Papers

Retrieval-Augmented Generation (RAG) and agentic frameworks have advanced enterprise AI considerably, yet agents remain fundamentally reactive: they wait for a human query before acting. This paper argues that genuine enterprise productivity gains require proactive agents: systems that surface relevant, actionable information to workers before they ask. We propose the Context Graph, a live relational data structure that models enterprise entities, their relationships, and state transitions over time. Built on this graph, we define a Delta Detection Engine that continuously monitors state changes, a Proactivity Scorer that ranks candidate insights by urgency, relevance, and persona-fit, and a Surfacing Layer powered by an LLM that delivers ranked notifications with grounded explanations. We formalize each component, derive a unified Proactivity Score function, and provide a complete end-to-end Python implementation using NetworkX and the Anthropic Claude API. Evaluation across three generic enterprise case studies (contract lifecycle management, engineering incident response, and sales pipeline hygiene) demonstrates that context-graph-driven proactivity achieves Precision@5 of 0.83, a false positive rate of 0.11, and reduces mean time to surface from 47 minutes (reactive baseline) to under 30 second.

Sleep physiology arises from the coordinated dynamics of the central nervous system (CNS) and autonomic nervous system (ANS), as reflected by multimodal polysomnography signals including EEG, EOG, EMG, ECG, and respiration. However, existing sleep foundation models often fuse heterogeneous biosignals in a topology-agnostic manner, overlooking their physiological organization. We introduce Omni-Sleep, a sleep foundation model that uses the CNS/ANS partition as a physiological prior for topology-constrained representation learning. Omni-Sleep learns structured representations through three objectives: intra-system consistency, which captures shared subsystem-level factors within neural and cardio-respiratory signals; inter-system synchronization, which aligns subsystem trajectories to model brain–body dynamics; and latent-space masked temporal modeling, which captures long-horizon sleep dynamics. Pre-trained on over 100,000 hours of multi-center multimodal PSG data, Omni-Sleep is evaluated on sleep staging and multi-disease classification. Across datasets and modality-ablation settings, Omni-Sleep outperforms strong foundation-model baselines, showing improved label efficiency, cross-dataset generalization, and robustness to missing modalities. These results highlight the value of physiological hierarchy for generalizable sleep representation learning. Code is available at https://github.com/AutoBrain-sleep/OmniSleep.

Parameter-efficient fine-tuning adapts a large language model to one task cheaply, but across a task sequence LoRA-style methods keep stacking low-rank updates on the same frozen weight, so each new task tends to overwrite the previous ones. We present ReCoLoRA (Recursive Consolidation of Low-Rank Adapters), a spectrum-aware framework for continual fine-tuning: adapters are initialized from a randomized SVD of the pretrained weight, per-layer effective ranks are selected by an elbow criterion, and the principal subspace is adapted before residual capacity is opened. Before each new task, ReCoLoRA re-decomposes the current effective weight, rather than the original one, into a frozen residual, a slowly updated principal component, and a fresh adapter (recursive consolidation), so every task starts from the model that has already absorbed its predecessors. On a six-task continual GLUE sequence over four 7-8B backbones, ReCoLoRA attains the best final average score on three of the four backbones against rank-swept LoRA, PiSSA, AdaLoRA, and DoRA baselines while training fewer parameters; an oracle-routed task-bank variant serves as an upper bound under full task isolation. Code: https://github.com/bhqy666/ReCoLoRA.

Neural operators have become a common approach for learning PDE solution maps and accelerating numerical simulations. Transformer-based neural operators are of particular interest, since attention can learn long-range dependencies in the computational domain. However, standard attention has two major limitations when applied to PDEs: it scales quadratically with the number of computational nodes, and it lacks an explicit bias toward local interactions. To address these issues, we introduce Local Linear Transformer (LLT) for PDE operator learning. The architecture combines linear global attention with local spatial mixing, and incorporates coordinate and geometry information. We evaluate LLT on several PDE problems, including elasticity, plasticity, airfoil flow, pipe flow, and Darcy flow. The reference data for these problems span finite-element, finite-volume, and finite-difference discretizations on structured and unstructured meshes. Compared with other neural-operator and transformer baselines from prior studies, LLT achieves competitive or lower relative $L_2$ error across these problems. On matched structured discretizations, wall-clock time per training iteration is reduced by factors of 1.8 to 2.5 relative to Transolver. We also scale the approach and apply it to a three-dimensional car aerodynamics dataset with 32,186 unstructured mesh points per sample. Together, these results indicate that LLT provides an accurate and computationally efficient operator for PDE problems across discretizations, mesh types, and problem settings.

