We describe the system submitted by Telefónica Innovación Digital to the Speech-COSER Challenge 2026, which features spontaneous speech from elderly speakers recorded in rural Spain. For ASR, we perform CTC-based pseudo-forced alignment scores audio-text match quality to guide a two-stage LoRA curriculum, followed by post-hoc magnitude pruning of the LoRA matrices to enhance generalization. An ensemble of seven systems combined via ROVER achieves 36.25\% unnormalized and 22.17\% normalized WER.
We study stochastic bandits with reward-dependent delays, possibly infinite, whose reports omit the originating round. For each c[0,1), Balanced Historical Optimism (BHO) has a delay term controlled by the largest suboptimal-arm quantile at level 1-cΔ_i, without requiring optimal-arm reports. Lower bounds give an inverse-square slack cost in the logarithmic statistical regime and linear minimax regret at c=1. A family with growing K forces worst-case regret Ω(GK) for every learner, where G=_iΔ_i h_i is the largest gap-weighted suboptimal cutoff. Pending-triggered balance preserves BHO's weighted-cutoff bound for c<1 and attains the historical-greedy exploration coefficient with deterministic rewards, a unique optimum and fixed bounded delays. Gap-weighted comparisons and experiments show when shared waiting helps and what balanced exploration costs.
Predictive algorithms are increasingly used to guide hospital discharge decisions, yet these models can exhibit censoring bias for vulnerable patients due to insufficient historical data or past discharge decisions. When biased readmission risk predictors are used to inform discharges, the training data generated by discharge decisions may perpetuate similar biases when predictors are re-trained, leading to a vicious cycle. We extend a queueing-based Markov decision process for hospital discharge to incorporate periodically retrained readmission risk predictors. Our key insight is that cost-optimal, "capacity-aware" operational decision policies can naturally diversify the lengths of stay at which patient outcomes are observed, enabling consistent estimation of readmission risk without explicit fairness constraints. We prove that repeated retraining under discharge-up-to policies converges to unbiased risk estimates under well-specification. In synthetic experiments and calibrated simulations based on patient-level data from a partner hospital, we find that where commonly used threshold-on-risk policies can remain persistently biased, discharge-up-to policies reduce learning error over status quo decision policies (such as thresholding on risk) by 24%-25%. These improvements to estimation also mitigate some forms of bias, reducing readmission cost disparities (adjusted for baseline risk differences between demographic groups) by 38%-63%. Our results demonstrate how capacity-aware decision policies can improve both predictive performance and fairness by addressing censoring bias.
Many studies of animal vocalizations require clean recordings, but animals are usually recorded in the field, with background noise from weather, vehicles, and other animals. Animal vocalization denoising is challenged by the scarcity of clean training data and the extreme diversity of animal vocalizations. Classical spectral gates leave residual noise, especially without voice activity detection (VAD), while detector-free neural denoisers designed for bioacoustics remove less noise. We show that a general-purpose source separation foundation model, SAM-Audio, removes more noise than either of these types of methods, when used without training and prompted with time spans from a simple animal VAD. To turn this heavy method into a practical tool, we distill it into CallCleaner, a 19M-parameter denoiser that needs no detector and runs ~90x faster than real time on a laptop CPU. We find that SI-SDR has limited utility in our setting as it penalizes regenerated output and conflates noise removal with call fidelity, a trade-off our separate metrics reveal. CallCleaner leaves 11 dB less noise around the call than Biodenoising, the leading neural bioacoustic denoiser with public weights, and among fast, CPU-capable systems it pushes the Pareto frontier of noise removal vs. call fidelity, without requiring a detector. We release code and weights for CallCleaner, which enables fast, private, local denoising of single-caller recordings without specialized hardware.
