Gradient equilibrium (GEQ) is a recent framework in online learning that requires the time-averaged gradients along the learner's trajectory to vanish. We study GEQ when the learner is constrained and ask when such an equilibrium remains achievable in an adversarial environment. We first show that constrained GEQ (CGEQ) can be impossible even under simple and slowly varying constraints, motivating imposing local temporal structure on the adversary. We then develop an approachability framework for CGEQ under locally constrained adversaries, and obtain necessary and sufficient characterizations in the spirit of classical Blackwell approachability. Beyond these theoretical characterizations, we provide an algorithmic principle that converts window-wise no-regret learners into CGEQ algorithms under a local feasibility condition with margin. Experiments on synthetic and real data demonstrate that the resulting algorithms perform strongly in both achieving GEQ and respecting the constraints.
Speech Emotion Recognition (SER) is a fundamental task in affective computing that enables applications such as mental health assessment, human-computer interaction, and automated customer support systems. Accurate emotion classification from speech remains challenging due to variations in speaker style, acoustic conditions, and the high computational cost of modern deep learning approaches. Many existing systems depend on large transformer-based architectures that require substantial computational resources, making them less suitable for real-time or resource-constrained environments. This research proposes EmoNet, a lightweight deep learning framework for speech emotion classification across seven emotion categories. The model is built on a four-block depthwise separable CNN encoder with squeeze-and-excitation attention and residual connections, combined with a Bidirectional LSTM for temporal feature learning and a multi-head attention pooling mechanism for improved representation. A comprehensive 134-channel acoustic feature set alongside data augmentation techniques including audio-level perturbations and SpecAugment is employed to improve robustness under diverse acoustic conditions. Evaluated on a combined dataset of TESS, RAVDESS, and SAVEE, EmoNet achieves a test accuracy of 96.75% with only 4.11 million parameters. The proposed model outperforms its reproduced baseline and remains competitive with significantly larger state-of-the-art architectures.
"Data quality" is widely invoked in instruction tuning, but the term is often underspecified. We study it as an operational property by asking which simple example-level metrics, drawn from prior work on data selection, text quality, and instruction tuning, best predict training utility under a fixed finetuning budget. For each metric, we sort instruction-response examples by score, partition them into low, middle, and high bins with roughly equal token mass, finetune separately on each bin, and compare validation perplexity and LLM-as-judge scores for the models' generated answers as a function of cumulative training tokens. Our results suggest that answer-dense supervision, as captured by response ratio, is a stronger and more portable predictor of instruction-tuning utility than commonly used alternative proxies.
Generative Flow Networks (GFlowNets) are trained to sample objects from compositional spaces with probability proportional to a reward function, but ineffective exploration can cause learned samplers to concentrate on a few high-reward regions. An intuitive remedy is to train multiple GFlowNets independently in parallel and combine their samples. However, this does not prevent multiple models from collapsing onto the same high-reward areas, leaving large regions of the state space unexplored. With this in mind, we propose Hierarchical Cooperative GFlowNets (HCG), an approach that dynamically partitions the search space to coordinate discovery without relying on hand-crafted boundaries. HCG imposes an ordering in which the highest-ranked model may explore the entire space, while lower-ranked models are progressively pushed toward regions not covered by higher-ranked ones. We provide a theoretical characterization of how specific learning objectives can shift probability mass away from designated areas, instantiating this mechanism within our architecture. A shared classifier coordinates the hierarchy by partitioning the search space, keeping the GFlowNets independently parameterized and enabling asynchronous training. Furthermore, we derive a normalizer-weighted aggregation rule for the ensemble and prove that it recovers the target reward-proportional distribution under the stated assumptions. Finally, experiments across multimodal distributions, sequence generation, combinatorial optimization, and biological sequence design show that HCG improves mode coverage and target approximation, translating increases in the number of coordinated samplers into broader, complementary exploration.