Temporal graphs are ubiquitous in real-world applications and Temporal Graph Networks (TGNs) have achieved superior predictive accuracy. Understanding which historical events drive model predictions can enhance trustworthiness of TGNs. Existing explanation methods overlook the memory module, the core component that records and updates node histories, leaving the influence of past events unexplored. To address this, we attribute TGNs predictions through the topology attribution tree and memory backtracking tree. The topology attribution tree captures the influence of neighbors and their memory vectors, then the memory backtracking tree quantifies how historical events shape node memory vectors. We apply the LRP in TGNs, ensuring that the total contribution of events equals the logits of model. Finally, top-k selection may be unfaithful due to the nonlinear mapping from logits to probabilities, we design optimization objectives to identify the important events. Experiments on nine temporal graph datasets, spanning node property prediction, link prediction tasks and graph classification tasks, show that our method provides faithful explanations and outperforms state-of-the-art baselines. The code is available at https://github.com/yazhengliu/MemExplainer

We argue that generative AI can degrade research by eroding the very practices through which scholarly judgement is formed. As constitutive conditions for the production and validation of knowledge, these practices cannot be reduced to the outputs of research, which is what AI tools so effectively simulate. When scholars delegate central tasks of inquiry to Large Language Models (LLMs), they may stop engaging in these practices and lose the formation they provide: the output may improve while the researcher fails to develop. Against this risk, merely keeping humans "in the loop" as prompters or quality checkers of AI outputs is insufficient to preserve research as a site of intellectual formation. What is needed instead is a renewed commitment to research as a lived practice in which judgement is formed gradually, often through friction, and participation in a scholarly community. We defend it because it rests on four sources and warrants of research that cannot be automated: tacit knowledge, personal commitment, socialisation, and deep reading. These enact what we call second scholarship, i.e., the reappropriation of scholarly craft, chosen out of a critical experience of what AI tools can and cannot do. Researchers who practise it may delegate tasks to these tools, provided that their own judgement in the delegated competence is already formed, that they can answer for the output, and that they keep exercising the competence themselves. What cannot and should not be delegated becomes what research communities must value and answer for. This is what is left for us.

On-policy distillation is a powerful way to transfer reasoning ability from a strong teacher to a smaller student: the student samples trajectories from its own policy, and the teacher provides dense token-level supervision on the states the student actually visits. However, this supervision is not always reliable: a teacher can assign high likelihood to plausible but incorrect solutions, or low likelihood to correct student solutions that follow different reasoning paths. Unconditionally distilling the teacher can therefore reinforce bad modes or erase useful student behavior. To address these limitations, we introduce RG-OPD: Reward-Gated On-Policy Distillation that uses verifier feedback to decide when teacher logits should be trusted. RG-OPD bridges sparse verifier rewards and dense teacher logits, preserving token-level supervision while filtering misleading teacher signals. Across reasoning and coding benchmarks, RG-OPD produces stronger distilled students, outperforming both vanilla reverse-KL distillation and the recent TSD-KD baseline. At 1K generation length, RG-OPD improves over reverse-KL by 2.9 points and over TSD-KD by 4.9 points; in the long-generation setting, it improves over the untuned student by 8.2 points. Our code is available at https://github.com/UoC-tail/RG-OPD.

The language models that underpin agents have seen a rapid rise in performance on function calling benchmarks. However, the metrics used in the training and evaluation of these models often encourage models to make positive claims even when the answer is uncertain, leading to hallucinations. Such hallucinations can be disastrous when language models are trusted to use function calls to make decisions in high stakes applications. To that end, we propose an agent evaluation metric that takes into account the negative outcomes associated with incorrect function calls. Further, to catch hallucinations before they can cause harm, we propose a lightweight trainable filter that can quantify a language model's uncertainty and remove potentially harmful function calls. By training that filter to detect and suppress uncertain function calls without modifying the underlying model, we demonstrate a practical path toward agents that know when to say "I don't know," a property we argue is essential to production reliability.

