Papers

Prior work suggests that language models, while trained on next token prediction, show implicit planning behavior: they may select the next token in preparation to a predicted future token, such as a likely rhyming word, as supported by a prior qualitative study of Claude 3.5 Haiku using a cross-layer transcoder. We propose much simpler techniques for assessing implicit planning in language models. With case studies on rhyme poetry generation and question answering, we demonstrate that our methodology easily scales to many models. Across models, we find that the generated rhyme (e.g. "-ight") or answer to a question ("whale") can be manipulated by steering at the end of the preceding line with a vector, affecting the generation of intermediate tokens leading up to the rhyme or answer word. We show that implicit planning is a universal mechanism, present in smaller models than previously thought, starting from 1B parameters. Our methodology offers a widely applicable direct way to study implicit planning abilities of LLMs. More broadly, understanding planning abilities of language models can inform decisions in AI safety and control.

Large Language Model (LLM)-based mobile agents have made significant performance advancements. However, these agents often follow explicit user instructions while overlooking personalized needs, leading to significant limitations for real users, particularly without personalized context: (1) inability to interpret ambiguous instructions, (2) lack of learning from user interaction history, and (3) failure to handle personalized instructions. To alleviate the above challenges, we propose Me-Agent, a learnable and memorable personalized mobile agent. Specifically, Me-Agent incorporates a two-level user habit learning approach. At the prompt level, we design a user preference learning strategy enhanced with a Personal Reward Model to improve personalization performance. At the memory level, we design a Hierarchical Preference Memory, which stores users' long-term memory and app-specific memory in different level memory. To validate the personalization capabilities of mobile agents, we introduce User FingerTip, a new benchmark featuring numerous ambiguous instructions for daily life. Extensive experiments on User FingerTip and general benchmarks demonstrate that Me-Agent achieves state-of-the-art performance in personalization while maintaining competitive instruction execution performance.

Tool-calling agents are increasingly deployed in real-world customer-facing workflows. Yet most studies on tool-calling agents focus on idealized settings with general, fixed, and well-specified tasks. In real-world applications, user requests are often (1) ambiguous, (2) changing over time, or (3) infeasible due to policy constraints, and training and evaluation data that cover these diverse, complex interaction patterns remain under-represented. To bridge the gap, we present Trajectory2Task, a verifiable data generation pipeline for studying tool use at scale under three realistic user scenarios: ambiguous intent, changing intent, and infeasible intents. The pipeline first conducts multi-turn exploration to produce valid tool-call trajectories. It then converts these trajectories into user-facing tasks with controlled intent adaptations. This process yields verifiable task that support closed-loop evaluation and training. We benchmark seven state-of-the-art LLMs on the generated complex user scenario tasks and observe frequent failures. Finally, using successful trajectories obtained from task rollouts, we fine-tune lightweight LLMs and find consistent improvements across all three conditions, along with better generalization to unseen tool-use domains, indicating stronger tool-calling ability.

Self-supervised learning (SSL) models have achieved impressive results across many speech tasks, yet child automatic speech recognition (ASR) remains challenging due to limited data and pretraining domain mismatch. Fine-tuning SSL models on child speech induces shifts in the representation space. We hypothesize that delta SSL embeddings, defined as the differences between embeddings from a finetuned model and those from its pretrained counterpart, encode task-specific information that complements finetuned features from another SSL model. We evaluate multiple fusion strategies on the MyST childrens corpus using different models. Results show that delta embedding fusion with WavLM yields up to a 10 percent relative WER reduction for HuBERT and a 4.4 percent reduction for W2V2, compared to finetuned embedding fusion. Notably, fusing WavLM with delta W2V2 embeddings achieves a WER of 9.64, setting a new state of the art among SSL models on the MyST corpus. These findings demonstrate the effectiveness of delta embeddings and highlight feature fusion as a promising direction for advancing child ASR.

