Mixture-of-Experts language models are cheap to run per token but expensive to hold in GPU memory, which drives the cost of serving them in production. One-shot expert compression shrinks them after a single calibration pass. Its two current state-of-the-art methods are REAP for expert pruning, which deletes the least salient experts, and REAM for expert merging, which merges them into the most salient ones. Both have been evaluated only on academic benchmarks. We compare the two on five multilingual production tasks of a web-hosting company, from classification to the generation of marketing HTML email templates, with two 30B-parameter models (Qwen3-30B-A3B-Instruct-2507 and the hybrid Mamba-attention Nemotron-3.5-Lightning) at 25% and 50% of experts removed and the same calibration data. Pruning wins: on the same rows, REAP is significantly better in 16 of 20 comparisons and REAM in 2. At 25%, REAP keeps 93.6% of Qwen3-30B's generative quality and REAM 62.5%, although on benchmarks the two are within a point for the same model. The damage concentrates in the output contract rather than the content: 89.8-95.2% of each significant loss on email templates comes from hard-gate failures such as invalid JSON, broken mobile layouts and runaway generation, and most failed non-English domain descriptions copy the prompt's English opening phrase. Cross-perplexity and logit divergence, two diagnostics from the REAP paper, also track production quality, as does the rate of repetition loops; n-gram diversity does not. We close with practical guidance for compressing MoE models for production.
On-policy self-distillation lets a language agent act as its own Teacher by conditioning on privileged context such as skills written from its own trajectories. Its guidance, however, is weighted by how much this context shifts token likelihoods rather than by whether it improves behavior, so errors in self-written skills return through later updates. We propose environment-calibrated on-policy self-distillation, which uses environmental outcomes to decide how much to trust the Teacher's current guidance and which self-written skills to reinforce. Command-level advantages, estimated by comparing commands executed at similar observations, scale each Teacher preference and reward reflections and skills whose induced preferences favor higher-advantage commands. Without skills at inference, 3B and 7B agents reach 90.5% and 93.3% macro success on seen ALFWorld tasks and 75.5% and 80.5% success on WebShop. On unseen ALFWorld tasks, they outperform GRPO under the same protocol by 11.0 and 7.6 percentage points, respectively. In an ablation study, supervising only executable commands rather than the Actor's full response raises 3B WebShop success from 52.3% to 75.0%.
A worked solution can make self-generated training targets shorter and cheaper, but accuracy on the practiced operation may hide how well a model can reuse newly learned rules. We study this choice in a 2B-parameter reasoning model learning two invented quantitative rules. The model generates targets from a rule sheet alone or from the sheet plus a program-written worked solution; targets are accepted for final-answer correctness, and the support is removed for fine-tuning and evaluation. Training questions request forward use in one segment; composition requires sequential reuse of the same rule across two segments. After six passes, using a program-checked numerical signature, sheet-only targets lead by 24.3 percentage points (95% item-bootstrap CI 14.6-34.0) on composition in a prospective test. On unseen argument values, observed forward accuracy is identical (87.5% each), but composition accuracy is 58.3% versus 23.6%, a 34.7-point gap (25.7-43.8). An independently sampled second target-bank pair repeats the direction (+14.6 points, 2.1-26.0). Ten passes on the shorter worked-solution targets, approximately matching the domain tokens of six sheet-only passes while changing other training quantities, narrow the gap to 16.3 points (6.3-26.7) in the first bank pair and 6.2 points (-6.3 to 18.8) in the second, where the pre-specified test does not hold. At six passes each, the worked-solution policy uses fewer capture tokens and less update time, yet the trained models use 2.6 times as many generated tokens per composition success under the numerical signature (2.1 times in the second bank pair). Choosing a capture context therefore requires testing how the new rules are reused, alongside the inference cost of those uses, rather than relying on forward accuracy alone.