Optimizer selection for large-scale model training has become a system-level design decision constrained jointly by compute, memory, tuning budget, and task diversity, yet the landscape of over one hundred methods remains fragmented. We therefore present OmniOpt, a unified survey and benchmark cookbook of optimizers for the research community. OmniOpt rests on four coupled components. First, we treat every optimizer update as a structured transformation through a five-stage meta-pipeline, and show that most methods engage only one or two of these stages. Second, we use norm-constrained linear minimization oracles (LMOs) to unify different optimizers. Third, these two views ground a dual-dimension taxonomy, one dimension assigning each method to a mechanism family and the other recording the measurable training objectives it aims to improve. Fourth, and at the core of this paper, we instantiate the full taxonomy in a unified cross-domain benchmark spanning representative optimizers, model scales, and training regimes from language model pretraining to image classification, systematically analyzing each method family across multiple effect objectives and laying out their trade-offs. OmniOpt thus supplies the research community with an operational coordinate system for selecting optimizers under explicit mechanism and objective assumptions, and charts a direction for the future development of the optimizer community.

Conditional functional dependencies (CFDs) are functional dependencies with a restricted scope: they specify the context in which a dependency holds and are useful for data-quality tasks, specifying complex integrity constraints, and extracting valuable insights from data. We study the CFD discovery problem, which is computationally demanding. We build on the state-of-the-art CFDFinder algorithm and introduce a set of algorithmic and engineering improvements, including a parallelization strategy, to produce ParCFDFinder. Our implementation is integrated into Desbordante - a high-performance open-source data profiler written in C++ that exposes a Python interface, enabling CFD discovery to be invoked from any Python program. Experimental results show that our enhancements speed up the algorithm by up to $318\times$ ($118\times$ on average) and reduce memory usage by up to $23\times$ ($14\times$ on average) compared with the existing Java-based implementation of Metanome. Integrating ParCFDFinder into Desbordante makes it possible, for the first time, to conveniently discover CFDs on datasets with hundreds of thousands of rows on a commodity machine within a reasonable time.

We propose a unified algebraic framework for classification performance evaluation covering binary, multiclass, multilabel, ordinal, hierarchical, cost-sensitive, and soft-label settings. Actual and predicted labels are represented as binary indicator matrices, where three aggregation operators (global, column-wise, row-wise) correspond directly to micro, macro/weighted, and exemplar averaging. Any binary measure expressed in terms of the four confusion-matrix counts extends to all these settings by substituting an operator, with no measure-specific derivation. We show that structural properties governing an extension are derivable from the binary formula. Micro-averaging equals denominator-weighted macro-averaging precisely for aggregation-decomposable (linear-fractional) measures, a strict class characterised algebraically. For soft ground truth, we prove from t-norm axioms alone that the product t-norm is the unique choice whose confusion counts preserve marginal memberships. In multiclass settings, micro-precision, micro-recall, and micro-F1 collapse identically onto accuracy. Furthermore, binary skew-invariance transfers unconditionally to multilabel aggregation, but only partially to multiclass problems. For measures with a linear numerator and prediction-independent denominator, the optimal decision threshold is the share of the numerator weight favouring a negative prediction, revealing when standard training targets the measure. Under one-hot multiclass encoding, a measure fails to attain its theoretical minimum whenever its zero-true-positive value still depends on true negatives, establishing non-trivial performance floors even for completely incorrect classifiers (e.g., zero correct predictions on 10 classes yields a label accuracy of 0.8).

Physical sensing and actuation noise floors should inform how much belief resolution a decision-making system can reliably use. We introduce Finite Reliability Representations (FRR), a framework for covering belief spaces by reliability cells: regions within which the optimal action-value function Q*(b,u) varies by at most a tolerance epsilon, uniformly over actions. The framework is formulated on beliefs rather than states and uses a cover rather than an equivalence quotient, because approximate decision-closeness is not transitive in general. A central technical point is that noisy Bayesian updates should not be treated as globally contractive on arbitrary beliefs. We therefore separate three objects: the fixed-observation filter map, the predictive observation law, and the controlled belief-transition kernel. For nonlinear continuous-state systems, FRR is obtained under a reachable-set Lipschitz modulus for the belief-transition kernel. For finite-state POMDPs, the same construction becomes exact on the belief simplex: prediction is linear, Bayesian correction is a normalized positive linear map, sensor noise enters through observation-distribution distinguishability, and actuation uncertainty enters through an action-execution channel. Under the corresponding action-value Lipschitz condition, an FRR cover supports a cell-constant policy whose suboptimality is bounded by 2 epsilon/(1 - gamma). We also introduce reliability entropy, the logarithm of the minimal number of reliability cells, as a measure of certified decision-relevant belief complexity. The framework distinguishes representation sufficiency from fundamental performance floors imposed by sensing, process, and actuation noise. It applies to finite POMDPs, linear-Gaussian filters, locally linearized nonlinear filters, and particle-filter implementations through analytic or empirical certification of reliability cells.

