The upcoming meeting, taking place on November 10 at Mila, will explore how we can collectively develop, govern, and deploy high-performing, reliable, and secure agentic systems by connecting academic researchers, industry experts, and practitioners.
Join us on the Venture Scientist Bootcamp, a full time, 4-month incubator at Mila, built specifically for deep tech founders with elite STEM backgrounds. You have until November 21, 2026 to apply!
We use cookies to analyze the browsing and usage of our website and to personalize your experience. You can disable these technologies at any time, but this may limit certain functionalities of the site. Read our Privacy Policy for more information.
Setting cookies
You can enable and disable the types of cookies you wish to accept. However certain choices you make could affect the services offered on our sites (e.g. suggestions, personalised ads, etc.).
Essential cookies
These cookies are necessary for the operation of the site and cannot be deactivated. (Still active)
Analytics cookies
Do you accept the use of cookies to measure the audience of our sites?
Multimedia Player
Do you accept the use of cookies to display and allow you to watch the video content hosted by our partners (YouTube, etc.)?
Publications
Imitation Learning via Discrete and Rhythmic Primitive Composition
Long-horizon robotic manipulation is inherently composed of semantically meaningful sub-tasks. While Movement Primitives (MPs) can model the… (see more)se components, their composition typically relies on manual engineering or sample-inefficient reinforcement learning. Existing supervised methods are limited by the lack of MP labels in standard datasets and often fail to generalize across discrete and rhythmic movements. We propose Neural Dynamic Composition (NDC), a supervised framework to compose Dynamic Movement Primitives (DMPs) using only raw trajectories for supervision, given a single example of each primitive. Our approach exploits parallel combinations of DMPs as a differentiable relaxation of categorical primitive selection, and utilizes a warp-invariant loss formulation to identify primitives despite spatial or temporal variations. A staged training curriculum is employed to learn the selection policy and DMP parameters, unifying discrete and rhythmic movements in one framework. Evaluated on 2D handwriting and 3D robotic tasks, our method outperforms baselines on most tasks and metrics. An ablation study demonstrates how each design choice contributes to the system's performance.
Large Language Models (LLMs) are increasingly used in high-stakes decision-making systems, where biased predictions can reinforce social and… (see more) economic disparities. Although prior work has examined intrinsic representational bias and unfair downstream behavior separately, it remains unclear whether mitigating intrinsic bias leads to fairer downstream outcomes. We introduce Fairness-Aware Concept Unlearning (FACU), a model-level mitigation method that adapts concept unlearning to fairness-oriented representation balancing. Unlike suppression-based approaches, FACU explicitly regularizes probability differences between stereotypical and anti-stereotypical associations while preserving predictive performance and language modeling quality. We evaluate FACU across three open-source LLMs, multiple intrinsic bias benchmarks, and three socio-economic classification datasets using both frozen LLM embeddings and LoRA-fine-tuned classifiers. FACU produces statistically significant reductions in intrinsic gender bias that are associated with downstream fairness improvements across most evaluated settings, datasets, models, and fairness metrics, without significantly degrading predictive performance. Combining FACU with extrinsic mitigation methods, particularly counterfactual data augmentation, yields further fairness improvements. These findings suggest that fairness-aware intrinsic mitigation can support fairer LLM-based decision-making and that bias mitigation should be addressed across both model development and downstream deployment stages.
2026-10-10
AAAI/ACM Conference on AI, Ethics, and Society (published)
The Promises and Limitations of Evaluation-Based AI Governance: Evidence from Expert Interviews
Lily Stelling
Jacob Charnock
Kevin L. Wei
David Krueger
We conducted 37 semi-structured expert interviews to
examine how frontier AI evaluations reduce societal-level
AI risks. Evaluations have be… (see more)come central to AI risk
governance, despite their shortcomings being widely
critiqued. Still, the case for how they reduce risks in
practice remains largely implicit. Drawing on expert
accounts, we reconstruct this case by identifying the
mechanisms through which evaluations are expected to reduce
risk, the assumptions underpinning those mechanisms, and
the challenges those assumptions face. Our findings
suggest that `evaluation-based governance' could contain
foundational gaps which undermine risk-reduction strategies
that depend on evaluations. Using thematic analysis, we
identified three key pathways through which evaluations
could reduce risk: by (i) making societal-level risks
concrete to decision-makers and the public (raised in 18
out of 37 interviews), (ii) serving as a basis for
governance commitments and regulation (18), and (iii)
incentivizing a better risk culture within AI companies
(16). However, these pathways often depended on shared
assumptions that participants themselves doubted. The main
challenges included, in order of frequency: (i) diminishing
informativeness of evaluation results as model capabilities
improve (21), (ii) measurement limitations that may be
inherent to the evaluation paradigm (18), (iii) a lack of
guarantees that evaluation results will be acted upon by
companies or governments (16), and (iv) the degree of
control and discretion AI companies retain over evaluation
design, disclosure, and response (12). Participants also
noted that evaluation-based governance could prove
counterproductive, by functioning as a permission structure
for developers' continued capability improvement (4). We
finish by discussing the implications of our findings for
the evaluation-based governance paradigm, and the
conditions that would improve its prospects.