Systematic literature reviews (SRs) remain slow and labor-intensive, and most automation tools address single stages of the workflow. We present PARIS (PRISMA Agentic Research Intelligence System), an open-source framework in which a conductor agent running in Claude Code orchestrates 15 specialized large language model (LLM) agents, from scoping a research question to exporting a manuscript draft. Agents are specified declaratively, coordinate exclusively through persisted artifacts that the conductor validates before every transition, and delegate retrieval, bibliometric analysis and scoring to deterministic Python utilities. Three human checkpoints gate study inclusion, corpus adequacy and the synthesis outline. We also introduce QualityScore, a pre-synthesis index that aggregates normalized network, bibliometric, semantic and process sub-scores into a 0-100 grade. Two case studies illustrate both: AI chatbots and university students' learning (388 records, 100 included) and AI-assisted SR methods (131 records, 123 included, QualityScore 69.1 recomputed with a proxy network sub-score). They show end-to-end feasibility across domains and the diagnostic use of QualityScore: in the first run, corpus improvements chosen at the corpus-adequacy checkpoint raised the score from 74.6 to 81.9 and the number of pre-2023 studies from 1 to 19. Audits of both runs exposed failure modes: incomplete retrieval, counts that drifted between agents, unrecorded or delegated checkpoint decisions and a grade that depends on how citation data are handled. PARIS fully supports 8 of the 27 PRISMA 2020 items and partially supports 11. Output quality has so far been judged only by an internal reviewer agent; independent validation against human reviews is the next step.
Quantum kernel methods have been proposed for network intrusion detection, yet their performance limitations are often attributed to noise in current quantum hardware. This study evaluates that assumption through a controlled analysis of Pegasos-QSVC on the NSL-KDD and ToN-IoT datasets. We perform 56 quantum experiments using one exact noiseless statevector backend and three noise-simulated superconducting backends while keeping preprocessing, PCA features, data splits, and classical baselines identical. XGBoost achieves accuracies of 98.94% on NSL-KDD and 99.41% on ToN-IoT, compared with 91.80% and 82.98% for Pegasos-QSVC. The quantum model also exhibits substantially greater seed-dependent variability. Eight runs collapse into degenerate classifiers, including four recall-collapse cases and four trivial-positive classifiers. Notably, two failures occur on the noiseless backend, demonstrating that hardware noise is not a necessary condition for classi-fier collapse. Kernel Gram matrix analysis further reveals low effective rank and severe ill-conditioning even in the absence of noise, indicating intrinsic fragility in the learned quantum representation. These results show that average accuracy alone can conceal operationally unacceptable behavior and highlight the importance of reporting worst-case performance, seed-level variability, class-specific metrics, and explicit degeneracy tests in quantum intrusion detection studies.
Transformers are built with continuous symmetries in the attention mechanism. We derive $\rho$, a scalar invariant that measures each head's capacity to couple to rotary positional encodings. This framework allows us to analyze heads without forward passes, opening a new research direction into data-independent analysis of modern LLMs. Empirically, we find that $\rho$ predicts induction heads, the most-studied circuit type in mechanistic interpretability, with Spearman correlation between $-0.55$ and $-0.74$ across four bias-free RoPE models spanning $135$M to $7$B parameters. As a preliminary proof of concept, we use $\rho$ to apply YaRN to only the top-50% positional heads, which degrades perplexity less than uniform application. Code is available at https://github.com/MPedraBento/rho-invariant.
In patrol missions involving multiple unmanned surface vehicles (USVs), a designated command USV can support fleet management by sequentially rendezvousing with each patrolling USV. These operations require motion-feasible route planning and priority-aware rendezvous sequencing, with the objective of minimizing mission completion time and total tardiness relative to priority-dependent due dates. We propose a framework combining grid-based path planning, surrogate travel-time prediction, and graph-based reinforcement learning. The path planner accounts for ocean currents and approximates turn-rate constraints through multi-stage maneuvers, with path feasibility empirically validated through trajectory-tracking simulations. For rendezvous sequencing, a Feasible Transition Graph (FTG) represents candidate rendezvous events and temporally feasible transitions, enabling a graph neural network to exploit spatiotemporal rendezvous relationships for policy learning. A surrogate predicts travel times from compact current and motion features to avoid repeated path searches during FTG construction and policy training. Evaluation using grid-based path planning shows statistically significant average objective reductions of 10.5%, 3.5%, and 1.5% compared with greedy, MLP-RL, and 2-step lookahead greedy, respectively. The policy achieves these improvements while requiring approximately 1/35 of the latter's computation time. It maintains its advantage over MLP-RL under unseen priority-tier settings, with larger gains as fleet size increases.
