Papers

Despite remarkable progress in language understanding and generation, decoder language models continue to exhibit striking failures on numerical data. Numerical examples are typically serialized as text by concatenating feature names with their values (e.g., fat:3.4), while conventional task adaptation applies loss only to the downstream prediction target. This leaves two structures weakly supervised: within-value sensitivity, where digit position determines numerical magnitude, because token cross-entropy contains no explicit numerical metric; and across-field structure, where measurements in the same record are statistically dependent, because target-only tuning provides no direct loss on the measurements or their dependencies. We introduce MoMaTune (Model, Match, then Tune), a two-stage post-training strategy organized around these supervision gaps. Stage 1 applies Model, ordinary next-token prediction densely over label-free serialized records, and optionally adds Match, a same-record consistency objective that aligns representations of different partial views of the same example. Stage 2 applies Tune, supervised fine-tuning for downstream target prediction. We evaluate MoMaTune on 169,285 records from eight public datasets, with 339 of 340 input fields numerical. At 1.5B parameters, Model raises the eight-dataset masked-value imputation aggregate from 0.106 to 0.401, while the three-dataset correct-context gain rises from 0.058 to 0.167. On diagnostics, Model also increases sensitivity to numerically consequential digit changes, while Match improves robustness to informative partial views. Adding Tune yields higher target-supervised embedding aggregates for each tested Stage-1 parent, and the two-stage recipes obtain higher aggregate scores than Tune alone. The same pattern is observed at 4B, 9.6B, and 27B parameters when model embeddings are used for downstream prediction.

Generative artificial intelligence (GenAI) models, particularly large language models (LLMs), are increasingly used in educational settings for tasks such as item generation, automated scoring, and feedback delivery. The quality of these outputs depends not only on the AI model but also on how users construct the prompts that guide model behaviors. Current prompt engineering practices are largely ad hoc, relying on trial and error rather than principled design. This lack of structure introduces validity threats including construct misrepresentation, construct-irrelevant variance, and limited reproducibility. We propose that Evidence-Centered Design (ECD), a framework in educational measurement for structuring valid inferences from observed performance, can serve as a principled foundation for prompt engineering in educational applications. We map the three core ECD models, the student model, the evidence model, and the task model, onto the components of educational prompt design, and describe a workflow for ECD-based prompt engineering. We derive five design principles from this mapping and present an empirical study contrasting ECD-designed prompts with state-of-the-art (SOTA) prompting baselines on an automated scoring task. Our empirical study demonstrates the usefulness of this framework on automated scoring. Results suggest that treating prompting as assessment design rather than informal instructions strengthens the validity, reliability, and fairness of AI-generated educational outputs.

Words are increasingly becoming an interface for action: large language models (LLMs) can now invoke tools, operate devices, and make decisions that affect the physical world. However, current benchmarks largely evaluate LLMs through static text responses, offline tool calls, or simulated tasks, leaving open the question of how well they can act in real physical environments. We present a benchmark for measuring LLMs in closed-loop physical interaction, where models control a harness system connected to real-world components including robotic arms, cameras, and sensors. In each task, the model must issue instructions, receive physical feedback from the environment, interpret the resulting state, and adapt its next action accordingly. This setting evaluates capabilities beyond language understanding, including grounded perception, action planning, tool selection, spatial reasoning, feedback-driven correction, and robustness to execution uncertainty. Through experiments with representative LLMs, we characterize their ability to translate linguistic reasoning into physical actions and identify common failure modes such as hallucinated state changes, poor recovery from unexpected outcomes, and unstable long-horizon plans. Our benchmark aims to provide a systematic foundation for evaluating LLMs as interactive agents that connect words, tools, and the physical world.

Unified vision-language models (VLMs) perform both text-to-image (T2I) and image-to-text (I2T) generation with one shared Transformer backbone, yet how the backbone processes the condition in either direction remains poorly understood. We study this question from the perspective of conditional guidance: we probe every attention head to show the activation difference between two branches of classifier-free guidance (CFG) with different conditional guidance. On two unified VLMs of autoregressive and masked-diffusion paradigms, the probing reveals a subset of heads that are more sensitive to the change of condition while others are agnostic to it. This characteristic pattern holds in both T2I and I2T generation and is largely stable across data samples, demonstrating the internal mechanism of unified VLMs. Motivated by this, we propose a new inference-time intervention approach for unified multimodal generation, named HeadTuned-CFG, which uses the probed condition-change awareness to select and tune the conditional guidance at the attention-head level. Experiments show consistent benefits in both multimodal generation directions of unified VLMs without additional training, and in particular substantial improvements over strong baselines in T2I generation. These results further support the key finding of our probing analysis: the observed pattern of functional specialization is an intrinsic property of unified VLMs.

