximal length of accessible sequences grows linearly with the prompt length, and (ii) beyond a critical threshold, the proportion of reachable sequences decays exponentially with sequence length. These bounds hold even with unbounded context and computation time, linking the expressivity limits of transformers to the geometry of their embedding space and the finiteness of their representational precision. Experiments using a “cramming” procedure confirm both the linear scaling and the post-threshold exponential decay.

Downloads

Next from AAAI 2026

Who Evaluates the Evaluators? Governance Challenges in AI Safety and Alignment Evaluations: A Review and Future Agenda

Similar lecture

On the Feasibility of In-Context Probing for Data Attribution