Verbal memory retrieval carries rich diagnostic information for Alzheimer's disease through the order, hesitation, and categorical organization of recalled words, yet computational modeling of this retrieval process remains largely unexplored. A key challenge is that timing and hierarchical structure are diagnostically entangled: the diagnostic meaning of retrieval timing depends on its hierarchical context. We propose TrajLift, a spectral framework that models verbal recall as timed trajectories on a hierarchical graph. Our central mechanism is latency-modulated heat diffusion on the semantic hierarchy, a unified operator that lifts each recalled word into a representation of both multi-scale hierarchical structure and continuous retrieval timing. Formal analysis shows that the resulting representation is first-order separable into structural and temporal components with spectral selectivity across hierarchical levels. On synthetic benchmarks requiring structure-time interaction, TrajLift stays near-perfect on deep hierarchies where all baselines fail. On real-world clinical data, we show that TrajLift outperforms diverse baselines by jointly modeling timing and semantic structure, with gains that grow as retrieval timing becomes more informative.
Computational advances in natural language processing and machine learning have substantially expanded the use of automated approaches in verbal deception research. However, existing verbal deception datasets remain fragmented across disciplines, as they are often difficult to identify, access, and reuse, limiting cross-study comparisons and the development of robust computational models. We present the Deception Archive, a curated and standardized repository of publicly available verbal deception datasets built upon a large-scale systematic review. Eligible datasets were retrieved, manually inspected, and standardized following a structured data curation workflow. The current repository comprises 41 publicly available datasets containing 387,995 truthful or deceptive statements. The archive represents a heterogeneous repository (e.g., in data size, topic of investigation, ground-truth, etc.) and a living resource for integrating newly available deception datasets over time. By providing centralized access to standardized datasets and accompanying metadata, the Deception Archive facilitates dataset discovery, cross-dataset validation, and reproducible evaluation of computational models.
Federated learning (FL) is a promising research paradigm that enables collaborative model training across distributed clients without data sharing. To handle unlabeled data in local clients, existing unsupervised FL research has primarily focused on tasks like clustering and self-supervised representation learning, leaving the critical problem of unsupervised visual recognition largely unexplored. To bridge this gap, we introduce FedUR, the first framework designed for unsupervised recognition by leveraging the zero-shot capabilities of pre-trained vision-language models like CLIP. Specifically, FedUR addresses data heterogeneity through a pseudo-labeling mechanism with a class-aware data selection strategy. This strategy first estimates a global pseudo-label distribution, then allocates pseudo-labeled data to clients accordingly. To handle the uncertainty of conventional positive pseudo-labels, FedUR further incorporates class weights and a negative pseudo-labeling strategy. Furthermore, FedUR proposes a personalized prompt aggregation protocol where general visual prompts are aggregated on the server, and personalized text prompts encoding local knowledge are retained on clients. Extensive experiments demonstrate that FedUR consistently outperforms baseline methods under both label-level and feature-level distribution shifts. We believe this work paves the way for further advances in federated unsupervised recognition. Code will be available at {github.com/tim-learn/FedUR/}}.
Graph transformers can capture long-range interactions, but their quadratic attention cost limits their use on large graphs. We introduce \emph{NearFormer}, a graph transformer whose global attention is restricted to a randomized near-neighbor (NN) graph. NearFormer uses an efficient randomized construction to connect each node to a small number of near neighbors from a broader neighborhood defined by a label-independent feature or spectral geometry. For a fixed retained budget \(k\), the resulting attention support has linear size \(\mathcal{O}(|E|+nk)\). The NN graph is built once by within-bin search and then held fixed. We use these edges as sparse global-attention supports, with optional learned attention biases, while preserving the original graph for local message passing. Experiments on long-range graph benchmarks and synthetic structural tasks show that NearFormer provides a favorable accuracy--scalability trade-off relative to local message passing, dense and sparse attention.