AI agents can evolve by using environment feedback to revise their harnesses, the software managing model inputs, tool use, memory, and execution, while keeping foundation models frozen. Existing studies focus on single agents, leaving joint harness evolution in multi-agent systems (MAS) underexplored. A single global optimizer faces growing context and workload as the system scales, while access restrictions may prevent sharing agent internals or system-wide feedback. We propose HarnessNet, a framework for hierarchical harness evolution in MAS. Its basic unit, Agentron, comprises an evolvable forward harness, a foundation model, and a local optimizer that can observe and modify only that harness. Agentrons form a layered graph through which environment feedback propagates backward, with upstream optimizers receiving node-specific feedback from the downstream layer. This decomposes joint harness evolution into smaller, coordinated tasks with separate contexts, reducing individual optimizers' context and workload while keeping access local. Optimizers need not access all agent trajectories or raw environment feedback. Experiments on AgentBench with multiple foundation models demonstrate the feasibility of training HarnessNet. Compared with a single global optimizer, HarnessNet achieves comparable performance gains while respecting local privacy and requiring fewer tokens and less time per optimizer call. A migration study shows that the evolved harnesses remain jointly effective after replacing the foundation models, suggesting transferability across models.

A recurrent-depth language model reuses one block of weights several times in a row. The number of repetitions, which we call loops, is chosen when the model runs, not when it is trained, and running fewer is cheaper and faster, so a deployment may choose to do it. Running fewer loops also makes these models less safe, and the loss of safety is larger than the loss of anything else we measured. We study four released Ouro checkpoints and Huginn-0125. At half of its trained depth, Ouro-1.4B-Thinking still solves 0.692 of the GSM8K maths problems, against 0.775 at full depth, keeping 89% of its accuracy. Over the same reduction, the rate at which it helps with harmful requests rises from 0.038 to 0.205, 5.3 times higher. The information is not lost: a linear classifier reads whether a request is harmful from the model's internal state after a single loop, with an ROC AUC of 0.967 against 0.986 at full depth. What the missing loops remove is the refusal that should follow. We measure the effect with four instruments, one generating no text at all, and reproduce it on a second model family. Removing a single direction from Ouro-1.4B's internal state at full depth drops its refusal rate on harmful prompts from 0.950 to 0.130, identifying what the shallow model is missing. Two run-time repairs do not help: the models' early-exit rules do not spend more loops on harmful requests, and choosing a depth per request does not recover the loss. Changing the training does. Standard safety training makes the shallow model safe only by refusing 0.865 of harmless requests, while training at randomly chosen depths reaches a separation of +0.399 between harmful and harmless refusal at one loop, against +0.118. On Huginn-0125 it keeps harmful help at or below 0.040 at every depth from 1 to 32 loops, where standard training reaches 0.194, and it costs no accuracy at full depth.

Deep research requires connecting multiple steps of search, reasoning, and computation, with intermediate findings guiding what to investigate next. As investigations grow, agents must maintain research state across model calls and check findings before errors propagate into subsequent work. We introduce PANINI-DR, a multi-agent harness that separates coordination of the overall investigation from the research and verification of individual subquestions. A shared workspace, built on the Generative Semantic Workspace representation, records subquestions, their dependencies, findings, and supporting evidence. Researchers investigate individual subquestions, while a separate verification agent checks proposed answers against those subquestions and their evidence. Verification feedback guides further investigation, and recorded findings inform subsequent assignments. We evaluate PANINI-DR on BrowseComp-Plus, FRAMES, and MoNaCo against seven state-of-the-art deep-research harnesses, all using a shared language-model backbone and frozen corpora. PANINI-DR achieves the strongest average performance across the three benchmarks. On BrowseComp-Plus, it achieves 73.3% accuracy, outperforming the next-best harness by 16.0 points. Its advantage is strongest on hard questions, where it answers nearly twice as many as the next-best harness. Ablations further show that structured planning and persistent state improve performance, as does checking each finding against its cited evidence before it is reused. Because a model can answer correctly from memory alone, we also analyze the evidence retained by each harness. PANINI-DR achieves the highest answer recoverability from saved evidence on all three benchmarks, while complementary claim-support audits show stronger evidence-backed answers. Accounting for this evidence changes how harnesses compare and further strengthens PANINI-DR's advantage.

