Supervised fine-tuning (SFT) on expert trajectories is a standard paradigm for training LLM agents. However, it suffers from a well-known training-inference distribution mismatch, calling for target-aware data synthesis strategies that ground the training data in the target model's native state distribution. To bridge this gap, emerging approaches incorporate the expert either to guide mixed rollouts alongside the target model, or to provide dense step-wise supervision. Nonetheless, existing methods apply expert assistance blind to the target model's inherent behavior: they either disrupt the rollout using sub-optimal random intervention heuristics, or they enforce indiscriminate uniform supervision that unnecessarily overwrites the model's already-correct actions. We introduce JudgeMix, a judge-guided data synthesis framework that dynamically interleaves the target model and expert based on step-wise semantic evaluations. JudgeMix queries expert guidance precisely when the target model struggles, ensuring high trajectory success rates while keeping the state distribution aligned with the target model. It additionally uses the judge evaluation to reweight expert supervision on different steps, preserving model's inherent valid behavior while correcting its errors. On two agentic tool-use benchmarks, AppWorld and AutomationBench, JudgeMix demonstrates consistent superior performance to existing methods, with improvements up to 10 pass@1 points.
Governments are making decisions about AI before they can measure its public effects, and those decisions determine whether a person can learn that AI influenced an outcome, challenge it, and see it corrected. Research has documented what AI governance produces: principles, strategies, registers, and audits. However, we lack shared language for understanding AI policy while it is still being made. We introduce the Society-Centered AI Governance Framework, which gives policymakers, researchers, and the public a way to assess instruments in any jurisdiction by whether a person can get notice, challenge a decision, and see it corrected. It reads a jurisdiction's instruments together by date and binding status, separates an agency duty from a right a person can invoke, and traces which later instrument must enforce each commitment. We evaluated it in North Carolina using interviews with nine AI Leadership Council members, conducted before and after the state's 2026 AI roadmap, and a provision-level reading, with human-reviewed model-assisted coding, of 49 AI governance documents from 19 states. For the state's own AI systems, we find visibility-first governance: visibility duties are in effect, notice exists as an agency duty with no right a person can invoke, and AI-specific challenge and correction remain advisory. The roadmap assigns owners and dates to the missing protections, so the gaps are documented but not yet closed. Recent federal preemption proposals focus on AI development, so recourse for a state's own systems is likely to remain a state decision. We identify "legitimacy before protection" as a testable risk of visibility-first governance and name the milestones that would test it. More broadly, judging AI policy by who can act on what it makes visible lets the field assess recourse while policy is still forming, before harm occurs.
Chemical mechanical planarization (CMP) is a critical semiconductor manufacturing process where material removal rate (MRR) is a key metric for process monitoring and advanced process control (APC). Virtual metrology enables MRR estimation from process time series signals, supporting run-to-run(R2R) control and process optimization while reducing reliance on physical metrology. Existing approaches typically rely on simplified linear assumptions or require handcrafted feature engineering. In this paper, we propose Global Regression Pretraining(GREP), a self-supervised, attention-based framework for CMP MRR prediction. GREP learns temporal representations from unlabeled sensor signals through global proxy-target regression, followed by supervised fine-tuning for MRR estimation. The framework is evaluated against a feature-engineered supervised baseline under three conditions: limited labeled data, sensor noise, and temporal process drift. Experiments on the PHM 2016 CMP dataset demonstrate competitive performance under limited labeled-data conditions, with improvements in selected cases and performance largely comparable to the baseline across different labeled-data fractions. Both models exhibit relatively small changes in prediction error with increasing noise, while GREP achieves a 17.0% RMSE reduction in the final temporal interval.