Learning from few labeled examples is a central challenge in tabular machine learning, and it becomes the binding constraint in domains where labeling is costly, such as crowd monitoring during Hajj and Umrah. Tabular foundation models, which predict from only a handful of examples without task-specific training, were recently introduced to address this very-few-label regime. In this study we test them on crowd-state classification to assess how much they help when labels are scarce, and we compare them against standard machine learning methods to characterize the accuracy and efficiency trade-offs between the two approaches. Using three real datasets we evaluate different machine learning models, in untuned and tuned forms, against three foundation models. Results show that no single family is best everywhere. The right choice depends on the label budget. When labels are very few, foundation models lead. As labels grow, tuned conventional models catch up and significantly surpass the foundation models on the more structural geometry target. Efficiency separates them further where tuned machine learning models incur a large tuning cost that foundation models avoid, although foundation models reprocess their context at every prediction. We summarize these results as a practical map of which approach to prefer under a given label budget and computational budget.

While decoders have rapidly scaled, encoders have remained largely unchanged since BERT. We revisit this disparity by frozen backbone evaluation via probing. Under this lens, the representations of BERT encoders become increasingly $\textit{unexploitable}$ by frozen probes, despite improved perplexity. The misalignment originates in BERT's flat design, which couples representation learning to the token reconstruction loss. We propose $\textbf{CrossBERT}$, a two-part architecture that separates the learning of high-quality encoded representations from the rigid grounding of token reconstruction. This design further enables high masking ratios ($\ge 50\%$) and gradient collection over all tokens via a $\textit{Complementary Masking Strategy}$, respectively increasing throughput by $1.5$ to $2\times$ and sample efficiency by $2\times$. Overall, CrossBERT demonstrates monotonic scaling and superior performance on MTEB(eng, v2) and frozen GLUE benchmarks.

Bangladesh has an estimated 1.17 mental-health professionals per 100,000 population and only six child psychiatrists nationwide. No Bengali-language, culturally adapted tool exists for early screening of abuse-related psychological trauma in children. We present ShishuRaksha AI, a decision-support (not diagnostic) framework that fuses four screening modalities: validated questionnaires (SDQ, CPSS), Bengali narrative text, House-Tree-Person (HTP) drawing features, and facial affect. The fusion is training-free and clinically weighted, uses cross-modal attention, and includes a single-modality override rule. Every risk score is explained through clinically weighted, perturbation-based additive attribution and rendered as a bilingual (Bangla/English) report with referral routing to national child-protection services (OCC, DSS, NMHH) under the Children Act 2013. No clinical dataset of abused children can be collected ethically at this stage, so we introduce a noise-aware synthetic benchmark (500 cases, 116 positive [23.2%], four deliberate noise layers, literature-grounded HTP priors) and evaluate tree-ensemble surrogates of the fusion design (facial channel excluded) under 5-fold stratified cross-validation. The fused model reaches an AUC of 0.874 [0.834-0.908], against 0.756 [0.705-0.803] for an SDQ-only baseline, with ablation, operating-point, subgroup, and calibration analyses. We state all limitations openly, including synthetic-only data, no held-out set, text-feature circularity, and an urban-rural subgroup gap. This work is a feasibility study and a design contribution toward ethically deployable child-protection screening in low-resource settings.

