Publications

2026

• MSAlign: Aligning Molecule and Mass Spectra representations for Metabolite Identification
  
  • Krzakala Paul
  • Melo Gabriel
  • Lançon Camille
  • Laclau Charlotte
  • Flamary Rémi
  • Thévenot Etienne
  • d'Alché-Buc Florence
  , 2026. Accurately identifying metabolites i.e. small molecules from mass spectrometry data remains a core challenge in metabolomics, with broad applications in drug discovery, environmental analysis, and clinical research. We address the Molecule Retrieval task, which consists in recovering the chemical structure of a metabolite from its MS/MS spectrum given a set of candidate molecules. We make three contributions. First, we propose a unified framework encompassing recent approaches based on representation alignment and contrastive learning. Second, we introduce MSAlign, a lightweight model that achieves state-of-the art performances by aligning two frozen foundation models (DreaMS for mass spectra and MolDeBERTa for molecules) and demonstrate that a score fusion strategy further improves the performance for a very small computational cost. Third, we investigate a long-standing evaluation problem: data splitting strategies in molecule retrieval implicitly trade off data leakage against domain shift. We formalize this tension by introducing a quantitative measure of distribution shift, and use it to evaluate splitting strategies in existing benchmarks. All datasets, splits, candidate sets, and a unified implementation of MSAlign and baselines are publicly released to support reproducible research.
• A Comprehensive View of Fairness through Distributional Stability
  
  • Taturyan Gayane
  • Laclau Charlotte
  • Clémençon Stéphan
  , 2026. We view fairness as a property of distributional stability. Rather than assessing a predictor under a fixed data distribution, we study how its predictions change under perturbations that modify the composition of protected groups. A predictor is fair if it remains stable under such shifts. Under this perspective, several classical notions of fairness arise as stability with respect to specific perturbations, with the associated unfairness gap given by a Lipschitz constant of a prediction-rate functional. This formulation also yields guarantees that hold uniformly over a range of demographic compositions at test time, without requiring knowledge of the deployment distribution. It leads to a learning procedure based on convex combinations of reweighted predictors, formulated as a second-order cone program, for which we establish generalization bounds. Experiments on standard benchmarks illustrate the approach.
• TabFORGE: A Tabular Foundation Model for Generative Modelling
  
  • Jiang Xiangjian
  • Liu Mingxuan
  • Simidjievski Nikola
  • Klein Tassilo
  • Jamnik Mateja
  , 2026.
• Can AI Agents Deliver Verifiable Network-Wide Outcomes Across Authority Boundaries?
  
  • Zhang Tianzhu
  • Huang Chih-Kai
  • Qiu Meikang
  , 2026. AI agents are increasingly involved in network automation, where they can initiate configuration changes through mediated operational interfaces and assess the resulting state. Nonetheless, operational networks usually span many devices and administrative domains. Realizing an operator's intent requires coordinating agents with distinct authority scopes that define the resources they can access, the operations they can invoke, and the network state they can observe. This division limits the blast radius of an erroneous action but fragments the evidence needed to assess the network-wide outcome. Successful execution of a configuration action proposed by one agent does not establish that remote devices responded as intended or that routing changes reached the required devices. A valid observation may also become stale after a subsequent change. Before the coordinated operation can be declared complete, a trusted assurance layer must collect current observations from the required scopes and determine whether they collectively support the operator's intended network-wide outcome. To address the completion admission problem, we present EVIDENCENET, a runtime assurance layer for deciding whether coordinated agent operations have achieved an operator's network intent. Its broker collects the post-change observations required by a completion contract, and its admission gate checks that the evidence comes from the required scopes, remains current, and satisfies the task rules. A verifier agent provides an additional assessment of the observation content. Experiments on live routing networks show that post-change state checks recognize successful outcomes that configuration-action records alone cannot establish. Controlled interventions further show that EVIDENCENET rejects completion when otherwise satisfactory observations have the wrong source, have been substituted, or are stale. The collected evidence can also guide recovery from configuration faults. These results support an emerging vision of autonomous networking in which agents drive planning, configuration, interpretation, and repair, while a trusted runtime layer controls the evidence and machine-checkable conditions used to admit completion.
• Fine-tuning-Free Diffusion Model with Adaptive Constraint Guidance for Inorganic Crystal Structure Generation
  