We present Fikiria, a two-stage SFT+GRPO recipe for Swahili mathematical reasoning. Its reward multiplies a format gate by a weighted sum of answer accuracy and the ChrF\texttt{++} similarity of the Swahili reasoning to an in-language reference chain of thought, so that training targets the surface fidelity of the reasoning trace and not only the final answer. We train on a 2.3k-problem Swahili dataset created by translating GSM8K with GPT-4.1 and filtering with SSA-COMET. On AfriMGSM Swahili, Fikiria-AfriqueQwen-14B reaches 79.2\%. Under the same fixed prompt and strict answer-format scoring, GPT-4o reaches 68.0\%, partly because it often fails the required answer format; IrokoBench reports 78.8\% for GPT-4o using a best-per-model prompt, a figure numerically close to ours but not directly comparable. In a small human evaluation with co-author annotators, GPT-4o's outputs were preferred overall. ChrF\texttt{++} measures surface fidelity, not the validity of individual reasoning steps.
Subsurface schematisation supports geotechnical decisions but remains challenging because sparse site-investigation data leave substantial gaps in information about ground conditions. This study explores TabPFN, a foundation model pretrained on general synthetic tabular data, for this task. Using only grid coordinates and soil behaviour type index observations from sparse cone penetration tests (CPTs), we first reconstruct two-dimensional sections without geotechnical fine-tuning or predefined layer boundaries. We evaluate the approach on synthetic sections and at two field sites in the Netherlands, comparing it with SchemaGAN, a conditional generative adversarial network trained on synthetic geotechnical sections, and with reference interpolation methods. In the baseline comparisons, TabPFN achieved mean errors close to SchemaGAN's on synthetic sections and lower mean errors at both field sites. On synthetic sections, TabPFN's mean absolute error decreased at each tested increase in CPT count, and, under the baseline input setting, its pointwise prediction intervals had mean coverage close to nominal levels. Subsequent fine-tuning on the synthetic sections used to train SchemaGAN improved synthetic test accuracy but increased mean errors at both field sites. These findings indicate that general-purpose tabular foundation models offer a promising route for updating subsurface predictions from sparse and evolving site-investigation data without retraining.
We explore the use of AI agents for solving computational problems in algebra. For the problems we consider, the main challenge is not correctness but \emph{time}: general-purpose algorithms are guaranteed to provide correct solutions but can be prohibitively slow, while reformulations that exploit special structure often allow solving a problem much faster. We propose to use this setting to evaluate how effectively agents discover and exploit structure within and across tasks to reduce computational costs. More broadly, computational algebra supplies infinitely many problems with independently verifiable answers and systematically controllable structure and difficulty. To illustrate these possibilities, we use benchmark families of systems with controlled shared structure and evaluate agents in three settings: solving independent instances, solving online streams of related instances using persistent memory, and compiling a fast reusable solver for a known problem family. Our experiments show that (i) agents can extend the range of systems solvable within a fixed budget, although reasoning overhead can sometimes outweigh computational savings; (ii) memory can reduce the cost of repeated solving but may preserve suboptimal strategies; (iii) compilation produces faster reusable solvers but at substantial upfront cost. Overall, computational algebra offers concrete, measurable targets for improving how agents discover and reuse structure.
Federated learning (FL) systems increasingly rely on blockchain based reward mechanisms to incentivize participation while preserving data privacy. We design and empirically evaluate two complementary Solidity smart contracts for FL reward distribution: the first one that allocates rewards proportionally to a publicly disclosed quality score, and a privacy preserving variant that hides individual quality scores behind a zero-knowledge (ZK) proof of a threshold condition. Both contracts mint a non transferable, use-and-burn token maintaining a trustless incentive structure. We evaluate the two designs through MNIST training with adversarial behavior, including Sybil attacks, zero gradient and random noise, across three independent random seeds. Our results reveal a fundamental tradeoff between cryptographic privacy and incentive fairness. The privacy preserving architecture strictly degrades system robustness, doubling the adversary's average reward capture while introducing severe outcome variance. This vulnerability is driven by a structural self-inclusion bias within cosine-similarity scoring: as the global model converges, legitimate gradient updates naturally diminish, rendering them statistically indistinguishable from bounded noise. Adversaries exploit this dynamic to systematically breach the ZK threshold, triggering a late stage failure cascade that culminates in absolute adversary monopoly over the reward pool and the total exclusion of honest contributors. Alongside a closed-form theorem formalizing this privacy--Sybil resistance tradeoff, we map strict mainnet deployment boundaries via linear gas scaling models, explicitly incorporating the post-Fusaka EIP-7825 per transaction gas cap. Ultimately, while cryptographically secure, static threshold ZK gates fail as incentive layer fairness mechanisms under realistic FL convergence dynamics on standard benchmarks.
