Singapore

Modern agentic frameworks operate under the fallacy of a Stationarity Trap: the dangerous assumption that meaning in underlying knowledge bases is static and user contexts are invariant. In this talk, evidence is presented from two distinct research streams to demonstrate that this assumption guarantees system failure. Applying our EchoCodes framework, we prove that “ground truth” content (e.g. data such as medical ontologies, intelligence signals, and software APIs) can undergo rate-independent semantic drift; and this can be detected before causing agents to execute employing meanings that have shifted. In the second dimension, as part of our work on the AiVisor system, we illustrate a “Personalization Paradox,” where agents optimizing for user utility incur statistically significant penalties in semantic consistency, effectively decoupling reasoning from lexically grounded baselines. Together, these findings define a non-stationary agent tradespace, where contextual drift (user alignment) and semantic dynamics (meaning discontinuity) converge to produce un-grounded agency. The convergence of these two phenomena, shifting ontological/ground-truth concepts and expanding contextual liabilities, creates a state of semantic insolvency. When such insolvent systems are granted the power to act, the result is un-grounded agency: high-confidence execution untethered from reality. Consequently, agentic deployments risk devolving into blind autonomy. We conclude by proposing a novel Semantic Guidance System to resolve the Stationarity Trap tradespace, moving the field beyond static evaluation toward dynamic semantic grounding for long-lifecycle agents.

AAAI 2026

Un-Grounded Agency and the Stationarity Trap in Agentic Workflows

computational dynamics

agentic ai

large language models

semantics

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.<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.

We assume a category theoretical approach to reasoning that is separate from but integrates with conventional logical systems and linear algebraic techniques such as employed by large language models, and reinforcement learning. The agenda is motivated by the ability to reason topological-ly in an expanded vocabulary of types not motivated by conventional ontological needs. One high value application is reasoning over open world situations and causally dynamic systems where essential facts are missing, situated influence is substantial, and significant implicit influ-ence is present. We are particularly looking at dynamic, unstable, possible outcomes, and situa-tions where outcome engineering is desired. The program is motivated by collecting open influences as salients, and focusing on type systems to support the concept that can have topological induction. The focus of this paper is on the na-ture of a type system where salience is a first class citizen and category theoretical topology cen-tric induction is enabled. While the research is not motivated by a specific use case, we focus on central nervous system modelling where fear memory extinction is modelled, so that we can leverage prior work. But ap-plications are expected to be rather broad, characterised by open world insights and frangible possibilities. Our group builds systems in Haskell frameworks so our a type system needs to be consistent with what can be supported programmatically. Because our group is particularly interested in human/machine navigation of structures, a consistent visual grammar sensitive to these type def-initions is desirable. This paper will include a brief survey of the literature with an emphasis on situation theory as a framework for sale and influence. Potential sensor technologies are consid-ered.

Categorical Type Systems for Salient Influence

Maintaining human oversight and broad stakeholder participation remain key challenges for trust management in human-AI teams. With the increasing deployment of agentic AI and large language models, the risk of unpredictable yet highly consequential outcomes could, if left unaddressed, undermine public and expert confidence in such teams. This study revisits that debate to articulate a framework that secures fairness, legitimacy, and trust. Drawing on theoretical and empirical analyses of citizen participation in fields adjacent or parallel to AI, it shows that the weaknesses of black-box AI have analogous counterparts in the isomorphic processes of black-box value elicitation and implementation decisions. Specifically, this study focuses on circumstances in which experts operate outside their domain of expertise to develop ad-hoc theories of reality. Examples of such “inexpert expert judgments” include situations where judges must evaluate complex statistical evidence in court, and where algorithm specialists must select features to be included in models based not on relevance to the goal but on availability or intuition. In cases like that, even something as trivial as including or excluding an interaction term in a regression-based model could lead to highly problematic outcomes. Bypassing domain specialist expertise and substituting it with stand-in knowledge is unnecessary when such expertise could be incorporated with modest effort.