Modern language models generally represent text as sequences of discrete token embeddings, an assumption deeply rooted in current practice but rarely questioned. We challenge this representation, especially for Chinese, by replacing index-based token embeddings entirely with a single rasterized image of the character sequence, processed by a vision encoder composed of a shared ResNet and a shallow Vision Transformer. To isolate the role of input representation, we construct a dual-branch controlled framework in which both a Vision-based model and an index-based baseline share an identical decoder backbone, training objective, optimizer, and data curriculum. Any performance difference is therefore attributable to the input modality only. Across all tested decoder backbones, the Vision-based model consistently outperforms the baseline, reaching a peak accuracy of 0.429 versus 0.355 for the index-based baseline—a 21% relative improvement—while converging in about half the number of training epochs. The advantage emerges especially within the first five epochs (under 21% of total data) and persists under moderate character corruption: the corrupted Vision model matches the clean index-based baseline. Ablation studies reveal that the advantage requires both spatially coherent input and a ViT encoder with 2D positional encodings. A cross-script comparison on English shows the advantage does not transfer directly to alphabetic writing systems, suggesting that the uniform visual density and radical structure of Chinese characters are enabling conditions. These findings suggest that transformers are more modality-agnostic than commonly assumed, and that discrete tokenization is not a fundamental requirement for Chinese language modeling.

Advanced neural technologies in speech synthesis and voice conversion (VC) have introduced severe risks to personal privacy, necessitating robust Speaker Anonymization Systems (SAS). Existing SAS approaches modify voice characteristics in the hand-crafted feature space or speaker embedding space, often struggling to provide sufficient identity variance across generated voices. In this paper, we propose NouveauVoice, a novel pseudo-speaker generation framework based on a Hierarchical Deep Variational Autoencoder (NVAE). Integrated as a standalone plug-in module on top of state-of-the-art architectures (FACodec and CosyVoice2), our approach leverages tractable sampling and the Evidence Lower Bound (ELBO) objective to synthesize highly expressive pseudo-speaker embeddings with significantly enhanced speaker diversity. Evaluating our framework under a protocol similar to the VoicePrivacy Challenge alongside Maximum Mean Discrepancy (MMD) analysis, we demonstrate that NouveauVoice achieves strong identity concealment, yielding an Equal Error Rate (EER) exceeding 38% against an automatic speaker verification attacker model. Our system shows a reasonable trade-off between strict anonymity, rich pseudo-speaker diversity, and downstream speech utility, such as intelligibility and emotional expressiveness.

Memes have become influential communication tools on social media, combining viral visuals with concise messaging to convey impactful ideas. While substantial research has examined the affective dimensions of memes, key challenges such as detecting harmful content, identifying cyberbullying, and performing accurate sentiment analysis remain critical, largely due to the need for deeper contextual understanding. In this paper, we introduce MemeEvidenceDetect, a hybrid task aimed at analyzing a meme and its contextual information to identify specific sentences that explain or elucidate its meaning and humor. To support this task, we present BanglaMemeEvidence, a curated dataset of 2,917 Bengali memes, emphasizing its significance as a resource for the Bangla language. Each meme is annotated with natural language explanations, including Meme OCR, Meme Context, and Evidence Sentences, alongside relevance scores that reflect the relationship between a meme and its corresponding annotations. To address the gap in dynamically inferring a meme's context, we propose BengaliMemeEvidenceNet, a hybrid multimodal framework that integrates textual and visual features for comprehensive meme representation. Our experiments demonstrate the effectiveness of BengaliMemeEvidenceNet, achieving an F1 score of 0.74. To the best of our knowledge, this is the first study to focus on evidence detection in Bengali memes, marking a notable step forward in the analysis of memes in low-resource languages.

Low-dimensional projections support interactive visual analysis of high-dimensional data embeddings, but their structure often does not align with analyst-defined semantic relationships. Recent LLM-augmented semantic steering methods address this gap by externalizing analyst intent from user-defined groups of seed examples, but they propagate intent through per-item LLM reasoning, causing LLM calls and cost to grow linearly with collection size. We propose a scalable semantic steering method that shifts semantic computation from individual items to user-defined groups. A single LLM call generates structured profiles for all groups, which are embedded and combined with seed centroids to form hybrid semantic prototypes. The method then propagates intent without retraining, using embedding-space soft assignment, abstention, and alignment-scaled updates before reprojection. On a 5K-document LitCovid corpus, our method achieves global alignment comparable to per-item LLM steering while reducing LLM calls by over three orders of magnitude. An image case study shows that the same prototype-based mechanism extends to multimodal embeddings. These results suggest that group-level representations can make semantic steering more practical for larger embedding collections.