There is no best worker in general, only a best worker for a particular firm. We develop a measure of worker-firm match quality that treats fit as a property of the pair, learned from the survival of employment relationships. A two-tower neural network embeds workers, from their career histories and current positions, and firms, from their characteristics and business descriptions, into a common space, and match quality is the cosine similarity of the two vectors. The measure takes its functional form from the skill-weights model of firm-specific human capital, and its only estimated objective, a separation hazard, from the view of the job match as an experience good. Estimated on more than a billion person-periods covering 101 million workers and 8.4 million firms, a 0.1 increase in match quality lowers the odds of separation by 48%. About a quarter of the variance of match quality is specific to the worker-firm pair, which worker-, firm-, and occupation-level measures set to zero by construction. Workers earn more when better matched, especially early in a relationship, and firms whose workforces score higher grow faster, become more profitable, and are issued more, and more valuable, patents over the next five years than their industry peers. Because it scores pairs that never formed, the measure also predicts where movers go, and acquirers record a larger share of the price as goodwill when the target's workers would fit the acquirer better than they fit the target.
Genetic Programming (GP) is a search paradigm inspired from natural evolution for the automated discovery of expressions, functions, and computer programs. Cartesian GP (CGP) is a flavor of GP that uses a graph-based model to encode candidate programs, as contrasted to the conventional Tree-based GP (TGP). Since its inception two decades ago, CGP has predominantly been analyzed empirically, thus little is known regarding its theoretical performance guarantees. This paper analyzes CGP for the Max problem, which has been only rigorously studied for TGP with a proven expected runtime superlinear in the optimal program length $n$. We prove that the expected time for the ($1$+$1$)~CGP employing common mutation operators to evolve an equivalent optimal of the same output is polylogarithmic in $n$, specifically at most $O(\log^9{n})$. This is remarkable, as it shows an exponential speedup by switching to CGP. Our results shed light on the benefit and the compactness of the graph-based representation for programs, and on how CGP can navigate its search and fitness spaces efficiently. This is a first proven performance guarantee for CGP on an established benchmark problem. Experiments complement our theoretical findings.
Language moved from many task-specific models to one unified model: a single generative model with a single generative training task that handles all of them within the same backbone, and keeps improving with data, parameters and compute. We believe the bitter lesson applies just as much to computer vision, 3D and spatial understanding, in the end to visual intelligence: the effort should go into general architectures that solve all these tasks in a unified way, and into scaling them, rather than into task-specific designs. We test this on monocular geometry, a task that has belonged to specialised architectures for years. Iris-Z is a general image generator that produces appearance and geometry together. A single backbone denoises, reasons about and understands the image and its depth at the same time, and one set of weights generates images with their depth, estimates depth from photographs and paints images onto a given depth. At enough scale it surpasses state-of-the-art specialists on established depth benchmarks while keeping the generative abilities of the backbone it started from without specialized layers. We release two models for the community.