The Hidden Black-Box: Expert Labor and the Case for Participatory AI;

Large Language Models (LLMs) excel at single-turn tasks such as instruction following and summarization, yet real-world deployments require sustained multi-turn interactions where user goals and conversational context persist and evolve. A recurring challenge in this setting is context drift: the gradual divergence of a model’s outputs from goal-consistent behavior across turns. Unlike single-turn errors, drift unfolds temporally and is poorly captured by static evaluation metrics. In this work, we present a study of context drift in multi-turn interactions and propose a simple dynamical framework to interpret its behavior. We formalize drift as the turn-wise KL divergence between the token-level predictive distributions of the test model and a goal-consistent reference model, and propose a recurrence model that interprets its evolution as a bounded stochastic process with restoring forces and controllable interventions. We instantiate this framework in both synthetic long-horizon rewriting tasks and realistic user–agent simulations, such as in tau-bench, measuring drift for several open-weight LLMs that are used as user simulators. Our experiments consistently reveal stable, noise-limited equilibria rather than runaway degradation, and demonstrate that simple reminder interventions reliably reduce divergence in line with theoretical predictions. Together, these results suggest that multi-turn drift can be understood as a controllable equilibrium phenomenon rather than as inevitable decay, providing a foundation for studying and mitigating context drift in extended interactions.

Drift No More? Context Equilibria in Multi-Turn LLM Interactions

Personalization of large language models (LLMs) has recently gained attention as a key approach to enhancing their usability and adaptability to individual users. However, when personalization relies on interaction histories, inconsistencies often arise between past and current user preferences, causing the model to fail to properly adapt to the user’s most recent intent. We propose a simple preference representation that encodes each preference as an intensity and polarity vector and aggregates past and recent preferences arithmetically to form a contradiction free, up to date personalized prompt. This enables the LLM to dynamically update user preferences at test time, linearly combining past and current preferences to produce a consistent, contradiction free personalized prompt. The proposed method does not require explicit ground truth data and can be applied to various downstream applications. Preliminary experiments using datasets demonstrate that our approach improves preference alignment compared to a non personalized model and achieves dynamic optimization over time.

Preference Descriptions for Dynamic Personalization of Large Language Models

As large language models (LLMs) become increasingly integrated into daily life, privacy concerns are on the rise. Driven by the appeal of convenient, universal solutions, current practices reflect a drift toward a one-size-fits-all approach to privacy. Unfortunately, we argue that a unified framework for prompt privacy is elusive and can instead mislead, creating even greater risk due to a false sense of safety. We identify five desirable properties of an effective privacy framework for prompts, namely protection guarantees, performance, efficiency, domain adaptability, and user accessibility, and highlight that existing frameworks, such as sanitization, differential privacy, cryptography, and contextual integrity, only satisfy a subset of these properties. Beyond individual frameworks, we find underlying tensions between these properties that preclude the development of a unified framework. We recommend two critical paths forward: emphasizing context-specific and application-specific evaluation of privacy frameworks, and fostering user awareness and privacy literacy.

A Unified Framework for Prompt Privacy is Elusive and Misleading

We propose a novel framework for persona-based language model system, motivated by the need for personalized AI agents that adapt to individual user preferences. In our approach, the agent embodies the user’s “persona” (e.g. user profile or taste) and is powered by a large language model (LLM). To enable the agent to leverage rich contextual information, we introduce a Knowledge-Graph-enhanced Retrieval Augmented Generation (Graph RAG) mechanism that constructs an LLM-derived graph index of relevant documents and summarizes communities of related information. Our framework generates personalized prompts by combining: (1) a summary of the user’s historical behaviors and preferences extracted from the knowledge graph, and (2) relevant global interaction patterns identified through graph-based community detection. This dynamic prompt engineering approach allows the agent to maintain consistent persona-aligned behaviors while benefiting from collective knowledge. On the LaMP benchmark, our method improves news categorization F1 by 11.1%, movie tagging F1 by 56.1%, and reduces product rating MAE by 10.4% over prior methods.