2026-10-10
AAAI/ACM Conference on AI, Ethics, and Society (published)
Large language models (LLMs) can produce harmful content while generating a response. A useful safeguard should not only detect this risk du… (see more)ring generation but also intervene while leaving safe responses largely unchanged and provide an auditable record of the internal patterns supporting its decisions. We propose TRACE (Temporal Risk Assessment and Concept Editing), a framework that tracks harmful-response risk during generation, selectively edits token activations, and leaves an auditable trace of its detection and steering decisions. TRACE learns a codebook of recurring activation patterns, with each codebook vector representing a latent concept. At inference time, it assigns each token activation to the nearest codebook vector to obtain the corresponding concept. A gated recurrent unit (GRU) processes the activations and their assigned concept features to predict completed-response harmfulness and a monotone cumulative streaming risk. During steering, activations assigned to harmful concepts are steered along directions defined by nearby benign targets. An external LLM labels the learned concepts post hoc, making detection sequences and harmful-to-benign edit records interpretable. Experiments across three generators and the WildGuard and S-Eval benchmarks show that TRACE improves streaming F1 over Kelp in five of six comparisons while remaining competitive on completed-response detection. TRACE's selective steering reduces Harmbench attack success with little change in refusal on originally safe responses. The post-hoc audit reveals recurring semantic structure in the learned concepts and distinct traces associated with harmful detection, safe generation, and activation edits.
The increasing penetration of renewable energy sources in microgrids demands accurate day-ahead forecasting to support optimal battery energ… (see more)y storage system (BESS) scheduling. Yet the conventional predict-then-optimize paradigm trains forecasters on statistical accuracy metrics that are agnostic to the asymmetric and time-varying cost impacts of forecast errors. This work proposes a decision-focused learning (DFL) framework that trains gated recurrent unit forecasters for both load demand and photovoltaic generation directly on the downstream operational cost of a two-stage energy management system, where day-ahead BESS commitments are fixed prior to real-time balancing. Gradients are backpropagated from the real-time operational cost through both optimization stages to the forecaster weights via implicit differentiation, enabling the forecasters to internalize the penalty structure of the scheduling problem. The framework is validated on the MG2Lab experimental microgrid testbed at Politecnico di Milano across two operational scenarios of increasing grid constraint and eight forecasting configurations comprising single-source and combined-forecaster settings. Results illustrate that the DFL-trained pipeline reduces operational cost by 3.78% under normal grid conditions and by 30.60% under grid-constrained conditions relative to the accuracy- optimized baseline, while nearly eliminating load shedding events. Notably, the DFL forecasters exhibit higher statistical errors than their accuracy-trained counterparts yet achieve consistently lower operational costs, confirming that forecast accuracy and decision quality do not share a monotonic relationship in constrained microgrid scheduling.
Existing governance approaches in decentralized financial networks rely on manual review processes conducted by human committees, which intr… (see more)oduces subjective bias, creates a scalability bottleneck under high network volume, and fails to capture complex behavioural patterns and evolving network dynamics. To address these limitations, we propose \name, a federated learning-based pairwise ranking framework in which multiple institutions collaboratively train a ranking model without sharing their private transaction data. This collaborative learning enables the model to capture network-wide behavioural patterns that signify strong network participation and directly contribute to the network effect, including diversity of counterparties, usage of multiple applications, and frequency of transactions. The proposed system ranks providers based on the authenticity of their actions, promoting genuine contributors and demoting those exhibiting suspicious behaviour. We evaluate \name\ on a real-world blockchain transaction dataset. Results demonstrate that the federated approach achieves ranking performance modestly below the centralized baseline (NDCG@15:
Computer-use agents are capable of completing complex tasks, increasing the use of automatic judges to determine success, either for trainin… (see more)g or for evaluation without human involvement. Despite their flexibility, their reliability on long tasks spanning multiple applications remains unclear. A trajectory, composed of long sequences of screenshots and actions, may appear complete, but in reality violates constraints from the instruction or introduces an unwanted side effect. To identify these errors, a judge needs to examine the trajectory with respect to the user's instruction. To this end, we introduce AgentHorizon, a benchmark of 1,373 computer-use tasks (instruction-trajectory pairs) drawn from 166 hours of human-recorded trajectories spanning three operating systems. By recording trajectories for closely related instructions, we can construct negative tasks by swapping the instructions. This paired design evaluates judges on their ability to distinguish a successful trajectory from one that completed a similar (but incompatible) request. We release the benchmark under three splits: a frontier split, AgentHorizon (AH), a simplified split, AgentHorizon-Simple (AH-S), and a development split, AgentHorizon-Development (AH-D). We further evaluate eleven judges by (1) directly passing the full trajectory (with up to 300 screenshots and actions), and (2) using them as coding agents across five agent harnesses. We find that our best agentic judge, GPT-5.5, achieves 80.9% balanced accuracy on the AH subset. We find that tool-use improves certain models but results in worse performance for open-weight models, and that judges differ drastically in their ability to accept a valid trajectory and reject failed ones. Our findings highlight the need for judges that are capable of locating and verifying often hidden evidence that a task was properly completed inside long interaction histories.