State-of-the-art neural network verifiers evaluate linear relaxations over axis-aligned bounding boxes. However, verification domains in cyber-physical systems are rarely boxes. Reachability specifications, obstacle-avoidance constraints, and neural Lyapunov certificates typically yield oblique or non-convex polyhedral regions, over which box-based bounds become extremely loose. This paper proposes Fit-LRP (Fitting Linear Relaxations to Polytopes), a framework the implements exact optimization of affine bounds over non-convex domains (represented by unions of polytopes) by solving a sequence of small, low-dimensional linear programs. Our approach seamlessly enhances the final optimization step of LiRPA and $\alpha,\! \beta$-CROWN branch-and-bound (BaB) algorithms. In BaB verification, approximating oblique constraints with axis-aligned splits causes severe combinatorial blowup and risks non-termination. By optimizing the final affine bound of a nonlinear function over the exact polyhedral state-space constraints our method eliminates box-induced over-approximations, enabling earlier search-tree pruning. On ARCH-COMP and neural Lyapunov benchmarks, Fit-LRP reduces branching on the hardest region by $47\times$, cuts worst-case verification time by $15\times$, and with region-wise partitioning, verifies on a single CPU $6.1\times$ faster than $\alpha,\! \beta$-CROWN on a GPU.
Applications built on large language models assign the model a role through its system prompt, intending to shape how it responds for their use case. We introduce a validation framework for extracting and injecting occupational role vectors, activation-level directions that capture how someone in a given occupational role would distinctively respond to a task or question. We validate these vectors through two complementary lenses, on Qwen2.5-7B-Instruct and Llama-3.1-8B-Instruct. The interpretability lens asks whether each vector is a stable, separable, domain-consistent representation of its role: extracted vectors are highly reproducible across repeated runs and sampling temperatures, remain separable even for very similar roles, and agree more with themselves across in-domain and out-of-domain questions than with other roles' signal. The control lens asks whether injecting a role vector during generation makes the model express the role, and sustain it over a conversation, as well as a role prompt does. On Qwen, injecting the vector alone makes the model express the intended role substantially more often than unsteered, but at best matches a role prompt; on both models, adding the vector to the prompt strengthens the role in single answers but did not slow the role's decline over 30 turns. Together, the two lenses show why both are needed: out of domain the role is expressed only subtly, yet the activations still encode it; and a vector that is stable and separable can still barely steer the model's behavior.
Learning from paired modalities is commonly equated with aligning them: two encoders are trained until the embeddings of a pair coincide in a shared space. We study whether cross-modal retrieval and classification can instead be obtained while the two encoders' own embedding spaces remain directly unaligned. DuoJEPA trains one encoder per modality with the self-supervised LeJEPA objective and appends to each encoder a single transformer layer that translates its embedding into the other encoder's space, supervised by a contrastive objective computed entirely within that target space. No loss term compares the two own spaces directly, although the encoders are optimised jointly through the translation objective. In our experiments, the translator's residual connections determine where the cross-modal pressure is absorbed: with a residual translator the two encoders become directly aligned, whereas with a skipless (residual-free) translator a paired-cosine test detects no alignment in the final models, and at most a weak one in the ablations, at comparable cross-modal performance. The learned class-token translation is well approximated by a single scaled rotation of the source space. We evaluate DuoJEPA on two modality pairs that differ in nearly every respect. On chest radiographs and reports from CheXpert Plus, it attains a zero-shot AUC of 86.6 and a cross-modal retrieval score of 78.5, compared with 60.0 and 64.2 for our reimplementation of LeVLJEPA with the same encoders and data. On Sentinel-2 optical and Sentinel-1 radar imagery, the frozen encoders fall within the range of published Sentinel foundation models under our reproduction of the Copernicus-Bench linear-probe protocol, the translator retrieves the paired image of the other sensor among 4,000 candidates in 77-79% of cases, and a patch-level extension transfers the within-image similarity structure of the scene.