PersonaAgent with GraphRAG: Community-Aware Knowledge Graphs for Personalized LLM

Large Language Models (LLMs) have demonstrated exceptional capabilities but face two critical deployment challenges: high computational costs and scarcity of personalized domain training data. We address these dual challenges through a comprehensive framework that combines synthetic data generation with inference optimization techniques. Our approach employs LLMs for zero-shot and few-shot synthetic dataset creation while applying structural pruning, knowledge distillation, quantization, and prompt caching for computational efficiency. We evaluate three architectural paradigms: encoder-only, encoder-decoder, and decoder-only on synthetic building permit classification and assess optimization techniques on public benchmarks. Our systematic evaluation reveals that strategic architectural selection based on task characteristics is more critical than model complexity: encoder-only models provide superior efficiency-accuracy trade-offs for high-throughput scenarios, demonstrating that understanding problem requirements enables effective deployment without requiring the largest available models. Knowledge distillation emerges as the key optimization technique for personalized domains, recovering pruning-induced performance degradation with Llama3-8B+KD achieving 87.1% accuracy at 20% pruning while exceeding unpruned baselines. Complementary optimization strategies, including dynamic caching (5-18% latency reduction with zero performance loss) and hardware acceleration (up to 200× speedup) enable flexible deployment configurations for domain specific applications.

Domain-Specific LLM Adaptation: Bridging Personalization and Efficiency Through Synthetic Data and Optimization

As AI becomes more personal, e.g., Agentic AI, there is an increasing need for personalizing models for various use cases. Personalized federated learning (PFL) enables each client to collaboratively leverage other clients' knowledge for better adaptation to the task of interest, without privacy risks. Despite its potential, existing PFL methods remain confined to rather simplified scenarios where data and models are the same across clients. To move towards realistic scenarios, we propose FedMosaic, a method that jointly addresses data and model heterogeneity with a task-relevance-aware model aggregation strategy to reduce parameter interference, and a dimension-invariant module that enables knowledge sharing across heterogeneous architectures without huge computational cost. To mimic the real-world task diversity, we propose a multi-modal PFL benchmark spanning 40 distinct tasks with distribution shifts over time. The empirical study shows that FedMosaic outperforms the state-of-the-art PFL methods, excelling in both personalization and generalization capabilities under challenging, realistic scenarios.

Not All Clients Are Equal: Collaborative Model Personalization on Heterogeneous Multi-Modal Clients

Personalizing Large Language Model (LLM) agents requires conditioning them on user-specific data, creating a critical trade-off between task utility and data disclosure. While the utility of adding user data often exhibits diminishing returns (i.e., submodularity), enabling near-optimal greedy selection, real-world personalization is complicated by structural constraints. These include logical dependencies (e.g., selecting fact A requires fact B), categorical quotas (e.g., select at most one writing style), and hierarchical rules (e.g., select at most two social media preferences, of which at most one can be for a professional network). These constraints violate the assumptions of standard subset selection algorithms. We propose a principled method to formally model such constraints. We introduce a compilation process that transforms a user‚Äôs knowledge graph with dependencies into a set of abstract macro-facets. Our central result is a proof that common hierarchical and quota-based constraints over these macro-facets form a valid laminar matroid. This theoretical characterization lets us cast structured personalization as submodular maximization under a matroid constraint, enabling greedy with constant-factor guarantees (and (1‚àí1/e) via continuous greedy) for a much richer and more realistic class of problems.

Structured Personalization: Modeling Constraints as Matroids for Data-Minimal LLM Agents

Large Language Models (LLMs) are typically aligned for safety during the post-training phase; however, they may still generate harmful outputs that pose risks to users. 
This challenge underscores the need for robust safeguards that operate across both model inputs and outputs. In this work, we introduce RoGuard 1.0, a state-of-the-art instruction fine-tuned LLM designed to enhance the safety of LLM systems through comprehensive input‚Äìoutput moderation, using a pipeline of LLMs to enhance moderation capability. 
Built on the Llama-3.1-8B-Instruct backbone, our model is instruction fine-tuned to generalize across previously unseen safety taxonomies and demonstrates strong performance on out-of-domain safety benchmarks. 
The instruction fine-tuning process uses a mix of synthetic and open-source safety datasets, augmented with chain-of-thought (CoT) rationales and input inversion to enhance contextual understanding and decision making. 
To support systematic evaluation, we also release RoGuard Evaluation, a new benchmark featuring an extensible safety taxonomy to assess the effectiveness of LLM guardrails and moderation frameworks.

Downloads

Next from AAAI 2026

Categorical Type Systems for Salient Influence

Stay up to date with the latest Underline news!

PRESENTATIONS

CONFERENCES

COMPANY

RESOURCES

.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

Categorical Type Systems for Salient Influence

Stay up to date with the latest Underline news!

PRESENTATIONS

CONFERENCES

COMPANY

RESOURCES

Downloads