Assessing psychological, neural and AI-based representational geometries across multiple similarity properties
Johan Alejandro Gamba
Megan A. K. Peters
Vincent Taschereau‐Dumouchel
Brian Odegaard
Assessing similarity between objects is a fundamental cognitive process, and similarity judgments can assess many different attributes, incl… (see more)uding objective properties (color, shape) or subjective appraisals (preference, scariness). Here, we investigated how psychological similarity across eight different properties relates to neural representational geometries and artificial models. We first collected large-scale pairwise similarity ratings for 280 objects. Using representational similarity analysis (RSA) and multidimensional scaling, we identified three mental representational spaces: one defined by multiple object features and general similarity, a second by affective evaluations, and a third by color. Combining fMRI and RSA in 80 participants, we found that object feature ratings significantly correlated with representational structures in visual and prefrontal cortices. Unexpectedly, affective ratings also correlated with patterns in visual cortices. Crucially, regional variability strongly shaped these correspondences: visual cortices exhibited high noise ceilings and consistent alignment with behavior, whereas associative regions showed greater across-subjects variability and weaker correlations. Finally, algorithms and neural networks with high classification performance on specific properties strongly correlated with behavioral and fMRI RSA matrices. Together, these results demonstrate that similarity is a multidimensional construct that comprises multiple representational spaces whose neural correspondence is shaped by representational content, regional variability, and modeling methodology.
Large audio-language models (LALMs) fuse an audio encoder into a large language model (LLM) through multi-stage training. This coupling mean… (see more)s that a new domain or a stronger LLM requires retraining, and their answers cannot be traced to what the model heard: a chain-of-thought is a post-hoc account. We propose Listen-to-Reason (L2R), an interpretable-by-design pipeline that passes audio to the LLM through an explicit, human-readable tree: small heads on frozen expert encoders map each chunk of a clip to semantically meaningful nodes on the tree (for speech, music and environmental sound), and a frozen text-only LLM answers from these nodes and an ASR transcript without hearing the clip. Every answer can therefore be traced to the nodes and transcript it read, and the nodes are causal: replacing the deciding node with a distractor overturns 78% of correct answers on SAKURA. With a 7B reader, L2R outperforms all LALMs we compare against on SAKURA and trails them by 6-12 points on MMAU and MMAR, despite training about 1,400x fewer parameters on orders of magnitude less audio data. However, because any LLM can serve as the reader, we show that a stronger reader narrows this gap without retraining any audio component. A new domain is added with one small head: with five labelled clips per species, it outperforms QLoRA fine-tuning of an LALM on the same clips by 13-26 points.
As LLM agents are increasingly deployed in high-stakes enterprise settings, understanding how they handle sensitive information in agentic d… (see more)ialogue has become critical. Privacy leakage in such settings requires reasoning about whether information flows respect the contextual norms under which they originated. We extend CIMemories to CIMEMORIES DIALOGUE, a benchmark of 4,410 two-agent dialogue instances grounded in Contextual Integrity theory. Using this benchmark, we systematically evaluate 7 open-weight models from the Qwen3 and Llama 3.1 families across conversation lengths of 1 and 6 turns under three prompting conditions. We find that task capability and privacy leakage are fundamentally coupled: more capable models are systematically leakier, with no model in our evaluation able to decouple the two. Thinking models consistently leak more than their non-thinking counterparts, a pattern we attribute to the extended reasoning surface created by chain-of-thought. Leakage compounds with conversation length, rising up to 3x from 1 to 6 turns. Our results suggest that the coupling between task capability and privacy leakage is not an artifact of misconfiguration but a structural feature of current agentic systems.
Whole slide image (WSI) analysis in computational pathology follows a multiple instance learning (MIL) pipeline where patch embeddings are e… (see more)xtracted independently and aggregated for slide-level prediction, but within-slide variance from staining, scanner, and local texture can overwhelm the discriminative signal. We propose Masked Feature Encoding for Multiple Instance Learning (MFE-MIL), a feature-space masking framework that trains a lightweight MLP adapter jointly with a window-based masked reconstruction branch and a MIL classification head. The two objectives are complementary. Classification guides the adapter to suppress within-slide patch variance, while window-based masked reconstruction provides an auxiliary regularizer for the adapted features without using patch coordinates, coordinate graphs, or segmentation preprocessing. The raster patch-extraction order is used only as a weak implicit prior. At inference, the decoder is removed, leaving only the adapter and MIL head. Across CAMELYON16/17, PANDA, and TCGA-BRCA with four diverse encoders, MFE-MIL improves ACC/F1 for nearly all tested aggregator-encoder settings and AUC in most, outperforms coordinate-based spatial methods (CAMIL), and achieves higher AUC than 2DMamba on three of four datasets (UNI). On five TCGA survival cohorts it improves the average concordance index for every aggregator tested, its most consistent gain. Code is available at https://github.com/AtlasAnalyticsLab/MFE-MIL.