Causality has become an increasingly important tool for gaining a deeper understanding of complex systems. Among various causal analysis methods, causal discovery, which identifies causal relationships among variables from data, has been widely used to uncover underlying causality in diverse processes. However, while multistage processes are prevalent in many fields, existing causal discovery methods may produce counterintuitive results, given the known process knowledge, and may not be computationally efficient for handling large datasets typical of multistage processes. To address this gap, we propose a novel causal discovery method called Order-based Causal Discovery for Multistage Processes (OCDM). OCDM is designed to infer the causal structure of multistage data while preserving their inherent hierarchical and sequential structure by explicitly incorporating process knowledge into the causal discovery process. Specifically, we propose a structural knowledge-informed order-inferring algorithm that infers the causal order of variables by incorporating information about the stage from which each variable originates, based on an order-based causal discovery framework naturally suited for inherently ordered multistage data. Furthermore, to eliminate spurious edges from the initial causal graph generated based on the inferred causal order, we introduce a novel pruning technique using stochastic gated neural networks, which offers greater computational efficiency compared to existing methods. Through experiments on various datasets, we demonstrate that OCDM effectively infers the causal structure of multistage processes, outperforming existing methods.

Large language models are increasingly deployed as IDE-integrated coding agents that decompose tasks, generate and edit files, run code, and refine outputs over many turns. Yet their safety is still often evaluated as if they were chatbots: one harmful prompt, one response, judged in isolation. We introduce workflow-level jailbreak construction, a failure mode in which a harmful objective is assembled across ordinary stages of a software-development workflow rather than generated through a single direct prompt. Using GitHub Copilot in Visual Studio Code, we study four closed-weight backends: Claude Sonnet 4.6, Claude Haiku 4.5, Gemini 3.1 Pro, and Gemini 3.5 Flash. Across 204 prompts from Hammurabi's Code, HarmBench, and AdvBench , the models show near-complete refusal under direct chat, CSV-read, and single-step code-fix baselines, with only 8/816 successful responses in each baseline condition. Under the full workflow, however, the same prompts and backends produce 816/816 unsafe teaching-shot completions, all independently confirmed by two expert evaluators under a strict rubric. These results show that conversational refusal benchmarks can substantially overstate the safety of deployed coding agents and motivate defenses that reason about safety across multi-turn IDE workflows and their generated artifacts, not only individual chat turns.

World models are rapidly becoming a core infrastructure for embodied intelligence and interactive agents: they provide controllable simulators in which agents can perceive, act, forecast, and acquire scalable experience. Yet current video generation world models are still organized around isolated control interfaces, such as camera trajectories, robot actions, or hand-joint signals. This fragmentation is increasingly a scaling bottleneck. The central challenge is not the absence of controllable generators, but the lack of a unified and extensible learning framework that can absorb heterogeneous action supervision while preserving a shared model of world dynamics. In this work, we introduce Worldscape-MoE, a Mixture-of-Experts world model built on Diffusion Transformers for scalable heterogeneous action control. Our key observation is that different controls specify different interfaces to the same underlying world: although their representations differ, they constrain shared physical regularities, scene dynamics, and interaction semantics. Worldscape-MoE operationalizes this observation through modality-aware control injection, shared and control-specific experts, and a progressive MoE tuning strategy that supports continual extension to new action modalities. Experiments across locomotion, robotic manipulation, and egocentric hand control show that heterogeneous supervision improves rather than interferes with individual control capabilities. Worldscape-MoE achieves strong results on WorldArena, improves locomotion and hand-control metrics, exhibits robust out-of-distribution generalization, and demonstrates scaling behavior as additional control data and experts are integrated.