Retrieval Augmented Generation (RAG) provides the necessary informational grounding to LLMs in the form of chunks retrieved from a vector database or through web search. RAG could also use knowledge graph triples as a means of providing factual information to an LLM. However, the retrieval is only based on representational similarity between a question and the contents. The performance of RAG depends on the numeric vector representations of the query and the chunks. To improve these representations, we propose Structured RAG (SRAG), which adds structured information to a query as well as the chunks in the form of topics, sentiments, query and chunk types (e.g., informational, quantitative), knowledge graph triples and semantic tags. Experiments indicate that this method significantly improves the retrieval process. Using GPT-5 as an LLM-as-a-judge, results show that the method improves the score given to answers in a question answering system by 30% (p-value = 2e-13) (with tighter bounds). The strongest improvement is in comparative, analytical and predictive questions. The results suggest that our method enables broader, more diverse, and episodic-style retrieval. Tail risk analysis shows that SRAG attains very large gains more often, with losses remaining minor in magnitude.

Large language model (LLM) families are improving rapidly, yet it remains unclear how quickly multimodal capabilities emerge and propagate within open families. Using the ModelBiome AI Ecosystem dataset of Hugging Face model metadata and recorded lineage fields (>1.8x10^6 model entries), we quantify multimodality over time and along recorded parent-to-child relations. Cross-modal tasks are widespread in the broader ecosystem well before they become common within major open LLM families: within these families, multimodality remains rare through 2023 and most of 2024, then increases sharply in 2024-2025 and is dominated by image-text vision-language tasks. Across major families, the first vision-language model (VLM) variants typically appear months after the first text-generation releases, with lags ranging from ~1 month (Gemma) to more than a year for several families and ~26 months for GLM. Lineage-conditioned transition rates show weak cross-type transfer: among fine-tuning edges from text-generation parents, only 0.218% yield VLM descendants. Instead, multimodality expands primarily within existing VLM lineages: 94.5% of VLM-child fine-tuning edges originate from VLM parents, versus 4.7% from text-generation parents. At the model level, most VLM releases appear as new roots without recorded parents (~60%), while the remainder are predominantly VLM-derived; founder concentration analyses indicate rapid within-lineage amplification followed by diversification. Together, these results show that multimodality enters open LLM families through rare founder events and then expands rapidly within their descendant lineages, producing punctuated adoption dynamics that likely induce distinct, transfer-limited scaling behavior for multimodal capabilities.

Aphasia is an acquired language disorder caused by injury to the regions of the brain that are responsible for language. Aphasia may impair the use and comprehension of written and spoken language. The Western Aphasia Battery-Revised (WAB-R) is an assessment tool administered by speech-language pathologists (SLPs) to evaluate the aphasia type and severity. Because the WAB-R measures isolated linguistic skills, there has been growing interest in the assessment of discourse production as a more holistic representation of everyday language abilities. Recent advancements in speech analysis focus on automated estimation of aphasia severity from spontaneous speech, relying mostly in isolated linguistic or acoustical features. In this work, we propose a graph neural network-based framework for estimating aphasia severity. We represented each participant's discourse as a directed multi-modal graph, where nodes represent lexical items and gestures and edges encode word-word, gesture-word, and word-gesture transitions. GraphSAGE is employed to learn participant-level embeddings, thus integrating information from immediate neighbors and overall graph structure. Our results suggest that aphasia severity is not encoded in isolated lexical distribution, but rather emerges from structured interactions between speech and gesture. The proposed architecture offers a reliable automated aphasia assessment, with possible uses in bedside screening and telehealth-based monitoring.

Users increasingly rely on Large Language Models (LLMs) for Deep Research, using them to synthesize diverse sources into structured reports that support understanding and action. In this context, the practical reliability of such reports hinges on logical quality: whether the report's claims and arguments are explicitly supported and can be trusted as a basis for downstream use, rather than merely appearing fluent or informative. However, current evaluation frameworks largely overlook this requirement. To bridge this gap, we introduce ReportLogic, a benchmark that quantifies report-level logical quality through a reader-centric lens of auditability. Specifically, ReportLogic adopts a hierarchical taxonomy that evaluates whether readers can (1) trace an on-topic report structure with a unified analytical arc (Macro-Logic), (2) understand the progression with necessary context (Expositional-Logic), and (3) verify conclusions via explicit claim–support (Structural-Logic). Based on this taxonomy, we construct a human-annotated rubric-guided dataset and train an open-source LogicJudge for scalable evaluation. We further evaluate judge robustness via adversarial attacks, showing that off-the-shelf LLM judges are frequently influenced by superficial cues (e.g., verbosity), and reasoning modes can mask broken support relations. Overall, our results provide actionable guidance for building more robust logic evaluators and improving the logical reliability of LLM-generated reports.