The Well, a 15-terabyte dataset spanning 16 physical simulation domains, provides an extensive benchmark for developing and evaluating machine learning surrogate models for complex physical systems. The baselines released with the benchmark were presented by their authors as illustrative examples obtained at small computational cost, trained with zero to very limited hyperparameter tuning and under a highly constraining 12-hour per-run wall-clock limit. These results have subsequently been treated as fully tuned benchmarks, and in some cases effectively interpreted as unconstrained performance baselines. We perform a short non-exhaustive hyperparameter search for the Tucker-factorized Fourier Neural Operator (TFNO) under the same 12-hour per-run wall-clock limit, while using only a single input timestep rather than the history of up to four used by the published baselines. Despite this more restrictive setup, tuning alone reduces VRMSE below both published FNO and TFNO baselines on 15 of 17 tasks, sometimes by several orders of magnitude. This sensitivity to a modest tuning effort underscores that the released neural operator baselines should not be treated as architectural benchmarks.
AI foundation models play a pivotal role in advancing heart disease diagnosis from electrocardiogram (ECG) data, particularly for rare diseases where the lack of sufficient labeled data hinders supervised learning. However, most self-supervised learning methods treat ECGs as generic sequence or image data without exploiting underlying physiological structure. To address this limitation, we propose QECGBERT, a self-supervised ECG language model that explicitly encodes physiological semantics by tokenizing the P, QRS, T, and isoelectric background waveforms. Vector-quantized variational autoencoder (VQ-VAE) tokenization is applied to each 12-lead ECG, with tokens enriched by continuous waveform features to preserve morphology lost during quantization without increasing sequence length. This design enables compact BERT masked language modeling with short sequences, improving scalability and reducing computational cost relative to long-sequence approaches. QECGBERT applies BERT pre-training to the MIMIC-IV dataset containing over 800,000 unlabeled ECG recordings, and fine-tuning on labeled PTB-XL, CPSC2018, and Chapman--Shaoxing--Ningbo datasets for evaluation. The results demonstrate that QECGBERT achieves comparable or superior performance to published benchmarks, offering a scalable, physiologically aware approach to ECG foundation modeling for cardiac diagnosis.
Tool-using large language models (LLMs) must decide whether external information warrants action. We investigate whether their expressed credibility judgments reliably indicate behavioral resistance to misinformation. We evaluate five LLMs on 196 constructed false notices paired with legitimate counterparts and 204 archived attacks adapted to simulated tool use. When asked to act on the constructed notices, models issue harmful tool calls in 27%--99% of cases they have rated untrustworthy. Removing the preliminary rating step increases harmful execution by 15 percentage points on average, showing that credibility elicitation itself changes the behavior being evaluated. Interventions that improve credibility discrimination do not consistently yield safer actions. A simple forewarning significantly reduces harmful execution in four of five models, but neither tested generated-example prebunk outperforms it in the main evaluation. These reductions can also compromise legitimate task completion. Supervised fine-tuning on generated false notices reduces harmful execution on held-out archived attacks under the rating-first protocol, with model-dependent utility costs and weaker protection without rating elicitation. Our findings motivate evaluating misinformation defenses through both harmful actions and legitimate task completion, while accounting for the behavioral effects of the evaluation protocol.