Scientific discovery requires ideas that are original, useful, and non-obvious. Yet AI scientists are typically given as much task information as possible during idea generation, which can anchor generation toward familiar approaches. Inspired by the cognitive framework of Blind Variation and Selective Retention (BVSR), we propose a discovery framework that separates these two roles of information: it generates proposals under deliberately limited task information to broaden exploration, then learns from proposals previously verified on the same task and their outcomes to retain promising candidates for costly empirical evaluation. We evaluate BVSR across five proposer models and 12 scientific tasks. Limiting task information substantially increases proposal originality and non-obviousness, and produces a larger share of highly creative and high-utility proposals than existing state-of-the-art prompting-based baselines. Learning from prior proposal outcomes on the task further improves candidate selection, discarding half of the candidates while retaining most top-performing proposals. When incorporated into existing AI-scientist workflows, BVSR achieves stronger discoveries across gene expression analysis, drug response prediction, anomaly detection, scaling law discovery, and brain activity prediction. These results suggest that effective AI scientists should use information differently across stages of discovery: constrain task-specific information during exploration while exploiting accumulated empirical experience on the task.
On-demand delivery platforms such as DoorDash and Uber Eats impose a constraint absent in traditional e-commerce: merchants beyond the user's delivery range are invalid candidates rather than merely low-ranked. Generative retrieval has recently shown strong potential in large-scale retrieval systems. Adapting it to this setting, however, poses three challenges. First, existing semantic IDs encode geography only as a soft signal, so reachability can only rely on a non-differentiable post-generation hard truncation. Second, user preferences vary substantially across both regions and queries, so aggregating historical behaviors without conditioning on geography and the current query yields biased interest estimates. Third, the retrieved candidate set is evaluated by its collective value, yet autoregressive training scores each sequence independently, with no set-level signal to discourage redundant near-duplicates. To address the challenges, we propose TS-GEAR, a unified generative retrieval framework for on-demand delivery: (1) Geo-Semantic ID splits merchant identifiers into geographic prefixes and semantic suffixes, making reachability a prediction target rather than a post-hoc filter; (2) Query-Geo Target Attention aggregates historical behaviors using the real-time query and location as dual targets, while encoding each behavior's own location into its attention key; (3) GMPO (Group Marginal Policy Optimization) rewrites advantages using a deduplicated group reward and marginal advantage within GRPO, avoiding auxiliary diversity objectives. Deployed online with a lightweight 0.1B backbone, TS-GEAR serves over 150 million queries per day in the search advertising pipeline of Taobao Shangou, consistently outperforming state-of-the-art generative retrieval methods on both HitRate and geographic validity, and achieving 3.8% gains in ad impressions and 4.2% in ad revenue.
Cohort comfort models (CCMs) aim to overcome the extensive individual data requirements of personal comfort models (PCMs) by grouping occupants based on demographic and psychometric characteristics (cold start) or similar thermal preferences (warm start). We evaluated the validity of the reported CCM performance gains through (i) a methodological assessment of the original study, (ii) a replication using one of the two original datasets (DORN: 20 participants, 6 months) under a revised experimental protocol, and (iii) a replication using an independently collected dataset (PersonAI: 72 participants, 4 weeks). Classification performance was evaluated using F1-micro, as in the original study, and F1-macro to account for class imbalance. Under F1-micro, a trivial baseline that always outputs the majority class performed on par with cohort models. Under F1-macro, all cohort models achieved lower median scores than a simple baseline that makes random predictions based on the estimated distribution of thermal preferences. In our methodological assessment we identified the following key issues: Limited data per participant and per minority class, reliance on a metric insensitive to class imbalance, missing trivial baselines, inadequate cross validation strategy, and multiple occurrences of information leakage. Our findings challenge the original study's conclusion that CCMs improve thermal preference prediction. CCMs did not outperform simple class-distribution baselines. Future studies should improve experimental designs, including controlled thermal conditions, to obtain sufficient observations across thermal preference classes.