Software is increasingly written by AI agents that produce an implementation from a natural language specification, a practice known as agentic coding. Code is now produced faster than it can be reviewed, and has been shown to exhibit its own characteristic failure patterns. Property-based testing (PBT), a long-standing technique in which invariant properties of a program are checked against many generated inputs, has recently been revisited with LLM agents authoring the properties, thereby removing the high-effort derivation of invariants that has long hindered its adoption. Existing methods, however, derive these properties from the implementation alone (Code-driven PBT). At the same time, a thorough specification such as a product requirements document has become routine within agentic coding, yet its role in testing the resulting code remains underexplored. We study the value of the specification as a source of properties for testing agent-written code. Our key insight is that the specification, unlike the implementation, records what the code was meant to do, and invariants derived from it are not biased by particular implementation choices. We propose Spec-Driven PBT, an agent that first expresses invariants in natural language from the specification and only then inspects the implementation to translate them into concrete tests. Across 25 applications from ViBench, Spec-Driven PBT finds a novel set of bugs that Code-driven PBT misses, while writing 2.4x fewer property tests. Though the two approaches surface a roughly equivalent number of bugs, there is little overlap: nearly 60% of the 169 total bugs are found by only one. Qualitatively, we find that the bugs unique to Spec-Driven PBT tend to violate documented product workflows whereas those to Code-driven PBT instead concentrate on implementation robustness, offering complementary value for bug discovery in agent-written code.
Cultural heritage sites are increasingly monitored by sensor networks, but querying their data requires database expertise. Large Language Models (LLMs) offer a natural-language interface, yet over raw sensor records they lose track of which sensor, site, and time each value belongs to, and silently miscompute. We present HeriKGQA, a knowledge-graph question-answering method that confines the LLM to interpreting questions and phrasing answers, leaving computation to SPARQL and deterministic code. Questions are parsed into intents constrained by a heritage ontology aligning W3C SOSA/SSN with CIDOC-CRM, compiled into SPARQL, and answered from query results and temporal, spatial, and anomaly facts derived by a grounding layer, with error correction guarding each LLM step. On 500 questions over real data from five heritage sites, HeriKGQA reaches 99.0% accuracy, 18.2 points above the strongest of three baselines using the same LLM and receiving the same anomaly rules and answer conventions, with comparable answer quality while running almost four times faster. Without grounding, accuracy drops by 12.2 points and on anomaly questions from 99% to 66%. We report lessons from development and live deployment.
Recent advances in agentic AI have demonstrated the potential to automate complex high-energy physics (HEP) workflows. We focus on the regime of HEP tasks where frontier models still struggle to provide correct and complete solutions, and introduce an open-ended optimization task format with continuous, automatically evaluated rewards. Unlike conventional benchmarks focused on task completion, these tasks target reinforcement learning (RL) environments to generate agentic improvement on state- of-the-art HEP challenges. We benchmark frontier reasoning models on several HEP tasks and demonstrate an initial proof of concept in which a small language model improves through RL using automated feedback without human intervention. By connecting challenging HEP problems with verifiable scientific rewards and RL, we aim to advance agentic AI from reproducing established workflows to autonomously discovering better scientific solutions, paving the way toward self-improving scientific agents.
Generative AI agents are increasingly being embedded into visualization systems that can search data, invoke tools, summarize evidence, and produce recommendations. However, when agents rely solely on their own knowledge, the visualization problem is not only whether the agent provides a useful answer, but also whether users can inspect how that answer is grounded. We present Food Critic, an agentic visualization system for diners' restaurant decision support that integrates public health inspection records with crowd review data. The system grounds agents under three conditions: a database-only ReAct agent, a retrieval- augmented agent with access to food-safety regulations, and an ontology-guided agent that produces structured hypotheses and typed evidence. Food Critic externalizes the agent's reasoning within these groundings as inspectable visual evidence. We evaluate LLM models with several use cases under each grounding of our agentic system. Our work positions restaurant choice as a concrete case for studying human-agent collaboration, visual provenance of agent reasoning, and the design of visual interfaces for grounded GenAI decision support.