Coding agents are given embedding search and code-graph tools on the premise that a repository's structure is something the model lacks and the tool supplies. We test that premise on SWE-bench Verified with small open-weight models, stratifying tasks by a deterministic hop depth (the shortest code-graph path from code the issue mentions to code the gold patch changes) that a companion paper defines and validates. First, we ablate a localization agent's tool set (grep only; +embedding search; +graph tools; all) on a depth-balanced 160-task split and on all 500 tasks for ten models (six Qwen3 sizes from 0.6B to 32B with five decodings each, an FP8 control, three Gemma 4 sizes; 79k episodes). Structured tools double grep-only accuracy for Qwen3-8B and Gemma 4 E4B; their gain peaks at 17-21 points at 1.7-4B and falls to 4-5 points at 32B (odds ratio 0.77-0.82 per doubling of parameters). From 8B up, about half of that gain is the form of the submission rather than finding the target; the fall from 14B to 32B is a fall in tool uptake at equal grep-only accuracy; and the same 4B size gains or does not gain from a tool depending on the checkpoint. Each extra hop multiplies the odds of correct localization by 0.50-0.67 at every size, and no tool set flattens that gradient: traces show the graph tools' gain is name lookup at hop 0 and one expansion at hop 1, and agents chain expansions in only 12-14% of episodes. A fixed repair model resolves 6-13 of 160 tasks from these localizations (oracle 26). Second, linear probes find direct call edges decodable under a within-anchor, degree-stratified protocol (AUC 0.74; 0.61-0.66 inside agent conversations), but below a lexical baseline (0.99), with two-hop edges at 0.57. Small agents carry a one-hop, largely lexical picture of their repository, in weights and in use. We release the agent, probes, logs and traces.
We propose a joint sparse sample weighting and parameter estimation (SWAPE) framework that jointly optimizes model parameters and sample weights to train a model while simultaneously assigning an individual weight to each training data point. Our framework accommodates several widely used loss functions for the model parameters, including convex, weakly convex, and difference-of-convex loss, and weakly convex sparsity functions to control the complexity of the sample weights. To solve the resulting bivariate problem, we propose an alternating minimization method in which the subproblems for updating the model parameters and weights are each approximated using majorize-minimize techniques. We provide several approximation methods for the loss functions considered in this paper, and show the overall algorithm's convergence. Numerical experiments on various learning tasks and data anomaly settings demonstrate the effectiveness of the proposed framework in minimizing the impact of undesirable data points compared to benchmark methods.
Continual learning is the ability of LLM agents to improve over time by acquiring, refining, and applying knowledge from experience while retaining useful prior knowledge. It is essential for agents learning over long time horizons, yet current evaluations largely assess individual task performance rather than improvement through accumulated experience. We introduce Continual Learning Exam (CLE), a large-scale benchmark developed by 50+ domain experts to measure experience-driven improvement in realistic professional workflows. CLE comprises 120 human-authored tasks, each organised as a five-level ladder, yielding 600 levels spanning 89 professional roles across 12 occupational domains. Agents must pass each level to unlock the next, with later levels building on experience accumulated during earlier assignments. An evaluation agent assesses each submission against human-authored rubrics. We compare a single-submission setting with one that provides diagnostic feedback after failure and allows one revision. Across 17 harness-model configurations using frontier LLMs, feedback and revision yield a median completion gain of 5.3 percentage points. Yet the best-performing system clears only 47.7% of levels and completes all five levels in just 23.3% of tasks, highlighting the difficulty of sustained learning from experience for existing frontier LLMs. As such, the dataset can also be used to post-train LLMs to strengthen their continual learning capabilities.
Large language models increasingly act as partners in everyday communication. In these interactions, useful assistance may require not only answering, but also asking the right questions that surface knowledge gaps, assumptions, and alternatives. In this study, we refer to such epistemic question-asking as model curiosity, and introduce a theory-grounded taxonomy that distinguishes curiosity-oriented questions from generic model questioning. Across four conversation datasets, we find that curiosity-oriented questions remain rare (17-28% of all questions), but are associated with more exploratory and epistemically rich subsequent user responses, with varying patterns across datasets. We further conduct a human study to evaluate the usefulness of model curiosity, finding that curiosity can encourage exploration, especially in advice, creative, and research contexts. These findings have implications for building LLMs that ask more selectively to support deeper exploration and human-AI collaboration.