Existing evaluations often test persona expression rather than whether persona information persists in extended decision-making. We introduce THEATRON, a controlled benchmark that uses financial replay to evaluate persona persistence in long-horizon sequential decision-making. In these environments, models conditioned on fixed behavioral-finance personas observe shared daily market evidence and trajectory-specific account-state feedback and make sequential decisions over multi-month trajectories. We evaluate Zero-Shot, One-Shot, and Chain-of-Thought persona injection across four LLMs, using cumulative-return and action trajectories to diagnose within-persona stability, cross-persona separability, and action collapse. Persona conditioning produces systematic behavioral shifts beyond a no-persona baseline, and persona-level separation generally exceeds the variation induced by semantically equivalent descriptions. Nevertheless, these shifts do not reliably form coherent persona-specific policies: apparently stable separation can arise from stereotyped action collapse, while diverse policies can exhibit noisy trajectory switching. These findings expose a gap between observable persona-induced behavior and persistent, context-sensitive persona-conditioned decision policies. Code and benchmark data are publicly available at: https://github.com/yuyan12138/theatron
We present MicroSim, a system for binary code similarity detection that represents each function as a graph built from IDA Pro Hex-Rays microcode and compares such graphs with a contrastive Siamese graph neural network (GNN). We evaluate on the TREX cross-optimization benchmark, a corpus of 849,072 function-level microcode graphs spanning optimization levels -O0 through -O3, under a strict class-disjoint split in which no function key is shared between train and test. Under this protocol, MicroSim reaches 0.9972 area under the receiver operating characteristic curve (ROC-AUC) and 0.9762 Matthews correlation coefficient (MCC), outperforming Gemini (0.9332 ROC-AUC), SAFE (0.9752), and jTrans (0.9740) retrained on the same split, as well as a structure-free Bag-of-Words (BoW) baseline (0.9484). We analyze the gap between pair classification and ranked retrieval in graph transformers, showing that Cross-domain Similarity Local Scaling (CSLS) under unified tie-resolved evaluation raises GraphGPS mean reciprocal rank (MRR) from 0.2542 to 0.2559 (R@1 = 0.1639, R@10 = 0.4561), strictly satisfying the theoretical MRR upper bound. Finally, over 5 seeds with equal compute, the pairwise cosine-embedding loss attains a slightly higher ROC-AUC (Δ = -0.0256, 95% confidence interval [-0.3282, +0.2770]), whereas the supervised-contrastive information noise-contrastive estimation (InfoNCE) loss is significantly better on thresholded classification and ranked retrieval.
In multi-institutional biomedical data analysis, comprehensive disease profiles are often fragmented across different data silos. Furthermore, local data distributions typically exhibit varying levels of granularity or even complex nested structures. Although Hierarchical Clustering (HC) is a powerful tool for exploring such data topologies, its cumbersome recursive process incurs prohibitive computational overhead and renders it highly susceptible to privacy leakage. This paper, therefore, introduces an efficient one-shot hierarchical FC framework that performs client-end (institutions) distribution exploration and server-end distribution aggregation through one-way prototype-level communication from clients to the server. Although the designed one-shot prototype transmission guarantees minimal computational overhead and mitigates privacy risks, it inevitably incurs information loss. To counteract this, a multi-granular competitive learning mechanism is proposed on the server to fuse the clusterlets, effectively maximizing the retention and mining of global distribution topologies under limited, one-way communication constraints. It turns out that the complex cluster distributions across clients can be efficiently explored, and extensive experiments comparing state-of-the-art methods on public datasets demonstrate the superiority of the proposed framework in balancing robust clustering, privacy preservation, and computational efficiency.