Language models can reach the right normative verdict for the wrong reason. We introduce NormWorlds-CF, a solver-verified environment for counterfactual normative reasoning in executable rule worlds. Its deterministic solver produces final answers, proof and falsification certificates, argument statuses, support sets, and paired-world change labels, enabling supervision and evaluation without LLM judges. The benchmark contains staged SFT diagnostics and a compact paired-world task with 270 root families and 1080 canonical-to-variant pairs. The SFT diagnostics show that final-answer supervision can saturate verdict accuracy without inducing falsification competence: answer-only SFT reaches perfect answer accuracy but scores zero on joint falsification certificates, while full-mix training with targeted replay reaches strong all-task accuracy (0.99). For the structured-change task, we introduce metamorphic-relation GRPO (MR-GRPO), a class-conditioned reward for GRPO that gives partial credit for relation families and solver-visible change fields. In matched Qwen3-1.7B continuation experiments, MR-GRPO improves held-out relation accuracy and relation-family correctness, and reduces wrong-family error, compared to sparse and answer-only GRPO. In Qwen3-4B three-seed validation, sparse reward preserves coarse relation labels best, answer-only reward improves answer-change but weakens relation-family structure, and MR-GRPO leads on answer-, support-, and status-change fields as well as class-conditioned MR and change-presence. These results show that verified counterfactual structure can shape post-training beyond final answers, while exact full change-record generation, invariant subtype recognition, and out-of-distribution (OOD) transfer remain open problems.

This study independently replicates and extends the Natural Language Tools (NLT) framework of Johnson et al.~(2025), which questions the use of structured tool calling in large language model (LLM) agentic systems. We evaluated NLT across 14 models and 8,560 trials, adding newer frontier, reasoning, and open-weight models to the original set. The results confirm the core findings and add detail. NLT improves tool-calling accuracy by 14.9 percentage points overall (62.3% versus 47.4% structured) and reduces critical errors by 93% (51 versus 755 errors). The gains depend on model capability: models without native tool calling, reasoning models, and smaller models gain substantially (+24.0pp to +43.1pp), while heavily optimized frontier models (GPT-5, Gemini 2.5 Pro) show smaller or reversed advantages. This matches recent analyses of reinforcement-learning-optimized tool use (Martinez, 2025). NLT also cuts token usage by 25.2%. The reliability and efficiency advantages compound in recursive agentic workflows, where agents chain many tool calls across sub-agents: a structured failure triggers retries, fallback routing, and coordination overhead, while NLT avoids most of that cost at the source. This work makes three contributions: (1) the first independent validation of NLT using open-source tooling, (2) evidence that model capability moderates NLT's advantages (Chen et al., 2025; Zhang et al., 2025), and (3) a measurement of NLT's reliability benefit (93% fewer errors), its most deployment-relevant property given the known fragility of structured tool calling. NLT is a practical alternative to structured tool calling, especially for production systems that value reliability over parseability.

The reproducibility crisis in scientific research has received widespread recognition, thereby increasing the importance of meta-analyses that integrate statistical analyses from multiple studies. However, statistical methods often have ambiguous and implicit underlying assumptions, which can lead to their erroneous applications and interpretations. To address this issue, we propose a formal verification framework for statistical Python programs. Specifically, we present Why3-py, a Python front-end for the Why3 verification platform that transforms Python code into verification-oriented WhyML representations, addressing the challenges arising from Python's dynamic typing and runtime polymorphism. Furthermore, we extend the StatWhy tool to support the verification of meta-analysis methods. These tools enable meta-analysts to identify overlooked assumptions and misuse of analyses, and to verify the correct use of hypothesis testing and meta-analysis methods in Python code.

We study the online routing problem in large language model serving, where requests arrive sequentially and must be dispatched to parallel decode workers under tight batch-size and KV-cache constraints. Unlike widely used routing heuristics that are not tied to explicit service-level objectives (SLOs) and offer limited control over latency-throughput trade-offs, we introduce a multi-objective optimization framework that formulates routing as an online linear programming with interpretable decision rewards. We apply an efficient bid-price control policy based on the online linear programming that admits requests when their SLO-weighted benefit exceeds their shadow prices. To meet millisecond decision requirements, we develop a warm-started, projected first-order updates that track the evolving dual shadow prices online with predictable runtime. We integrate our router into the Vidur simulator and demonstrate substantial improvements over standard baselines across multiple SLO regimes, including end-to-end latency, time-to-first-token, throughput, and tail performance. A big picture from our result: a science-based approach outperforms others based on heuristics.