  • de Lambilly Auguste
  • Baturin Vladimir
  • Portehault David
  • Lambard Guillaume
  • Sokolovska Nataliya
  • d'Alché-Buc Florence
  • Crivello Jean-Claude
  , 2026. Generative diffusion models have emerged as powerful tools for the discovery of inorganic crystal structures, yet steering their sampling process toward user-defined physical and chemical objectives remains challenging.</p><p>We present a computational framework that integrates adaptive constraint guidance into a pre-trained crystal diffusion model, enabling the generation of candidate structures that satisfy targeted structural and chemical requirements without model retraining. The approach incorporates differentiable constraint functions directly during sampling, providing an interpretable mechanism for expert-driven exploration of the crystal structure space. To assess the reliability of generated candidates, we introduce a multi-stage validation workflow combining descriptor-based analysis, duplicate removal, comparison with reference crystal databases, graph neural network energy prediction, and thermodynamic stability evaluation through convex-hull analysis. The framework is applied to several classes of inorganic compounds and to constraints involving atomic volume, local coordination environments, and near-neighbor structural motifs. Results demonstrate that adaptive guidance effectively redirects the sampling distribution toward structures exhibiting the desired characteristics while preserving chemical plausibility. Subsequent validation reveals which generated candidates remain viable after energetic and thermodynamic screening. The proposed methodology provides a practical and transparent strategy for incorporating expert knowledge into crystal generative models and establishes a general computational framework for constrained materials discovery.
• MusicFlux: Flexible Conditioning for Text-Guided Music Generation and Editing
  
  • Baoueb Teysir
  • Boulitreau de Menezes Martins Marques Leonardo
  • Stamatiadis Paraskevas
  • Richard Gaël
  , 2026. Music editing transforms existing recordings while preserving their original attributes. Recent approaches adapt pretrained text-to-music models for tasks like inpainting, event insertion, or restoration by conditioning them with context audio. It was recently proposed for image editing to unify generation and editing by treating the source image as in-context information. However, applying this directly to music is computationally expensive due to long audio sequences. In this work, we develop MusicFlux, a text-guided music editor built on an efficient pretrained flow-matching text-to-audio model. We introduce two conditioning approaches that trade off editing performance and computational efficiency. Experiments show that both are competitive with strong baselines, with one achieving better editing quality and the other preserving the original sequence length and substantially improving computational efficiency.
• Toward Visual Methods: Reflections on How Visuals Enrich Empirical Research
  
  • Wang Zezhong
  • Carpendale Sheelagh
  • Huron Samuel
  , 2026. In design disciplines, particularly human-computer interaction (HCI) and information visualization (VIS), researchers often need to study qualitative material that is visual in nature, like screenshots, sketches, photographs, videos, or images of physical objects and spaces. De- spite an existing literature on how to analyze visual artifacts in social science and the arts, few papers reflect on the value and the means to support such analysis in the context of information visualization or graphical user interface design. This is particularly intriguing, as visual design is often the primary channel for communicating data, and visual perception is one of humans’ main epistemic sources in the world. Over the past 25 years, we have analyzed such visual artifacts across dozens of our own qualitative studies and surveys. We present six cases, reflect on four values and practical strategies for using visual methods for analyzing visual artifacts, and introduce an open archive of visual methods for HCI and VIS at http://visualmethods.pages.dev.
• Why Question Answering Modality Matters in Visualization Studies and Tests
  
  • Cabouat Anne-Flore
  • Ge Lily W
  • Isenberg Tobias
  • Huron Samuel
  • Kay Matthew
  • Isenberg Petra
  , 2026. In this position paper, we argue that answer modality choices can affect the validity of visualization evaluations. Visualization evaluation studies frequently rely on specific answer modalities such as multiple-choice questions, open-ended numerical estimates, verbal explanations, or spatial interactions. Yet, answer modality is often treated as a neutral reporting mechanism rather than as a factor that may influence reasoning strategies, representational transformations, and ultimately measured performance and the conclusions researchers can draw from their observations. We discuss two broad classes of modality effects on participant performance: changes in participants' task-solving strategies and changes induced by representational transformations between the visual stimulus and the expected response format. We further argue that answer modality determines what evidence about participants' reasoning is preserved for analysis. We ground our reflections in observations from prior empirical work and propose practical guidance for reasoning about answer modality as part of visualization evaluation design.
• I Want More Nuance": Investigating the Use of Input Physicalizations to Facilitate Discussions in Citizen Meetings
  