Swarm Learning improves resilience and removes the need for a single trusted coordinator by distributing model aggregation and validation across participating nodes. However, it also delegates the validation of model updates to peers whose private data may be biased, unrepresentative, or adversarial, and existing Swarm Learning protocols generally do not attest that data, leaving decentralized validation as an attack surface. We present a robust, privacy-preserving Swarm Learning framework that makes verifiable data quality the condition for admission to the learning task. A validator joins only after a zero-knowledge attestation proves, without revealing any raw sample, that a reference model ran correctly on its private data. The attested outputs also yield a label-free class profile, so the network assembles a collectively IID validator pool from individually skewed validators as accurately as an oracle with the true labels (total-variation distance to uniform 0.025, against 0.021 for the oracle and 0.21 for random selection). Admitted validators score client updates and, through a Tendermint-style Byzantine Fault Tolerant consensus, agree on a gradient mask and a quality-derived aggregation weight for each submission, replacing dataset-size weighting with a quality-aware signal. Adapting the cut-layer idea of Split Learning, a challenge-based audit lets peers re-verify that a mask was computed on the attested data without an expensive end-to-end zero-knowledge proof. Across four benchmarks under non-IID partitions the framework matches FedAvg on clean data; under a gradient-scaling attack that collapses FedAvg to random chance, it retains 94% of clean accuracy on CIFAR-100 and over 99% on FashionMNIST, where FedAvg achieves only 2.1% and 11%. The result is a private, robust, and auditable foundation for decentralized learning.
The small open embedder EmbeddingGemma-300M, used as released, finds the right passage first for only 51.7% to 60.8% of Belebele questions in Tajik, Kazakh, Uzbek and Kyrgyz. We present Saidzoda-CA-EmbeddingGemma-300M, one model for all four languages. It is our Tajik fine-tune of EmbeddingGemma-300M, trained for one more epoch on 279,904 pairs: questions in Kazakh, Uzbek and Kyrgyz that our own language model wrote about source passages, a Tajik refresher, and English and Russian replay. On Belebele, recall@1 rises by 14.4 to 26.0 points in each language, to between 75.2% and 77.7%. Against single-language models trained with the same settings, the one model stays within 2.8 points at recall@1 in Kazakh, Uzbek and Tajik, with overlapping recall@10 intervals, though it also had more training. It keeps the Tajik law recall@10 of its starting point (92.2 against 92.4) but loses 4.4 points of law recall@1. A model trained on one of these languages also improves the others on Belebele, but less than training on them. The model still trails the 560M XLM-RoBERTa embedders on Kazakh, Uzbek and Kyrgyz Belebele: the better of multilingual-e5-large-instruct and BGE-M3, used as released, is 9.4 to 10.2 points ahead at recall@1. Our scores of those released models agree with the published MTEB results within 2.4 points. Russian and English recall@1 fall by 3.0 and 4.2 points.
Generating accurate CAD models requires substantial engineering expertise and careful control of geometry, topology, and mechanical constraints. Text-to-CAD aims to simplify this process, but current LLM-based methods often generate CAD programs directly. Language models remain unreliable when producing long code while coordinating coupled geometric, topological, and dimensional constraints. Consequently, executable programs may still produce structurally incorrect models, which execution rate and rendered appearance may fail to reveal. To address these limitations, we propose GOPR-CAD, an agentic Text-to-CAD pipeline based on Graph-of-Operations Progressive Reasoning. A language model acts as a decision agent over encapsulated CAD operations and progressively constructs an executable operation graph, which a deterministic compiler converts into the final CAD model. This separation requires no additional training and supports different language models. To evaluate structural fidelity, we construct MHCAD, a benchmark of 300 standard mechanical part requirements derived from Machinery's Handbook. Their explicit structures enable reference-free STEP/B-rep evaluation. GOPR-CAD achieves 82.7% Structural Yield on MHCAD, and 87.2% Structural Yield on an external CADPrompt subset.