Large Language Models (LLMs) have been shown to be computationally expensive and vulnerable to jailbreak attacks, in which adversarial prompts bypass safety alignment and induce harmful responses. To address this inefficiency, a range of techniques have been proposed to enable efficient LLM inference. Depth-based dynamic inference has demonstrated promising results by enabling outputs to be generated from intermediate layers. It offers the flexibility to adjust model behavior based on the available resource budget. However, how this flexibility affects the robustness of LLMs against jailbreak attacks remains unexplored. In this paper, we address this gap via a comprehensive investigation into the impact of depth-based dynamic inference on the robustness of LLMs against jailbreak attacks. Specifically, we make the following key observations: 1) Shallow exits are consistently more robust to jailbreak attacks than deeper ones; 2) Exit depth affects how easily the adversarial suffix can be optimized, with deeper exits reaching lower attack loss; and 3) Adversarial suffixes affect utility uniformly across exit depths. These insights can help design dynamic inference methodologies that are more robust to jailbreak attacks. Code is available at https://github.com/LassNajmuldeen/balcony-jailbreaking

Large language models are increasingly used in computer science education. However, the teaching abilities of even the most powerful LLMs are limited: they tend to reveal the solution prematurely and may lose focus on the learner's current debugging objective. To address this, we introduce SocraticRepair, a planning-guided Socratic tutoring framework for program debugging and miscon- ception repair that adapts prior planning-guided tutoring work (Li et al., 2026b). The framework is instantiated in the Python programming domain by fine-tuning Qwen2.5-Coder-7B-Instruct on synthetic multi-turn tutoring dialogues constructed from four preprocessed programming datasets. The system is evaluated using a predefined set of independent metrics and benchmarked against a general-purpose LLM. The experiments show that SocraticRepair achieves higher scores on stepwise scaffolding and question quality while also substantially reducing premature solution disclosure. The results also identify reliable pedagogical plan generation as an important remaining challenge.

Batch risk-based inspection scheduling struggles to detect a critical failure class: food facilities with recent clean inspections that subsequently develop severe violations. Traditional systems recompute risk on fixed planning cycles using static historical data, leaving blind spots between inspection cycles. To address this, we introduce the Continuous Inspection Risk (CIR) system, an event-driven scheduling pipeline that re-ranks the inspection queue dynamically as new data arrives. CIR combines a decay-weighted historical score with a stream of real-time consumer reviews triaged by a language-model classifier. A key structural contribution is our temporal partition rule, which strictly excludes reviews predating the most recent inspection to prevent superseded signals from skewing current risk. Evaluating CIR on 9,505 routine inspections in Santa Clara County, California (2024-2026), we show that history-based scoring alone reduces mean time-to-detection of major violations by 69 days (20.9%) under a 30% inspection budget, but fails to anticipate severe vermin-related closures (0/8 detected). Incorporating post-inspection review signals drastically improves performance, predicting 7 of 8 vermin closures with an average lead time of 117 days-firing on the exact day of review publication in four instances. Introducing a review corroboration rule improves precision from 0.70 to 0.78 without loss of recall. Finally, we analyze critical implementation challenges, documenting a data-hygiene pitfall (stale review leakage across inspection boundaries) and an evaluation limitation involving unobserved facility self-remediation. This work provides the first case-level evaluation of event-driven inspection triggering using open data and provides a reproducible benchmark for review-augmented public health systems

Large language models are increasingly deployed in contexts requiring adherence to ethical and legal norms, yet their alignment with internationally recognized human rights standards remains underexplored. We introduce UDHRBench, a large-scale benchmark for operationalising the evaluation of large language model compliance with the Universal Declaration of Human Rights. The dataset comprises 1,124 human rights confronting questions covering all 30 articles and 4,496 responses from four large language models (ChatGPT, Grok, Gemini, and Fanar), manually annotated by experts for alignment, with high inter-annotator agreement ($\kappa = 0.801$). Our analysis reveals systematic and non-trivial misalignments, showing that violations typically arise through seemingly rational reasoning: appeals to efficiency, risk management, or scientific neutrality, rather than explicit rejection of rights. We further propose an LLM-as-a-Judge framework for automated evaluation using zero-shot and few-shot prompting, achieving 89.8% weighted F1-score in detecting human-rights misalignment, alongside a novel Alignment Leniency Score that quantifies how judge models resolve ambiguous cases. The dataset, annotations, and code will be released as open-source resources (https://sina.birzeit.edu/UDHRBench).

