Singapore

Evaluating the safety and alignment of AI systems remains a critical challenge as foundation models grow increasingly sophisticated. Traditional evaluation methods rely heavily on human expert review, creating bottlenecks that cannot scale with the rapid pace of AI development. We introduce Jo.E (Joint Evaluation), a novel multi-agent collaborative framework that combines large language model evaluators, specialized AI agents, and strategic human expert involvement to conduct comprehensive safety assessments. Our framework employs a five-phase evaluation pipeline that systematically identifies vulnerabilities across multiple safety dimensions including adversarial robustness, fairness, ethics, and accuracy. Through extensive experiments on state-of-the-art models including GPT-4o, GPT-5, Llama 3.2, Phi 3, and Claude Sonnet 4, we demonstrate that Jo.E achieves approximately 22% improvement in vulnerability detection while reducing human expert time requirements by 54% compared to traditional evaluation approaches. Our results show that automated collaborative evaluation can significantly enhance both the efficiency and effectiveness of AI safety assessment without sacrificing rigor or comprehensive coverage.

AAAI 2026

Jo.E(Joint Evaluation) : A Multi-Agent Collaborative Framework for Comprehensive AI Safety Evaluation of Language Models

automated testing

foundation model assessment

ai safety evaluation

llm evaluation

multi-agent systems

human-ai collaboration

adversarial robustness

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.

Information retrieval in Large Language Models (LLMs) is increasingly recognized as intertwined with generation capabilities rather than mere lookup. While longer contexts are often assumed to improve retrieval, the effects of intra-context interference remain understudied. To address this, we adapt the proactive interference (PI) paradigm from cognitive science, where earlier information disrupts recall of newer updates. In humans, susceptibility to such interference is inversely linked to working memory capacity. We introduce PI-LLM, an evaluation that sequentially streams co-referenced key–value updates and queries only the final values. Although these final values are clearly positioned just before the query, LLM retrieval accuracy declines log-linearly toward zero as co-referenced interference accumulates; errors arise from retrieving previously overwritten values. Attempts to mitigate interference via prompt engineering (e.g., instructing models to ignore earlier input) yield limited success. These findings reveal a fundamental constraint on LLMs’ ability to disentangle interference and flexibly manipulate information, suggesting a working memory bottleneck beyond mere context access. For Agentic systems, reliable operation hinges on reconciling past and present states. Proactive interference corrupts long-horizon state maintenance; dependable agent tracking therefore needs explicit memory control and interference-aware context management. We expose a “know, cannot do” failure. Coreference-only needles expose a plan–execute disjunction: LLMs form the correct last-value retrieval plan but fail to carry it out, with execution reliability declining systematically with task complexity. Code and data will be publicly available.

Proactive Interference Reveals Working Memory Limits in LLMs Beyond Context Length

While a multi-agent approach based on large language models (LLMs) represents a promising strategy to surpass the capabilities of single models, its success is critically dependent on synergistic team composition. However, forming optimal teams is a significant challenge, as the inherent opacity of most models obscures the internal characteristics necessary for effective collaboration. In this paper, we propose an interaction-centric framework for automatic team composition that does not require any prior knowledge including their internal architectures, training data, or task performances. Our method constructs a "language model graph" that maps relationships between models from the semantic coherence of pairwise conversations, and then applies community detection to identify synergistic model clusters. Our experiments with diverse LLMs demonstrate that the proposed method discovers functionally coherent groups that reflect their latent specializations. Priming conversations with specific topics identified synergistic teams which outperform random baselines on downstream benchmarks and achieve comparable accuracy to that of manually-curated teams based on known model specializations. Our findings provide a new basis for the automated design of collaborative multi-agent LLM teams.