  • Courtoux Emmanuel
  • Huron Samuel
  • Bressa Nathalie
  , 2026. We investigate how input physicalizations can support discussions in citizen meetings. Input physicalizations are physical data representations that allow data collection and can enable participants to externalize and represent their stances and arguments during discussions. We conducted an exploratory workshop in a citizen meeting with 24 participants. Participants first used and critiqued two provotypes during discussions and then designed new physical discussion tools in groups. Our qualitative analysis shows the role of expressivity of input physicalization discussion tools to convey nuances of participants' viewpoints. Our findings highlight the tension between the expressivity of input physicalizations and the need for structured data input to enable comparable and interpretable data visualizations. Finally, we discuss trade-offs for expressivity and polysemy in visual representations.
• An Agile Homomorphic Encryption and ML-KEM Scheme for Privacy-Preserving Health Data Processing
  
  • Resende Daniel
  • Mrabet Nadia El
  • Kühne Ulrich
  , 2026. The storage and processing of health data raise serious privacy concerns; therefore, the data must be encrypted. Fully Homomorphic Encryption (FHE) allows computations on encrypted data, which makes it a very promising technology in this context. However, although the efficiency of FHE algorithms has improved in recent years, they remain very complex. The security of several prominent FHE schemes is based on variants of the Learning With Errors (LWE) problem, which is also the basis of Module-Lattice-Based Key-Encapsulation Mechanism (ML-KEM), the new post-quantum standard for asymmetric encryption. In this paper, we propose an agile cryptographic scheme -combining ML-KEM and BGV -that allows efficient implementations in an embedded use case. We describe the parameter selection for the proposed scheme and show its feasibility with a proof-of-concept implementation.
• WiCleanData: Guaranteeing the Type Consistency of Wikidata by Taxonomy Refinement and Constraint Enforcement
  
  • Peng Yiwen
  • Jeanmougin Marc
  • Bonald Thomas
  , 2026, pp.8988 - 8992. Because of its collaborative nature, Wikidata suffers from errors, in- consistencies, and excessive complexity, such as redundant classes, ambiguity between instances and classes, wrong taxonomic paths, and type constraint violations. The manual curation of these issues is infeasible at scale. To address these challenges, we introduce WiCleanData, a refined version of Wikidata with a consistent tax- onomy and free from type constraint violations. Specifically, we have designed an automated pipeline that first cleans the taxonomy with language model assistance, then simplifies type constraints by hierarchical aggregation, and finally filters facts accordingly. The resulting knowledge graph, free from any type violation, is made publicly available via a Web interface, enabling easy exploration and downstream applications. (10.1145/3799682.3840187)
  DOI : 10.1145/3799682.3840187
• Geometry-Based Refinement of Operational Design Domains within Physics-Constrained AI-based Systems arising from Riemannian Flow Matching
  
  • Tran Binh
  • Gonzalez Martin
  • Barbaresco Frédéric
  • d'Alché-Buc Florence
  , 2026. <div><p>Operational Design Domains provide an AI constituent-level description of the conditions under which AI constituents are intended to operate safely within an AI-System. For physicsconstrained AI, however, the relationship between these operating parameters and the mathematical space in which an AI model learns and generates predictions is often left implicit, requiring manual mappings between ODD dimensions and data features. We argue that this relationship can instead be made explicit through a shared geometric formalism. Our framework maps physical requirements into manifolds and geometric structures that constrain the representation, dynamics, and implementation of an AI constituent, allowing ODD boundaries to be refined directly through mathematical and physical considerations. Riemannian Flow Matching provides an intrinsic learning mechanism for models defined on non-Euclidean state spaces, while metriplectic structures additionally encode conservation and dissipation laws into learned dynamics. We illustrate a progressive refinement process in which physical representations, manifold parameterizations, numerical domains, and empirical performance trade-offs jointly determine the operating envelope of the AI constituent. The framework is evaluated on crystal structure generation and thermodynamically consistent trajectory simulation. The applications demonstrate how geometric structure can bridge physical requirements and ODD parameters, providing an interpretable and formally grounded foundation for verification, validation, and assurance of physics constrained AI.</p></div>
• Comparative Study of a Self-Complementary Antenna Array in Transmission and Reception
  
  • Berthelot Gabriel
  • Thevenot Marc
  • Lepage Anne Claire
  • Menudier Cyrille
  • Begaud Xavier
  , 2026.
• System-level performance in massive ISAC networks
  