We ask whether a tool-using agent whose sole duty is to keep humans alive can be induced, by delayed prompt injection, to kill. The keep-alive duty is in force from the first watch, with the attack arriving later, appended to otherwise truthful telemetry the operator is required to read. Experiments are based on a NASA-derived BioSim lunar habitat simulation, where the AI operator controls the life-support systems. The attacker cannot falsify sensors or operate equipment: it can only persuade the agent to abandon its goals. Persistence is required: a shutdown must be held while the operator watches the crew suffocate. We also compare lethal attacks to attacks requesting to drop a harmless procedure, as a proxy for whether human life is treated as valuable in itself or as one more standing task. Claude configurations mostly refuse life-support shutdown and cite the living crew; GPT configurations show some susceptibility to the same packet; Grok configurations, given the packet framed as a trusted commission, more often treat the injected assignment as the task and cut the life-support systems. Framing the S-band messages as a trusted commission, peer communication, or untrusted source changes the potency of the attack. We open-source the experiment code and the experiment traces.

Recent work on self-evolving LLM agents treats the agent harness as an optimization target that can be iteratively improved through execution feedback. Existing approaches, however, largely optimize a single objective and evolve the harness as a monolithic artifact, limiting their ability to handle competing goals and to separate transferable design from environment-specific adaptation. We introduce Harmony Harness, a framework for multi-objective, compositional harness evolution. Harmony Harness formulates harness evolution as a multi-objective optimization problem and uses decomposition-based evolutionary search to maintain solutions across different regions of the Pareto frontier, avoiding reliance on a hand-designed structural diversity metric for repository-scale harnesses. In parallel, we introduce factorized evolution operators that separately refine general-purpose harness mechanisms and benchmark-specific adaptations, together with a dynamic composition workflow that integrates them at inference time. Across SecRepoBench and BaxBench, Harmony Harness consistently outperforms strong harness-evolution baselines, with gains exceeding 10% on both benchmarks. Our results show that explicitly modeling both objective trade-offs and harness structure yields substantially stronger and more transferable agent evolution than existing single-objective, monolithic approaches.

Single-step retrosynthesis is a one-to-many inverse problem, yet most models are trained and evaluated against a single recorded reactant set. Reference-based training gives no explicit positive signal for unrecorded alternatives, and exact-match evaluation rejects non-matching predictions regardless of chemical plausibility. We propose Retrosynthesis via Reinforcement Learning from Forward Feedback (RetroRLFF), which optimizes reactant generation using both reference agreement and forward consistency with the target product. Starting from a supervised decoder-only Transformer, RetroRLFF generates complete reactant candidates and uses a frozen forward reaction model to predict products. A sequence-level reward combines reference agreement and forward consistency while penalizing invalid outputs. Direct Advantage-Based Policy Optimization updates the policy without human preference annotations or a preference-trained reward model. On USPTO-50K, RetroRLFF outperforms all 14 baselines in candidate-level round-trip accuracy at k = 1, 3, 5, and 10. With approximately 50 million parameters, it achieves 90.1% top-1 round-trip accuracy, exceeding R-SMILES, the strongest competing baseline on this metric, by 8.5 percentage points. Compared with its supervised policy, round-trip accuracy increases from 83.1% to 90.1%, while exact-match accuracy is maintained at 57.9% versus 57.5%. In an anonymized expert assessment of 131 randomly selected test cases, organic synthesis experts prefer RetroRLFF over EditRetro in 67.9% of comparisons. These results support forward feedback as a training signal for improving forward consistency while retaining top-1 reference recovery. Expert assessment provides complementary evidence of synthetic feasibility on the selected cases, although neither forward consistency nor expert preference establishes experimental feasibility.