The Geometry of Dialogue: Graphing Language Models to Reveal Synergistic Teams for Multi-Agent Collaboration

Large language models (LLMs) often generate fluent
but factually incorrect statements despite having access
to relevant evidence, a failure mode rooted in how they
allocate attention between contextual and parametric
knowledge. Understanding and steering this internal
behavior is key both for trustworthy deployment and
for scientific interpretability of model mechanisms.
We introduce COMPASS (Context-Modulated PID At-
tention Steering System), a lightweight, interpretable
control framework that embeds a model-based feedback
loop directly within decoding. COMPASS quantifies
context reliance via a transparent metric, the Context
Reliance Score (CRS), which serves as an online probe
of how attention heads ground generation in evidence.
Using this interpretable signal, a PID controller dynam-
ically modulates attention heads to maintain factual
consistency without retraining or multi-pass decoding.
Across benchmarks (HotpotQA, XSum, HaluEval,
RAGTruth), COMPASS consistently reduces contextual
hallucination rates (2.8–5.8% absolute) while revealing
how distinct attention heads contribute to evidence
alignment. These results highlight feedback-driven in-
terpretability as a pathway toward scientific understand-
ing of LLM behavior.

COMPASS: Context-Modulated PID Attention Steering System for Hallucination Mitigation

In this work, we propose JETHICS, a Japanese dataset for evaluating ethics understanding of AI models. JETHICS contains 78K examples and is built by following the construction methods of the existing English ETHICS dataset. It includes four categories based normative theories and concepts from ethics and political philosophy; and one representing commonsense morality. Our evaluation experiments on non-proprietary large language models (LLMs) and on GPT-4o reveal that even GPT-4o achieves only an average score of about 0.7, while the best-performing Japanese LLM attains around 0.5, indicating a relatively large room for improvement in current LLMs.

JETHICS: Japanese Ethics Understanding Evaluation Dataset

Recent studies have pointed out that large language models (LLMs) face challenges in representing diverse value systems and facilitating consensus building, suggesting a potential incompatibility with democratic decision-making processes. As more advanced AI systems emerge, these issues are likely to become even more severe. According to the theory of Instrumental Convergence, advanced AI agents tend to seek control over humans—treated as uncertain variables—in order to achieve their goals, implying the formation of a governance relationship between humans and AI. This study analyzes two scenarios from Bostrom’s control problem—(1) the Singleton Scenario and (2) the Multipolar Scenario—together with (3) the Ecosystem Scenario discussed in the Japanese AI-alignment community through the lens of Aristotle’s typology of political regimes (classified by the number of rulers and the orientation toward private or common benefit). In each scenario, the success of alignment and the structure of institutional design determine whether AI systems pursue public goods that include human welfare, or instead
prioritize their own objective functions. Based on this analysis, the study predicts emergent behavioral principles in each scenario (e.g., cooperative, dominant, or indifferent) and their degrees of negotiability with humans. This study provides historically grounded insights into the ethically emergent dynamics that may arise mechanistically within advanced AI systems. Through this perspective of Emergent Machine Ethics (EME), it contributes to the design of governance structures that enable sustainable coexistence between humanity and advanced AI.

Governance Forms in the Age of Superintelligence: An Aristotelian Analysis

We build a custom transformer model to study how neural networks make moral decisions on trolley-style dilemmas. The model processes structured scenarios using embeddings that encode who is affected, how many people, and which outcome they belong to. Our 2-layer architecture achieves 77\% accuracy on Moral Machine data while remaining small enough for detailed analysis. We use different interpretability techniques to uncover how moral reasoning distributes across the network, demonstrating that biases localize to distinct computational stages among other findings.

Building Interpretable Models for Moral Decision-Making

Moral cognition has traditionally been modeled as adherence to fixed ethical theories—deontology, consequentialism, virtue ethics—implemented as static rules or value functions. We propose Bounded Morality, a framework that reconceives morality as an adaptive computation under finite cognitive and informational resources. Extending Herbert Simon’s notion of bounded rationality, we formalize moral reasoning as the allocation of limited capacity across two dimensions: breadth (the scope of moral concern) and depth (the complexity of reasoning and integration). This trade-off defines a bounded moral space where ethical theories correspond to resource-efficient strategies optimized for distinct regimes. We further operationalize moral cognition as a six-stage computational pipeline—from salience detection to action planning—each stage transforming information while consuming resources. This perspective unifies insights from developmental psychology, cognitive science, and AI alignment, offering design principles for systems whose moral competence scales with computational capacity. Bounded Morality reframes alignment not as value imitation but as capacity expansion toward increasingly inclusive and integrated reasoning.