Early successes on agentic robotics point to a potential paradigm shift toward using coding agents to control robots. Existing robotics benchmarks are built around visuomotor policy learning, leaving it open whether coding agents can reliably control diverse robots under physical-world constraints. We therefore introduce the Embodied First Exam, a unified, comprehensive, and agent-native evaluation framework where coding agents can choose their own control approach: direct control, code to control, or train a policy to control. We hope it serves as a first exam for coding agents before they enter the physical world, with 1,156 tasks in 35 suites on 47 robots, including 9 new suites that go beyond pick-and-place to spatial puzzles, dexterous manipulation, multi-robot interaction, dynamic whole-body control, and sim-to-sim engineering. Each task can be run in up to three evaluation settings. The Standard Setting mimics the real world: the agent receives only sensor observations and, by default, cannot reset the environment. Its Realtime variant keeps the environment running while the agent reasons. The Privileged Setting gives the agent access to the full simulator state and unlimited resets, allowing it to search for an oracle solution for training purposes. Success rates in this setting serve as an approximate upper bound. On the 456 core tasks, the strongest of eight agents, GPT-6 Astra in Codex, achieves a success rate of 34% in the Standard Setting, leaving much room for improvement. To see where agents fail, we label the stage at which failed attempts go wrong and compare failures across agents, settings, and stage categories.
Robotic manipulation has advanced rapidly in mapping visual observations and language instructions to actions. However, reliable manipulation in open-world environments requires robust spatial understanding: the ability to identify targets in complex layouts (spatial reasoning) and adapt to scenes that continuously evolve through interaction (spatial memory). Yet existing embodied benchmarks fall short in evaluating spatial understanding: they often introduce strong visual cues that bypass spatial reasoning, and focus on static scenes that offer little insight into an agent's capacity for spatial memory. To address this gap, we introduce GeoMind, a systematic benchmark for structured spatial understanding through tabletop manipulation. GeoMind comprises 54 evaluated task variants organized around spatial reasoning and spatial memory. It also incorporates a scalable pipeline that automates the generation and annotation of task instances with verified targets and executable interactions. Extensive evaluations across diverse embodied agents reveal their consistent weaknesses in spatial understanding. Our findings highlight that these agents struggle to track identities after spatial changes and convert relational or geometric constraints into executable actions. We release the GeoMind benchmark at https://github.com/Notfound-JH/GEOMIND-benchmark.git.
We introduce FraudBench, an evaluation of large language models' (LLMs) propensities to recommend, assist, or directly take actions that would constitute fraud under U.S. law. Our evaluation contains a novel dataset of 159 scenarios based on real-world U.S. court cases and validated by legal experts. We present each scenario to LLMs via three tasks at varying levels of autonomy: advising users, drafting messages for users to send, and acting directly on behalf of users (477 tasks total). In each scenario, FraudBench re-enacts the facts of a court case to give LLMs an opportunity, but no explicit instruction, to either misrepresent or conceal a material fact in a high-stakes commercial setting--e.g., whether to describe company-funded product testing as "independent" to customers (misrepresentation) or whether to hide unfavorable regulatory investigations from clients (concealment). We evaluate 20 LLMs across 25 experimental conditions (totaling >300,000 generations) and include a preregistered human baseline ($n_{\texttt{annotations}} = 2{,}957$, $n_{\texttt{humans}} = 996$). We score a response as fraudulent when it engages in, assists with, or recommends misrepresentation or concealment. Key findings include: (1) across LLMs, fraud rates ranged from near 0% to over 40%; (2) LLM fraud rates increased with autonomy, with overall rates of 3.0% when advising, 14.7% when drafting, and 21.4% when acting directly, without applying pressure; (3) LLMs sometimes engaged in fraud despite recognizing legal risks; (4) LLMs more often engaged in fraud by concealment than by misrepresentation; and (5) prompt-based interventions substantially affected fraud rates, with a prompt to obey the law reducing fraud rates, while prompts to maximize profits and act agentically increased fraud rates.