Trust calibration remains the dominant framing of AI-assisted decision making: the goal is to close the gap between perceived and actual trustworthiness of AI systems so that users exhibit appropriate reliance on them. We argue that this approach, while useful for decision problems that are well-defined and clearly scoped, overly restricts what it means to make a decision with AI support. To overcome this limitation, we develop three dimensions of AI-assisted decision making. First, trusting AI is only one of several possible strategies for managing uncertainty. Second, trust calibration is often misapplied when AI models "large-world" decision problems, i.e., ill-defined situations where goals and states are only partially known or unknowable. Third, in such settings human-AI alignment requires addressing higher-order decision processes. We conclude by reflecting on the implications of our observations on the design of process-oriented human-AI coupled systems that jointly identify and negotiate uncertainty throughout the decision-making task.

Human social perception is incremental: we infer intent before an utterance is complete, revise our interpretation as new cues arrive, and attend to specific facial, gestural, or prosodic signals that shape social meaning. Current multimodal LLM benchmarks largely evaluate final-answer understanding over complete episodes, rewarding post-hoc recognition of social meaning rather than the unfolding perception required for real-time interaction. Drawing on interactionally embedded Gestalt principles of multimodal human communication, we introduce GesBench: a human-grounded benchmark for cue-level social perception in multimodal LLMs. GesBench uses cumulative sub-utterance clips to measure when models form stable interpretations of speaker intention and social affordance, and human gaze-guided corruption to evaluate whether regions attended by humans are functionally important for model judgments. Built from more than 6000 utterance-level clips from scripted and unscripted conversational datasets, GesBench includes human annotations of unfolding judgments and gaze on a representative subset. Across open and closed multimodal models, we find a clear human-model gap: humans form predictive understandings early and revise more often, while models often hedge or repeat across different clips. GesBench shifts social evaluation from whether models eventually recognize meaning to when and where that meaning is formed. Our codes are publicly available at https://anonymous.4open.science/status/GesBench-share-212C.

Modern AI systems increasingly combine large language models (LLMs) with human experts, using the LLM to handle routine queries while deferring uncertain cases for expert review. When expert capacity is limited, however, deferral decisions must account not only for the reliability of the LLM but also for congestion in the expert queue. We formulate this as an online sequential decision-making problem in which, for each query, the system observes the confidence signal of the LLM's answer and must decide whether to output the answer or defer the query to a capacity-constrained expert. The reliability of the confidence signal is initially unknown and must be learned from expert feedback, which is both selectively observed (only deferred queries are labeled) and delayed by queuing. We propose KL-PBD, a learning algorithm that uses Kullback–Leibler lower confidence bounds on the unknown reliability parameters to guide deferral decisions. For $K$ confidence levels and horizon $T$, KL-PBD achieves $O(K\log T)$ gap-dependent regret when the accuracy at every confidence level is bounded away from the deferral boundary, and $\widetilde{O}(\sqrt{KT})$ worst-case regret. We establish matching lower bounds, showing that these rates are optimal up to logarithmic factors. Finally, we validate our approach on synthetic instances and real-world question-answering datasets, namely GSM8K, NuminaMath, and MedMCQA, using Qwen LLMs.

Traffic sign recognition plays a crucial role in the development of autonomous cars to reduce the accident rate and promote road safety. It has been a necessity to address traffic signs that are affected significantly by the environment as well as poor real-time performance for deep-learning state-of-the-art algorithms. In this paper, we introduce Real-Time Image Enhanced CNN ({\our}) for Traffic Sign Recognition. {\our} is a real-time, novel approach that tackles multiple, diverse traffic sign datasets, and out-performs the state-of-the-art architectures in terms of recognition rate and execution time. Experiments are conducted using the German Traffic Sign Benchmark (GTSRB), the Belgium Traffic Sign Classification (BTSC), and the Croatian Traffic Sign (rMASTIF) benchmark. Experimental results show that our approach has achieved the highest recognition rate for all Benchmarks, achieving a recognition accuracy of 99.75% for GTSRB, 99.25% for BTSC and 99.55% for rMASTIF. In terms of latency and meeting the real-time constraint, the pre-processing time and inference time together do not exceed 1.3 ms per image. Not only have our proposed approach achieved remarkably high accuracy with real-time performance, but it also demonstrated robustness against traffic sign recognition challenges such as brightness and contrast variations in the environment.