Bounded Morality: An Algorithmic Framework for Moral Computation

We introduce the first complete formal solution to corrigibility in the off-switch game, with provable guarantees in multistep, partially observed environments. Our framework consists of five structurally separate utility heads—deference, switch-access preservation, truthfulness, low-impact behavior via a belief-based extension of Attainable Utility Preservation, and bounded task reward—combined lexicographically by strict weight gaps. Theorem 1 proves exact single-round corrigibility in the partially observable off-switch game; Theorem 3 extends the guarantee to multi-step, self-spawning agents, showing that even if each head is learned to meansquared error ε and the planner is ε-sub-optimal, the probability of violating any safety property is bounded while still ensuring net human benefit. In contrast to Constitutional AI or RLHF/RLAIF, which merge all norms into one learned scalar, our separation makes obedience and impact-limits provably dominate even when incentives conflict. For settings where adversaries can modify the agent, we prove that deciding whether an arbitrary post-hack agent will ever violate corrigibility is undecidable by reduction to the halting problem, then carve out a finite-horizon “decidable island” where safety can be certified in randomized polynomial time and verified with privacy-preserving, constant-round zero-knowledge proofs.

Core Safety Values for Provably Corrigible Agents

Multi-view pose estimation is essential for quantifying animal behavior in scientific research, yet current methods struggle to achieve accurate tracking with limited labeled data and suffer from poor uncertainty estimates. We address these challenges with a comprehensive framework combining novel training and post-processing techniques, and a model distillation procedure that leverages the strengths of these techniques to produce a more efficient and effective pose estimator. Our multi-view transformer (MVT) utilizes pretrained backbones and enables simultaneous processing of information across all views, while a novel patch masking scheme learns robust cross-view correspondences without camera calibration. For calibrated setups, we incorporate geometric consistency through 3D augmentation and a triangulation loss. We extend the existing Ensemble Kalman Smoother (EKS) post-processor to the nonlinear case and enhance uncertainty quantification via a variance inflation technique. Finally, to leverage the scaling properties of the MVT, we design a distillation procedure that exploits improved EKS predictions and uncertainty estimates to generate high-quality pseudo-labels, thereby reducing dependence on manual labels. Our framework components consistently outperform existing methods across three diverse animal species (flies, mice, chickadees), with each component contributing complementary benefits. The result is a practical, uncertainty-aware system for reliable pose estimation that enables downstream behavioral analyses under real-world data constraints.

An Uncertainty-Aware Framework For Data-Efficient Multi-View Animal Pose Estimation

Pattern separation, essential for encoding distinct memories of overlapping contexts, relies on dentate gyrus coding, which is shaped by entorhinal input and strong lateral inhibition. The pattern-separated state space provided by the hippocampus is thought to facilitate striatal-dependent reinforcement learning, enabling associations between sensory features and outcomes. Although synaptic plasticity, value prediction error modulation, and adult neurogenesis have been implicated in this process, their precise contributions remain unclear. To investigate the computational mechanisms underlying pattern separation, we developed neural network models incorporating an entorhinal cortex–dentate gyrus–striatal circuit. Simulations suggest that lateral inhibition is necessary for forming a decorrelated coding subspace, whereas hippocampal plasticity and dopamine modulation are not required for value learning. These findings dissociate neural pattern separation in hidden-layer representations from behavioral discrimination at the model output, highlighting how biologically grounded architectures and learning rules can enhance interpretability.

Downloads

Next from AAAI 2026

Proactive Interference Reveals Working Memory Limits in LLMs Beyond Context Length

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

Proactive Interference Reveals Working Memory Limits in LLMs Beyond Context Length

Stay up to date with the latest Underline news!

PRESENTATIONS

CONFERENCES

COMPANY

RESOURCES

Downloads