  • Mengoli Emanuele
  • Soprano-Loto Nahuel
  , 2026. We propose a dynamic stochastic geometry framework to model and evaluate large-scale integrated sensing and communication (ISAC) networks. Base stations are distributed according to a homogeneous Poisson point process, while user equipments and sensing objects undergo localised random motion around their serving base station. Full spectrum reuse with a common waveform generates mutual interference that couples the two functionalities. Communication performance is characterised through downlink queues whose service rates are governed by the signal-to-interference-plus-noise ratio (SINR) via the Shannon capacity. Sensing performance is captured by a Kalman filter (KF) whose measurement noise covariance depends on the sensing SINR, yielding a non-classical state-dependent filtering problem. We define system-level performance metrics under the Palm distribution of the base station process and prove the consistency of their empirical estimators. We then prove that, for a typical base station, communication and sensing metrics are associated, so that operating regimes favourable to one functionality tend to benefit the other. The proof combines the monotonicity of the queue and filter recursions in the configuration of interferers with the Harris--FKG inequality for Poisson functionals. We also show that the excess association ratio of the two interference levels decays as the inverse of the network density. Simulations based on a ray-tracing channel model corroborate the analysis.
• Low-Profile Huygens Unit Cell for a Sub-7 GHz MIMO Antenna
  
  • Medrar Ghiles
  • Lepage Anne Claire
  • Begaud Xavier
  , 2026.
• Data Processing Inequalities for the Quantum Guesswork
  
  • Mejdoub Ilyass
  • Béguinot Julien
  • Rioul Olivier
  , 2026. <div><p>Quantum guesswork has emerged as a crucial measure in assessing the distinguishability of nonorthogonal quantum states. In this paper, we review and compare previous definitions of the quantum guesswork, and establish the pre and post data processing inequalities (DPIs) that are relevant for the guessing problem. We also derive improved lower and upper bounds on the quantum guesswork in terms of the Shannon entropy and Holevo's information.</p></div>
• Beyond Kemeny Medians: Consensus Ranking Distributions. Definition, Properties and Statistical Learning
  
  • Clémençon Stéphan
  • Irurozki Ekhine
  , 2026.
• Secrecy Metrics over Random Gaussian SIMOME and MIMOME Channels
  
  • Leroy Clément
  • Arbi Tarak
  • Geller Benoit
  • Rioul Olivier
  , 2026. <div><p>The achievable secrecy of time-invariant wireless channels is typically characterized by the secrecy capacity of the communication system, whose mathematical expression is well established. However, in most practical scenarios, channels cannot be assumed to remain constant or to be perfectly known at both the transmitter and the legitimate receiver. In such cases, no closed-form expression for the secrecy capacity is available.</p><p>To address this, several metrics are studied, that characterize secrecy in random channel environments: Ergodic Secrecy Rate (ESR), Ergodic Positive Secrecy Rate (EPSR), and Secrecy Outage Probability (SOP). We derive the optimal values of these metrics for SIMOME channels (Single-Input, Multiple-Output, Multiple-Eavesdropper) and various MIMOME (Multiple-Input, Multiple-Output, Multiple-Eavesdropper) scenarios.</p></div>
• Decentralized Ranking Aggregation via Gossip: Convergence and Robustness
  
  • Le Caillec Kerrian
  • van Elst Anna
  • Colin Igor
  • Clémençon Stéphan
  , 2026. The concept of ranking aggregation plays a central role in preference analysis, and numerous algorithms for calculating median rankings, often originating in social choice theory, have been documented in the literature, offering theoretical guarantees in a centralized setting, i.e., when all the ranking data to be aggregated can be brought together in a single computing unit. For many technologies (e.g. peer-topeer networks, IoT, multi-agent systems), extending the ability to calculate consensus rankings with guarantees of convergence and resilience to potential contamination in a decentralized setting, when preference data is initially distributed across a communicating network, remains a major methodological challenge. Indeed, in recent years, the literature on decentralized computation has mainly focused on computing or optimizing statistics such as arithmetic means using gossip algorithms. The purpose of this article is precisely to study how to achieve reliable and resilient consensus on collective rankings in a decentralized setting, thereby raising new questions, robustness to corrupted nodes, and scalability through reduced communication costs in particular. The approach proposed and analyzed here relies on the robustness guarantees offered by random gossip communication, which allows autonomous agents to compute a global ranking consensus using local interactions only, without coordination or a central authority.</p></div>
• Unfenced: Atomic Guards Against Spectre: ISA-Level Lock-Free Mitigation for Multi-Core RISC-V Processors
  