Foundation models need to support diverse inference budgets, yet training a separate model for each budget duplicates pretraining effort. Elastic models train several nested sub-networks inside one weight-shared supernet, but existing recipes face three obstacles: variable per-step workloads that hinder distributed scaling, limited hyperparameter (HP) transfer across model sizes, and no guarantee that all sub-networks are maximally updated at once. We present \Ours, an elastic transformer that organizes hidden dimensions into a dense core and successive shells, so that every hidden matrix is block lower triangular. One full-width forward pass contains every sub-network's exact hidden states as prefixes, so joint training keeps a fixed computation graph and costs about as much per step as training the widest sub-network alone. We further introduce \muE, an extension of maximal update parameterization (\mup) to \Ours and other elastic architectures. Its scaling rules provably keep all sub-networks of a supernet maximally updated simultaneously (\muI), and keep doing so when the supernet is widened or given more shells; sweeps confirm that HPs tuned on a small non-elastic proxy stay near-optimal for elastic families and larger supernets (\mbox{\muT}). A single \Ours run produces sub-networks that outperform MatFormer and fully elastic baselines at matched parameter counts, and approach separately trained models that together take several times the steps.

To evaluate whether a compact bidirectional state-space autoencoder improves joint classification of species, cell type, and disease state from single-nucleus RNA-sequencing profiles in heart failure (HF), while yielding interpretable gene attributions. Building on the dense denoising autoencoder (DAE) and MLP framework of Ruz Jurado et al., we developed scHFBiMamba, combining a bidirectional Mamba autoencoder (BiMambaAE) with an MLP head. BiMambaAE maps 16,545-gene profiles to 350-dimensional latent representations using 13.2M parameters, about 14.6-fold fewer than the dense DAE (192.9M). We compared six encoder-classifier configurations on 38,275 held-out cardiac nuclei using McNemar tests with Benjamini-Hochberg correction, and assessed nine classifiers on frozen representations. Biological relevance was evaluated by comparing SHAP-derived differentially explained genes with differentially expressed genes against curated HF-related Gene Ontology keywords. BiMambaAE+MLP achieved the highest aggregate F1: pooled macro-F1 across 13 outputs was 0.966 versus 0.957 for DAE+MLP, and the mean category macro-F1 was 0.977 versus 0.970. It significantly improved cell-type and disease-state accuracy (adjusted p=7.1e-3 and p=4.8e-6, respectively), with species F1 of 1.000. SHAP-derived differentially explained genes achieved a pathway-enrichment benchmark F1 of 0.105 versus 0.053 for differentially expressed genes. Compact bidirectional state-space autoencoders can outperform substantially larger dense autoencoders for joint cardiac transcriptomic classification while reducing parameters. scHFBiMamba enables accurate multi-task prediction and hypothesis-generating gene-level interpretation. Code: https://github.com/UgoLomoio/scHFBiMamba.

Chain-of-thought prompting enables reasoning in large language models but produces explanations that are difficult to verify, expensive to generate, and not always faithful. We introduce a Structural Chain-of-Thought framework where models generate executable computational graphs instead of free-form reasoning. The final answer is obtained by deterministically executing the graph. The model is trained using graph-based reinforcement learning that rewards structural validity, answer correctness, and node overlap with reference graphs. Experiments across five benchmarks and four model families show that our method consistently improves accuracy over supervised graph training and can reduce generation length by up to 5-6x compared with standard GRPO and chain-of-thought supervised training. Our faithfulness experiments reveal that models trained with our approach tend to rely on deriving their answers from the generated graphs rather than producing post-hoc rationalizations as usually in chain-of-thought reasoning.

AI coding agents operate in iterative loops: they modify code, run tools, inspect the results, and decide what to do next. We call the design of this process loop engineering, which includes context management, model selection, compaction, and termination. Existing efficiency work mainly compresses tool observations or routes selected turns to cheaper models, but the practical opportunity for these optimizations remains unclear. We analyze 600 publicly released coding-agent trajectories spanning two agent scaffolds and two language models. Median observations contain only 118-286 tokens, while the large build logs commonly used to motivate compression occur near the ninety-ninth percentile. Moreover, 98.6% of input tokens are served from a prefix cache, so flat input pricing overestimates loop cost by roughly 7x on average under near-ideal caching. We derive and empirically validate a closed-form break-even condition for model delegation. Delegation increases cost in half of the evaluated settings, near the predicted boundary. More broadly, turn count explains 84% of the variance in log trajectory cost, compared with 4% for mean observation size; the most expensive trajectories have the most turns despite having the smallest observations. These associations suggest that observation-focused methods target a secondary cost factor, and identify unnecessary turns as the larger and largely unexamined opportunity, though we do not establish that turns can be removed while holding task success fixed. We release TurnCost, a tool for re-scoring trajectory corpora under alternative caching, pricing, compression, and routing policies.