Pipelines that require structured outputs from hosted language models often specify a Pydantic schema in the prompt, validate the response, and retry when validation fails. Although native structured-output modes provide an alternative, application-side validation remains important for enforcing custom constraints. Yet pipeline reliability is often summarized by a single validity rate, which obscures why outputs fail and which interventions may help. We introduce a fine-grained taxonomy of schema-conformance failures and a deterministic classifier that assigns failure categories from model responses and Pydantic validation errors. We apply this framework to 3,997 generation attempts across nine OpenAI and Anthropic model/API configurations, using questions from HotpotQA and FinanceBench. Our results show that similar validity rates can mask different failure patterns, and that dominant failure categories vary with schema constraints and task. Markdown-fenced JSON is the most common failure: among 681 fenced failures, only 48.0% pass validation after fence removal, while the remainder conceal additional violations, predominantly field-length violations. In an exploratory prompt ablation, variants that suppress fences appear to shift failures toward length violations, but the shift is consistent with exposing violations that were already present, illustrating how aggregate validity rates can obscure changes in failure composition. We derive four engineering practices for validator, schema, retry and evaluation design, three supported by our data and one, on retry feedback, proposed but not tested, and release the classifier, completions, and annotations.
Text-to-SQL systems achieve remarkable accuracy on academic benchmarks, yet performance drops dramatically in enterprise deployments. This gap stems from a fundamental assumption: existing approaches require complete schema documentation, but enterprise databases typically lack comprehensive documentation, featuring cryptic naming and implicit relationships. We present the first text-to-SQL framework that operates without schema documentation, instead learning database structure from historical query logs. Query logs encode implicit structural knowledge: which tables are queried together, how they join, and what columns map to business concepts. We construct a knowledge graph from parsed query patterns and introduce entity-driven GraphRAG retrieval that grounds user questions in domain terminology rather than relying solely on embedding similarity. Validated on both an industrial deployment with expert evaluation and the Beaver benchmark, our approach substantially outperforms similarity-based retrieval methods, the dominant industry approach, in both executability and answer quality. On Beaver, we achieve 95.7% execution validity compared to 60.8% for embedding-based retrieval, with expert data engineers preferring our system 70% to 10% on internal deployment. Our work demonstrates that text-to-SQL can operate effectively in schema-ambiguous enterprise environments, enabling natural language database interfaces "in the wild".
Estimating unobserved forces in mechanistic systems is a central task of scientific machine learning. Approximate Gaussian filtering methods that synthesize mechanistic and empirical information have shown remarkable capacity in jointly identifying latent processes together with the system state in a single forward-backward pass. The resulting latent force ODE filter framework channels information from observational data specifically into the unknown parts of the system. This work addresses open questions about calibration and stability of the method. We show that the joint posterior over state and force is the reweighted conventional latent force posterior. This uncovers a crucial dependence of the stability of the method on the prior variance of the system state, which effectively controls the flow of information from observations to the latent force through the mechanistic system. From this insight, we construct an improved algorithm that provides stable and calibrated plug-and-play inference. We demonstrate the improved framework on several simulated experimental settings.
"Why is my period late?" Sometimes, the most useful response begins with another question. Yet conversational AI systems may respond to incomplete health queries without seeking essential context or grounding their guidance in verifiable evidence. This paper introduces HerHealthGPT, an evidence-grounded conversational AI system designed to address these challenges in women's reproductive health. HerHealthGPT integrates retrieval-augmented generation (RAG), context-aware clarification, and source validation within an interactive web interface. Rather than responding immediately to incomplete queries, the system maintains conversational context, asks targeted follow-up questions, checks for premature conclusions, and grounds its responses in retrieved evidence. We evaluate the AI components using 540 English messages derived from 90 reproductive-health cases and conduct a prototype questionnaire study with 30 adults. Evidence retrieval improves concern-category accuracy from 54.1% to 59.3% compared with fine-tuning alone. The clarification router achieves 85.6% recall and 86.2% specificity on constructed targets. In the user study, 83.3% of participants preferred the retrieval-enabled configuration, while 25 of 30 considered its follow-up questions relevant. These findings suggest that integrating evidence retrieval with interactive clarification offers a promising approach to more transparent, context-aware reproductive-health information systems, while further clinical validation remains necessary.