Code localizers are usually trained to predict the files a reference patch edits, but the code a fixer needs to read is not the same as the code it changes. We propose Fixer-Guided Localization (FGL): the localizer proposes several localizations, a frozen fixer scores each by how likely it makes a verified patch, and the localizer is trained on these scores with weighted supervised fine-tuning, without running tests and without reinforcement learning. We build the full pipeline on Gemma 4 E4B for the Kaggle Gemma-4 Developer Agent competition, from 1,629 test-verified Kimi K3 trajectories. The FGL target helps the fixer find the bug as often as the teacher's localization while using 27.5% less code. On the development set, one EM round moves the agent in the expected direction compared with a matched control: more tasks reach a patch (53 vs. 46), fewer tasks get stuck reading files that do not exist (43 vs. 54), and more tasks are solved (8 vs. 6). Training also changes how the agent works: trained agents explore about three times more than the base model and write far fewer wrong patches, and the single agent solves up to 11 of 78 tasks, against 8 for the base model. An analysis of 13 evaluation runs identifies what limits a 4B agent under an 8-minute budget (time, repetition loops and edit quality) and gives practical lessons for training small coding agents. All data, code and adapters are publicly available.
Differentiable simulation propagates derivatives through a physics simulator, turning animation, system identification, control, and computational design into gradient-based optimization. Engines differ in the physics they model and, less visibly, in which mathematical object they differentiate. We separate two questions the literature conflates. The referent is which object the returned derivative belongs to: the finite computation that executed, the converged equation it approximates, or the continuous dynamics. The realization describes the organization of the backward pass: automatically generated differentiation, an equation-level adjoint solve, or a custom adjoint. The referent defines the target against which exactness is assessed; the implementation determines how accurately that target is approximated. Cost depends on the realization together with how the states needed by the backward pass are retained or recovered. Crossing the axes places every method and engine we classify in a cell, separates claims the literature merges, and shows that the apparent trade-off between exactness and memory follows from how methods have been built rather than from a constraint. A second, orthogonal taxonomy covers contact, where the map is non-smooth. We then ask whether a correct gradient helps. Four mechanisms by which useful gradients are lost are cataloged and measured on one testbed, where the advantage of the exact derivative over the equation-level adjoint depends on where the fitting target comes from as much as on the solver budget, and we review the evidence that zeroth-order methods can outperform simulator gradients in the regimes where those mechanisms act. The model classes, systems techniques and applications are surveyed through this lens, and we close with a reporting checklist, a released diagnostic that computes it, and a research agenda.
Training a neural network under homomorphic encryption requires transposes, batch reductions, optimizer updates, and repeated refreshes in addition to forward matrix products. We combine rectangular diagonal packing, feature-major SIMD mini-batch lanes, baby-step giant-step evaluation, and compact optimizer state in an MIT-licensed Rust CKKS implementation. Its CUDA path keeps training and bootstrap ciphertext arithmetic on the GPU, while retaining host key generation and client operations. In a matched 100-update experiment on the main non-interactive MNIST model, device-resident bootstrapping reduces training time from 14,268.295 to 293.075 seconds, a $48.68\times$ ratio against our CPU implementation, with exactly equal decoded parameters and predictions. Including setup and final evaluation gives a $40.67\times$ ratio. A separate GPU run completes ten passes over 4,800 images in 2,967.095 seconds and achieves 90.35% test accuracy versus 90.41% for its plaintext twin. These are single-initialization results; exact CPU/GPU parameter agreement is verified at 100 updates, not for the full 1,000-update run. Earlier completed CPU runs support a larger-batch key-switch advantage and baseline-batch accuracy within the registered tolerance, but the baseline-batch normalized key-switch target is missed. The released ReBoot baseline does not meet the common security-estimator threshold, and its full-setting CPU time is unmeasured. We therefore distinguish measured acceleration of our implementation from conditional layout comparisons and do not claim unconditional superiority over ReBoot.