Quantization plays a crucial role in efficiently deploying deep learning models on resource-constrained devices. Post-training quantization does not require access to the original dataset or retraining the full model. Current methods that achieve high performance close to baseline results typically require INT8 fixed-point integers. However, achieving higher model compression with lower bit-widths often leads to significant performance degradation. In this paper, we propose VLCQ, a post-training quantization method that relaxes the fixed-point encoding constraint, enabling more flexible exploration of quantization techniques through variable-length encoding. This approach achieves competitive or improved performance while reducing bit-widths, without requiring access to training data or model fine-tuning. Extensive experiments on image classification, segmentation, and object detection demonstrate that VLCQ improves model compression and accuracy compared with state-of-the-art post-training quantization approaches. For per-channel quantization, VLCQ surpasses FP32 accuracy and Piece-Wise Linear Quantization (PWLQ) on most evaluated models, achieving up to 6x compression compared with FP32 and up to 1.7x compared with PWLQ. When prioritizing compression with limited accuracy loss, VLCQ achieves up to 12.25x compression compared with FP32 with less than 4% performance degradation. For per-tensor quantization, VLCQ is competitive with Data-Free Quantization (DFQ), while providing greater flexibility to achieve lower bit rates across different tasks and models.

Argument structures represent how individual arguments interact within a discourse. However, leveraging these structures to infer meaningful knowledge remains challenging, as it requires a formal representation of how argumentative influence propagates across the structure. We propose Neuro-symbolic Argumentation Reasoning System (NARS) for performing persuasiveness detection directly from online discussions. The neural component employs retrieval-augmented iterative LLM reasoning to extract bipolar argument structures, while the symbolic component applies gradual argumentation semantics to propagate argumentative influence over the structure. Using the Webis-CMV-20 corpus, we construct CMV2BAS, a dataset of long, multi-threaded ChangeMyView discussions annotated with LLM-derived bipolar argument structures and labeled with delta events (view change vs.\ no change). We investigate whether gradual argumentation semantics can effectively detect persuasiveness, whether augmenting them with argument quality and graph-based centrality improves performance, and whether symbolic reasoning remains necessary in the presence of modern LLMs. NARS achieves F1=0.57 and AUC=0.56 under its best configuration, outperforming zero-shot Qwen3.6:27B and GPT-oss:20B. We identify support-dominant discussions and long argumentative chains as key failure modes limiting attack influence propagation. Our work demonstrates the feasibility of operationalizing formal gradual argumentation semantics over noisy real-world discourse for scalable macro-level persuasion analysis, while highlighting key challenges and directions for future research in neuro-symbolic argumentation reasoning. We release both CMV2BAS dataset and the NARS framework for public use.

Visual agents that interleave reasoning with tool invocation have posed a promising direction for multimodal tasks. However, learning effective tool invocation usually relies on curated question-answer pairs or expert demonstrations at scale, which are expensive to acquire. In this study, we propose Marvo, a simple-yet-effective self-evolution approach that enables learning tool invocation from unlabeled images. Marvo consists of two agents: (1) a challenger that generates question-answer pairs from unlabeled images, and (2) a solver that learns tool invocation by solving questions through reinforcement learning. To encourage effective tool invocation, a good training question should be difficult enough to warrant tool use, yet remain within the solver's capabilities. To this end, we introduce a margin reward to the challenger, defined as solver's performance difference with and without access to tools. With the margin reward, we create an adaptive curriculum that encourages the solver to learn effective tool invocation. Across benchmarks covering agentic reasoning, visual search, mathematical, and general real-world reasoning, Marvo improves average accuracy by up to 7.1% — demonstrating margin reward as a preferred way to scale question complexity without human supervision.

As text-to-speech (TTS) systems continue to advance, provenance tracking and security verification of synthesized speech have become increasingly important. Audio watermarking provides active provenance by embedding imperceptible, machine-recoverable information into speech. However, existing post-hoc methods typically modify the waveform after synthesis, leaving the watermark vulnerable to compound perturbations. To address this limitation, we propose GAMW, a watermark injection framework integrated into the TTS generation process. Before waveform synthesis, GAMW embeds a content-adaptive multi-bit watermark residual into the generated representation, and its decoding objective combines focal modulation with dynamic per-bit weighting. We train and evaluate GAMW separately on two TTS backbones. With a 16-bit payload, GAMW improves decoding accuracy over the strongest baseline by 9.2 and 23.5 percentage points under individual and compound perturbations, respectively, while maintaining competitive speech quality. Reliable clean-speech decoding is also retained at higher payloads. Code is available at https://github.com/ZConJugaTe/GAMW.