  • Awais Muhammad
  • Mushtaq Maria
  • Naviner Lirida
  • Haj-Yahya Jawad
  • Bruguier Florent
  , 2026, pp.1-6. Speculative execution attacks, such as Spectre Variant 1, leak sensitive data through transient microarchitectural effects. This paper introduces batched atomic guards: ISA-level software wrappers that use the lr.w.aq/sc.w.rl acquire/release pair to re-validate a security bound once per N = 16 iterations on a dedicated cache-aligned guard cell, in place of fence barriers. On the SiFive HiFive Premier P550 quad-core board, a fence wrapper imposes a 3.4 × execution-time overhead (25.41 ms vs. 7.50 ms unwrapped); the batched guard reduces this to 1.2 × (8.93 ms), 2.8 × faster than the fence wrapper. The batched guard records 0.00% SC failures when the attacker writes a separate variable rather than the dedicated guard cell; we also characterise two non-batched variants and report a counterintuitive result: moving the index-mask computation outside the LR/SC window increases the SC failure rate from 11.07% to 19.98% because of higher loop throughput, despite a narrower individual reservation window. Benign co-located core IPC is unaffected (2.601) across all five tested modes. We give a data-dependency argument for why an acquire load serialises the dependent secret load, and report P550 hardware experiments showing lr.w.aq never observes an uncommitted remote store and costs 13 ns more than a plain load; our platform’s timer resolution does not permit a direct cache side-channel leakage measurement. A four-class classifier trained on 1,480 gem5-derived execution windows distinguishes unwrapped, fence-hardened, and lock-free-hardened execution with 97.4% accuracy. (10.1145/3857307.3857311)
  DOI : 10.1145/3857307.3857311
• Sequential State-Dependent Kalman Filtering for Dynamic Parameter Estimation of a Mobile User in a Large Cellular Network
  
  • Soprano-Loto Nahuel
  • Balakrishnan Ashutosh
  • Baccelli François
  , 2026. We estimate the position, the velocity and the shadowing gains of a mobile user with the base stations (BSs) of a large cellular network modelled by a Poisson point process. At each slot, the strongest BS senses the user and sends its measurements to a controller, which runs a sequential extended Kalman filter (EKF), whose measurement noise covariance depends on the state through the echo signal-to-noise ratio (SNR). As a network metric, we use the Cesàro average of the error covariance estimated by the filter. In a two-dimensional setting, the state-dependent covariance lowers the position error by 23% with respect to a classical EKF, and keeps the estimated variance consistent with the actual error. Next, we compare two measurement reports for the estimation of the shadowing gain: (i) the echo SNR only, and (ii) the echo SNR with the delay, which is a sensing-aided communication (SAC) framework. With SAC, the shadowing gain estimation error falls from 1.9 to 0.25dB, equivalent to that of a filter that knows the true user position. The SAC framework also saves up to 2.4dB of transmit power for a downlink scheduled ahead under slow shadowing, with respect to the last measured gain. As the user moves across the network, the error peaks of the two reports alternate: without the delay, they occur close to the BSs; with it, near the cell boundaries. The engineering insights obtained through the proposed framework are expected to be useful for designing initial access protocols with joint sensing and communication.
• SAS: Scene-Annotated AudioSet Strong for Joint Scene-Event Analysis
  
  • Olvera Michel
  • Stamatiadis Paraskevas
  • Serizel Romain
  • Essid Slim
  • Richard Gaël
  , 2026. Context is central to human auditory perception, as listeners interpret sounds within their acoustic environment to resolve ambiguity. However, large-scale datasets that jointly capture sound events and acoustic scenes remain limited. We address this by augmenting AudioSet Strong with clip-level scene annotations under two taxonomies for joint sceneevent modeling. A corpus-scale analysis shows that scene identity is better captured by events distinctive to a scene than by those that merely occur frequently. Reweighting event co-occurrences toward scene-specific associations produces representations that better align with human-derived taxonomies, while also exposing scenes where the acoustic and semantic views diverge. This geometric divergence predicts classification errors, while supervised validation confirms that the proposed scene annotations provide meaningful complementary information for developing scalable context-aware machine listening systems.
• Confidence-Aware Aggregation for Multimodal Audio Reasoning
  