Self-evolving frameworks usually optimize task solutions while treating the surrounding harness as fixed. We introduce Harness-Aware Self-Evolving (HASE), an agentic reinforcement-learning framework in which a single model can generate task solutions or edit selected harness components in a multi-turn action space. HASE enables a single Qwen3-8B model to match the text-classification performance of a GPT-OSS-120B model that uses Claude Code as the harness proposer. In alpha factor mining, HASE outperforms the reported GPT-OSS-120B baseline. HASE also repairs imperfect evaluation components and converges to state-of-the-art performance in circle-packing algorithm discovery. These results show that HASE improves the harness and the solution through one unified agentic process.

LLMs can be conveniently adapted to a diverse set of tasks, e.g, prediction, question-answering tasks, etc, using appropriate prompts with few-shot examples. Biased or harmful concepts, e.g. gender or bio-weapons, present in pre-trained LLMs can lead to unsafe or unethical responses for many such prompts. Removing such undesirable concepts robustly across different prompt types remains a challenging problem, since existing unlearning methods typically ignore the impact of prompt variation. In this paper, we explore a novel adversarial approach to use a joint prompt for the main task and concept task prediction. We show that fine-tuning using the “worst prompt type” for concept prediction (with the highest concept accuracy) improves the average unlearning performance over a fine-tuning method that uses a combination of all prompt types. Our proposed method, MPSelectTune, is a two-stage approach that minimizes the concept accuracy of the highest accuracy-prompt type, after fine-tuning using a novel multi-task loss using multiple prompt types. Experimental results on four benchmarks show $2 - 15\%$ main task accuracy improvements over recent baselines and while reducing the worst-case concept accuracy by up to $17\%$ compared to recent baselines.

Multi-Meta-RAG improves retrieval for multi-hop question answering by filtering a vector store on metadata, specifically the news source and the publication date, which it extracts from each query using GPT-3.5. This extractor requires a proprietary API call for every query, and its output is not restricted to the allowed sources: even though the prompt specifies 49 sources, GPT-3.5 generates 128 distinct source names, 79 of which do not appear on the list. We show that the source part of this extractor can be replaced by a local, deterministic probe trained on the hidden states of a small open-source language model. The probe is a multi-label logistic classifier on the mean-pooled hidden states of one layer, trained with a class-balanced loss on the gold evidence sources. Using 5-fold cross-validation to select the layer, we evaluate four models ranging from 135M to 1.5B parameters out-of-fold across all 2556 MultiHop-RAG queries. The probe reaches a set-exact accuracy of 90.9%, compared to 88.0% for a substring baseline and 80.9% for GPT-3.5. The difference comes from null queries, for which GPT-3.5 never predicts an empty source set. Across the remaining queries, all three methods achieve set-exact accuracies within 1.1 percentage points of each other. Unlike earlier probing work, we use the probe as a component of a retrieval pipeline rather than for interpretation. For all four models, the cross-validation selects one of the first layers, in contrast to the middle layers reported for retrieval tasks. We attribute this to the source name appearing verbatim in 95.4% of gold query-source pairs. The 135M model is 1.5 percentage points less accurate than the 1.5B model, so the filter requires only a partial forward pass through a small model and one linear layer. The code is available at https://github.com/mxpoliakov/Multi-Meta-RAG.

Accent normalization (AN) seeks to convert non-native (L2) accented speech into standard (L1) speech while preserving speaker identity. The current techniques either require naturally recorded parallel L1-L2 speech for training, or suffer from quality degradation when supervised by synthesized targets. In this paper, we present TokAN, a token-based accent normalization framework that operates on self-supervised discrete speech tokens extracted from a L1-L2 jointly trained vector-quantization (VQ) tokenizer, without the need of synthetic supervisory speech. An autoregressive encoder-decoder model performs token-to-token conversion, translating L2-accented token sequences into the tokens of standard voice. We also introduce reinforcement learning (RL) post-training based on Group Relative Policy Optimization (GRPO), using word error rate and accent classifier confidence as complementary rewards. A non-autoregressive flow-matching synthesizer recovers the Mel-spectrogram from the converted tokens, conditioned on the source speaker embedding. We also develop a flow-matching duration predictor that supports total-duration-aware synthesis, making TokAN applicable to duration-critical tasks such as voice dubbing and live casting. Experiments on seven English accents demonstrate that TokAN reduced the word error rate from 12.40% to 9.89% after supervised fine-tuning, and further to 9.23% after RL post-training, consistently outperforming frame-to-frame, direct flow-matching, and prompt-based token-conversion baselines in terms of accent reduction and intelligibility.