Large language models (LLMs) perform well on general coding tasks, but their capabilities in application-layer embedded software development remain underexplored. Existing embedded benchmarks primarily focus on hardware-in-the-loop validation in simulated environments. We introduce RealEmbedBench, a benchmark of 5,142 samples covering three application-layer tasks: code generation and completion, state machine modeling and implementation, and cross-OS code translation. We evaluate six representative LLMs and find substantial performance differences across tasks and models. Project context, especially the local code window, significantly affects code generation and completion. Models perform reasonably well at state machine modeling but struggle to translate these models into executable code. Cross-OS translation performance varies substantially by direction. Translation from Harmony to RT-Thread is particularly challenging, with an average pass rate of 0.1667. These findings suggest that developers should consider local code context when using LLMs for code generation and completion and select models according to the translation direction. They also motivate the use of more real-world state machine implementation data in LLM training to narrow the gap between modeling and executable code.
Audio deepfake detectors trained with known synthesis attacks can learn generator-specific shortcuts that fail as synthesis systems and acoustic conditions evolve. We investigate an attack-free alternative: a normality model trained exclusively on bona fide speech, which scores synthetic audio by its deviation from authentic representations. We propose MLN-WavLM, a method that applies shrinkage-Mahalanobis modeling to pooled hidden states from a frozen large-scale, pre-trained speech processing WavLM Base+ architecture. Evaluated on ASVspoof 2019 LA, our best attack-free variant achieves an 8.45\% Equal Error Rate (EER), closely tracking a fully supervised WavLM-HM-Conformer baseline (7.86%). However, when subjected to the ASVspoof 2021 DF domain shift, performance drops to 22.12% EER, exposing vulnerabilities in the bona-fide-calibrated operating point. Layer-wise analysis reveals that early WavLM representations dominate the ensemble. To test real-world viability, we simulated deployment streams using unlabeled threshold adaptation. While robust median/MAD recalibration easily handles in-domain contamination up to 10%, recalibrating under the 2021 DF domain shift triggers a total collapse of low-FPR recall. We conclude that while frozen source-domain normality can effectively rank deepfakes, it does not yield a stable open-world decision rule, and downstream threshold adaptation cannot rescue upstream ranking degradation.
The CoRL 2026 RoboEval challenge has eight bimanual tasks, and it ranks a policy not only on whether it succeeds but on how: every stage of the task, done safely, efficiently and with both arms in coordination, scored against human demonstrations. My solution has two parts. The generic part, carried over from my LeHome challenge solution, is off-policy RL on a π_{0.5}-based policy that is also its own value function. It is trained with AWR and RECAP on advantages of the exact official score, in a loop where every rollout is at the same time a Thompson-sampling trial of about 20 inference parameters. Out of the box, it took the policy to 0.85 RDS on held-out seeds in two days (and it was still very far from convergence). The competition-specific part adds data that local exploration cannot find: generated demonstrations that reach the full reward, and recovery episodes that start from the policy's own failures. The submitted policy reaches 0.990 RDS on held-out seeds and 0.987 on the official seeds.
Aligning two models without paired data requires a distance that both share, yet existing methods use Euclidean or spherical geometry. This is either an arbitrary choice or imposed by their loss function. We show that, under a precise form of the manifold hypothesis, the pullback of the data metric gives such a distance for any two models of one data-generating process. This holds whether the models are encoders, decoders or both, and whatever their architecture, loss, prior or training data. The alignment map preserves this distance and the distribution of representations, and is identifiable if and only if the data distribution breaks every symmetry of the data manifold. We show that this symmetry breaking holds generically. Identifiability makes the alignment coupling the unique optimum of Gromov--Wasserstein and of its weighted combinations with a linear program. The linear program alone has the same optimum if almost every data point has a distinct distribution of distances. The coupling, and with it the alignment map, is recovered exactly from permuted samples, even without identifiability, and consistently from independent ones. Our method, Neural Gromov--Monge Alignment (NGMA), uses no pairs or labels and outperforms every unpaired and label-informed baseline. It relies on pullback distances to align image models with different priors and single-cell models trained on disjoint cells, where Euclidean distances fail.