Automated bilateral negotiation underlies two-sided matching markets such as freight brokering, where many shippers and carriers negotiate price before a matching mechanism decides who is paired with whom. Prior strategy comparisons typically use isolated one-on-one duels or a single fixed market, so they cannot show how a strategy behaves against several live rivals under a changing balance of supply and demand. We built nine negotiation architectures spanning rule-based, search-based, optimization-based, imitation-learning, and reinforcement-learning families, and ran all nine simultaneously on both sides of a Gale-Shapley matching market under a balanced condition and two scarcity conditions rotated across all 36 exclusion pairs. In the balanced carrier market, an imitation-learned engine (ML2), a reinforcement-learning policy trained only on isolated negotiations (RL), and an expected-utility optimization engine (OPT) formed a statistically tied top group, and no architecture led in every condition and role. RL beat all eight carrier rivals when carriers were scarce, but its match rate collapsed from 62% to 16% when shippers were scarce and it had to compete with the other carrier agents for demand. A second reinforcement-learning policy trained against rival agents (RL2) kept its match rate above 91% in every condition, yet lost on realized payoff to all eight shipper-side rivals in two of the three conditions, traced to a confirmed train/deployment feature mismatch. We also tested nine large language models (LLMs), from 1 to 72 billion parameters, under the same protocol: they lost significantly to the two strongest engines (OPT2 and ML2) in 32 of 36 comparisons and won none. These results show that isolated or single-condition evaluation can hide failure modes that a many-sided, scarcity-varying, statistically corrected tournament exposes directly.
We introduce the BEATS dataset, a robust compilation of psycho-physiological data collected in the wild from 30 participants while they listened to curated playlists intended to promote well-being. An acronym for Biomarkers, Experience and Affect from Therapeutic Soundscapes, the BEATS dataset aims to support research on wellbeing, music psychology, affective computing and signal processing. More specifically, this dataset includes: (1) sensor data collected during music listening sessions using a research-grade wearable device in free-movement environments; (2) music listening activity records, stimulus metadata, acoustic features, and self-reports associated with the listening sessions (3) individual characteristics, such as demographics, music preference, musical skills, and personality traits; and (4) self-reported psychological measures before and after their experimental participation spanning for a number of weeks. By situating data collection outside of lab settings, the BEATS dataset offers the opportunity to study real-world music-listening scenarios, allowing for a richer exploration of the physiological, subjective and therapeutic effects of music under real-world conditions.
In pre-deployment testing of AI systems such as AI agents, a developer can often deal with known failure modes (e.g., leaking API keys, deleting important databases, or installing a malicious package) by implementing fixes or making contingency plans for when the same issue might occur again. What is more challenging are the unknown unknowns: the unseen failure modes that haven't surfaced during testing. The decision to stop testing and deploy therefore faces an inherent tradeoff: continue testing to observe more behaviors, at the cost of time and resources, or stop testing and deploy, at the risk of failure modes remaining undiscovered. How then does the developer decide when they have tested enough? To answer this, we develop an algorithm that certifies when the probability of observing an unseen failure mode is sufficiently small. The algorithm constructs an anytime-valid upper confidence sequence on this quantity, so the developer can repeatedly test a system, inspect the bound, and decide to stop when it falls below a chosen threshold without invalidating the statistical guarantee. Our problem can be interpreted as a sequential variant of the missing-mass estimation problem which gave rise to the classic Good-Turing estimator (Good, 1953). On SWE-Chat, a dataset of real-world coding agent traces (Baumann et al., 2026), we implement our procedure and show that it achieves tighter upper bounds than a simple multiple-testing corrected Good-Turing baseline, and even dominates a fixed-horizon (non-multiple-testing-corrected) Good-Turing upper confidence bound, which is not anytime valid. Additionally, to better match realistic evaluation pipelines, we extend our method to handle sampling from a different task distribution and imperfect automated failure labeling.