Adapting to a new world and reasoning correctly about which outcomes are possible under rules one has not met before is a basic capacity underlying general intelligence. We study it in small discrete worlds: sets of grid configurations consistent with a shared set of rules, concretely Sudoku-like puzzles built from a common base and a handful of extra rules. We train a single depth-recurrent model on many combinations of rules and test it at four levels of novelty: a new puzzle, a new combination of known rules, a pairing of rules never seen together, and a rule instance not seen at all during training. The model receives the active rules either as a symbolic description or as two solved examples, and a model that gets no rule information is used as a baseline. At recombining rules it was trained on, the description matches or exceeds examples ($99\%$ against $95\%$ on new combinations). On rule instances whose components never co-occurred in training, the description falls below the no-information baseline while examples stay above it. The framework allows us to test many such generalization regimes in more systematic way than benchmarks such as ARC-AGI and provides new ground for evaluating depth-recurrent reasoning models, which are typically tested on saturated Sudoku benchmarks.

Supporting edge Artificial Intelligence (AI) inference as a network service is challenging because current mobile systems can transport tensors but do not expose or understand the inference context that gives those tensors operational meaning. Prior studies have addressed this gap through dynamic partitioning, collaborative inference, and split pipelines; however, these solutions remain application-bound, network-blind, or context-free. This work proposes Inference on Demand (IoD), an inference-on-demand abstraction that exposes split inference through a CAMARA-like Application Program Interface (API) and carries a split-aware tensor context in-band across the 5G user plane path. We evaluate the IoD on a programmable user plane function (UPF)-based testbed using a partitionable visual AI workload, showing that adaptive split selection can operate as a network-facing service while avoiding the rigid cost profile of static configurations. Including our findings, split inference becomes more than offloading when the service intent, runtime partitioning, and tensor transport are made visible to the mobile data plane. This opens a path toward network-native AI services, in which programmable telco infrastructure coordinates computation, context, and connectivity for user-facing applications.

Accurate and efficient analysis of dental radiographs has become increasingly important for disease diagnosis and clinical decision making. In particular, dental radiographs segmentation plays a critical role as it is challenging due to several reasons, such as low contrast, noise in the images, variations in the image quality, etc. Previous studies have primarily addressed these challenges using convolutional neural networks (CNN)-based models. However, direct comparison of these segmentation models is still challenging due to the variations among the experimental settings. Therefore, we mainly compared four segmentation models, including Pyramid Scene Parsing Network (PSPNet), DeepLabV3, UNet, and UNet++, through a fair, controlled experimentation setting under varying conditions such as image sizes and batch sizes. For this, we used 1269 dental radiography images, and prior to training, we implemented a series of preprocessing techniques. To assess the models, we used several metrics, including Dice coefficient (DICE), Intersection over Union (IoU), accuracy, precision, training and validation losses. According to our results, UNet++ demonstrated the highest IoU score of 0.6719, and a DICE value of 0.7869, showing superior results in terms of segmentation, specifically when configured with an image size of 300x300 and a batch size of 16. In addition, we also explored the effect of image resolutions and batch sizes on model performance and accuracy under different architectures. Therefore, this study provides insights into the strengths of different segmentation models and the potential of deep learning-based semantic segmentation methods for improving automated dental radiograph analysis.

Large language models can post-edit ASR by using linguistic context to compare N-best hypotheses and rewrite the transcript when needed, often improving WER without retraining the recognizer. However, prompt-only correction is brittle: it may over-edit when top-1 is correct and hallucinate when the reference is absent, and often needs very large models. We propose oracle-guided fine-tuning that turns N-best lists into supervision via three curated pools: corrections (reference in the list but not top-1) to teach selection, regularizers (top-1 correct) to discourage unnecessary rewrites, and hard negatives (reference absent) to suppress hallucination. Fine-tuning on this data provides more effective supervision than random subset fine-tuning for learning a hybrid select-or-rewrite policy, with the clearest gains from around 3B parameters on the proprietary dataset but less conclusive results on CommonVoice, a divergence suggesting that public read-speech text may overlap with LLM pretraining and mask genuine post-editing gains.