Temporally correlated noise can substantially improve the utility of differentially private optimization, yet candidate mechanisms are commonly designed or evaluated using only partial information about the optimization process, such as workload structure, learning-rate schedules, or local curvature. We study whether persistent optimizer state provides additional information that can change the relative utility of privacy-valid correlated-noise mechanisms. We introduce MOMENT-DP, an optimizer-state-aware framework that models how temporally correlated privacy noise propagates through momentum-based optimization. For locally quadratic objectives, MOMENT-DP expresses privacy-induced final loss through an optimizer transfer operator that jointly captures curvature, learning-rate scheduling, momentum, and temporal noise covariance. NoiseCurve's curvature-aware quadratic criterion is recovered exactly as the zero-momentum, constant-learning-rate special case. Changing only momentum while holding curvature, schedule, privacy calibration, and candidate mechanisms fixed induces genuine mechanism-ranking reversals. The effect persists outside exact quadratic dynamics in per-example-clipped private logistic regression, while remaining unresolved in the neural regime we study. These results establish optimizer state as an under-modeled, decision-relevant axis in evaluating temporally correlated privacy noise.
Agentic autoregressive video generation combines semantic planning, sequential generation, and visual verification for controllable synthesis. Growing visual histories enlarge the key-value (KV) cache, increasing memory use and attention costs across candidate rollouts. Existing compression methods primarily optimize individual generators without the planner's semantic constraints, risking low priority for states supporting requested attributes or actions. We introduce SemanticKV, a training-free method for semantically aligned KV cache compression. To our knowledge, SemanticKV is the first KV cache compression method for agentic autoregressive video generation that explicitly uses planner-extracted semantic constraints to guide retention. Text cross-attention identifies constraint-related video positions, whose self-attention queries are collected into separate groups. Attention weights are averaged within each group and maximized across groups, giving each cached entry a high score when strongly used by any represented constraint. A single top-k selection retains high-scoring entries under the memory budget, leaving their keys, values, and positions unchanged. For fixed query groups and equal entry costs, this selection minimizes an additive upper bound on worst-group discarded attention. The method operates independently within each candidate without retraining or additional video generations. We evaluate four AR video backbones on three benchmarks for 5-, 30-, and 60-second generation. On Causal Forcing at a compression ratio of 0.7, SemanticKV obtains a VBench Total Score of 84.10, compared with 84.05 for the strongest compressed baseline and 83.99 for Full KV. It achieves a 1.22x speedup over Full KV and reduces KV-cache memory from 6.04 GB to 1.81 GB, a 3.35x reduction.
On-policy distillation (OPD) trains a student language model on its own responses, guided by a fixed teacher. Generating these responses is expensive, so systems use quantized inference and asynchronous rollouts to reduce the cost. Both, however, introduce training-inference mismatch between the rollout engine and the model being trained. Such mismatch can destabilize reinforcement learning (RL), but its effects on OPD remain unclear. This paper investigates how much mismatch OPD tolerates, where its effects arise, and how training systems can use this tolerance. Measuring all mismatch sources on one scale, we examine quantization across three student-teacher pairs, with additional tests of weight staleness and injected noise on one pair. OPD tolerates substantial mismatch without meaningful loss of training quality but degrades at higher levels, revealing an empirical mismatch budget. Both the quality cost of mismatch and the mismatch introduced by quantization depend on the model, with some layers particularly sensitive to reduced precision. In controlled comparisons, OPD remains stable under rollout configurations that destabilize RL methods. Under weight noise, mismatch also shifts the prefixes reached during generation, introducing errors that per-token corrections cannot address. To put these findings into practice, we introduce a lightweight admission check that screens rollout configurations against a calibrated budget without requiring a full training run for each candidate. Model-specific FP8 recipes preserve training quality while providing capacity for at least 1.93as many KV-cache tokens as BF16. Combining FP8 with asynchronous rollouts reaches the same training quality 1.62sooner than synchronous BF16 training. Code is available at https://github.com/benyucong/train-inference-mismatch-opd