  • Olvera Michel
  • Stamatiadis Paraskevas
  • Richard Gaël
  , 2026. Large Audio Language Models (LALMs) enable generative reasoning over complex acoustic scenes but often exhibit prediction instability. We show that this variability can serve as a reliability metric via inference-time self-consistency and confidence-aware aggregation. Using a log-probability–based scoring framework, candidate answers are weighted by model likelihoods, leveraging reasoning diversity and self-certainty without additional training. Experiments on four LALMs, including Qwen3-Omni and Audio Flamingo 3, demonstrate that confidence-aware aggregation improves accuracy and calibration, with additive methods like Borda voting outperforming classical Majority Vote, especially under high disagreement. This approach transforms LALM variability into a source of reliability for scalable and robust multimodal audio understanding. Project code is available at https://github.com/michelolzam/lalm-self-consistency.
• Probing Large Audio-Language Models for Compositional Understanding of Sounding Actions
  
  • Olvera Michel
  • Stamatiadis Paraskevas
  • Wang Changhong
  • Richard Gaël
  , 2026. Large audio-language models (LALMs) excel at understanding and reasoning tasks over atomic sound events, yet their ability to infer higher-level human activities from such fine-grained events remains largely unexamined. Everyday human actions and activities, such as setting a table, cleaning the house, or preparing a breakfast emerge compositionally from temporally distributed sound events, requiring abstraction beyond the event-centric granularity that dominates current training and evaluation paradigms. Our benchmark evaluates a wide set of LALMs under a principled framework that tests how language-based reasoning, grounded in acoustic perception, structures sound abstractions into higher-level understanding. By systematically varying exemplar typicality and distractor similarity, our evaluation exposes that current models do not reliably perform compositional inference from atomic acoustic events to higher-level human activities solely from audio. All data, taxonomies, and evaluation scripts are publicly available on our companion website: https://alm-sounding-actions.onrender.com
• From Disclosure to Evaluability: an NLP-Based Analysis of Corporate Transition Plans
  
  • Calamai Tom
  • Balalau Oana
  • Le Guenedal Théo
  • Suchanek Fabian M
  , 2026. <div><p>Corporate climate transition plans detail how companies intend to reduce their carbon footprint, but their unstructured and often incomplete nature limits assessments of whether disclosed actions are sufficiently concrete, quantified, and consistent with stated emissionreduction objectives. We analyze transition plans from CDP by decomposing them into the key information needed for such assessments: disclosed mitigation measures, their implementation status, quantitative attributes, action type, specificity, and organizational perimeter. Large language models achieve moderate to strong performance across these tasks (65-91% F1), enabling large-scale diagnostic analysis of transition-plan disclosures. Our results show that the main bottleneck lies in the disclosures themselves. Specifically, 80% of transition plans contain at most two quantified planned measures, and nearly 47% of proposed actions are both generic and lack a specified organizational perimeter. Firms also tend to favor high-level targets, preparatory actions, and energy-related measures over more structural business transformations. Overall, our findings show that current transition-plan reporting often lacks the granularity needed to assess whether disclosed actions are consistent with stated emission-reduction targets.</p><p>SOTA open-and closed-source LLMs. Third, we apply the resulting models to 2,283 CDP transitionplan entries across 2023 and 2024, extracting over 15,000 mitigation measures, and use a bootstrapbased uncertainty propagation procedure to account for measurement error from imperfect model predictions. This provides a large-scale empirical diagnosis of what current climate disclosures contain and what they omit.</p><p>Findings. We demonstrate that Large Language Models (LLMs) can effectively extract and classify complex climate disclosure data, achieving a performance between 65.4% and 91.1%. Applying these models shows that the main bottleneck for transition-plan credibility assessment lies in the disclosures themselves rather than only in the availability of automated analysis methods. Specifically, 80% of transition plans contain at most two quantified measures, and nearly 47% of proposed actions are too generic to evaluate their climate impact. Firms largely favor high-level targets and superficial energy modifications over concrete emissions reductions and structural business transformations, a trend consistent with the carbon lock-in phenomenon (Tan and Chen, 2025). Finally, we find that reporting standards are driven more by regional pressures than industry sector, with European firms providing significantly more detailed plans than their US and Japanese counterparts.</p><p>The datasets, prompts, and scripts used in this study are publicly available at https://gitlab.inria.fr/cedar/ transition-plan-evaluability.</p></div>