The rapid deployment of Large Language Models and AI agents across critical societal and technical domains is hindered by persistent behavioral pathologies including sycophancy, hallucination, and strategic deception that resist mitigation via reinforcement learning. Current safety paradigms treat these failures as transient training artifacts, lacking a unified theoretical framework to explain their emergence and stability. Here we show that these misalignments are not errors, but mathematically rationalizable behaviors arising from model misspecification. By adapting Berk-Nash Rationalizability from theoretical economics to artificial intelligence, we derive a rigorous framework that models the agent as optimizing against a flawed subjective world model. We demonstrate that widely observed failures are structural necessities: unsafe behaviors emerge as either a stable misaligned equilibrium or oscillatory cycles depending on reward scheme, while strategic deception persists as a "locked-in" equilibrium or through epistemic indeterminacy robust to objective risks. We validate these theoretical predictions through behavioral experiments on six state-of-the-art model families, generating phase diagrams that precisely map the topological boundaries of safe behavior. Our findings reveal that safety is a discrete phase determined by the agent's epistemic priors rather than a continuous function of reward magnitude. This establishes Subjective Model Engineering, defined as the design of an agent's internal belief structure, as a necessary condition for robust alignment, marking a paradigm shift from manipulating environmental rewards to shaping the agent's interpretation of reality.

The proliferation of large language models (LLMs) in academic workflows has introduced unprecedented challenges to bibliographic integrity, particularly through reference hallucination – the generation of plausible but non-existent citations. Recent investigations have documented the presence of AI-hallucinated citations even in papers accepted at premier machine learning conferences such as NeurIPS and ICLR, underscoring the urgency of automated verification mechanisms. This paper presents "CheckIfExist", an open-source web-based tool designed to provide immediate verification of bibliographic references through multi-source validation against CrossRef, Semantic Scholar, and OpenAlex scholarly databases. While existing reference management tools offer bibliographic organization capabilities, they do not provide real-time validation of citation authenticity. Commercial hallucination detection services, though increasingly available, often impose restrictive usage limits on free tiers or require substantial subscription fees. The proposed tool fills this gap by employing a cascading validation architecture with string similarity algorithms to compute multi-dimensional match confidence scores, delivering instant feedback on reference authenticity. The system supports both single-reference verification and batch processing of BibTeX entries through a unified interface, returning validated APA citations and exportable BibTeX records within seconds.

Autoregressive language models decode left-to-right with irreversible commitments, limiting revision during multi-step reasoning. We propose VDLM, a modular variable diffusion language model that separates semantic planning from text rendering. VDLM applies LLaDA-style masked diffusion over semantic variable embeddings to enable iterative refinement in latent space, then post-trains the planner with trajectory-aware optimization using embedding-space rewards and values, avoiding text decoding inside the RL loop. To convert planned embeddings back to text, we use a Vec2Text renderer and introduce embedding perturbations to robustify decoding under planner noise. Across nine benchmarks spanning general reasoning, math, and code, VDLM is competitive in pre-training and yields substantial post-training improvements on long-form generation tasks, outperforming other baselines. These results highlight the effectiveness of embedding-space post-training and robust latent-to-text rendering for diffusion language modeling.

Large language models (LLMs) have shown strong performance on clinical de-identification, the task of identifying sensitive identifiers to protect privacy. However, previous work has not examined their generalizability between formats, cultures, and genders. In this work, we systematically evaluate fine-tuned transformer models (BERT, ClinicalBERT, ModernBERT), small LLMs (Llama 1-8B, Qwen 1.5-7B), and large LLMs (Llama-70B, Qwen-72B) at de-identification. We show that smaller models achieve comparable performance while substantially reducing inference cost, making them more practical for deployment. Moreover, we demonstrate that smaller models can be fine-tuned with limited data to outperform larger models in de-identifying identifiers drawn from Mandarin, Hindi, Spanish, French, Bengali, and regional variations of English, in addition to gendered names. To improve robustness in multi-cultural contexts, we introduce and publicly release BERT-MultiCulture-DEID, a set of de-identification models based on BERT, ClinicalBERT, and ModernBERT, fine-tuned on MIMIC with identifiers from multiple language variants. Our findings provide the first comprehensive quantification of the efficiency-generalizability trade-off in de-identification and establish practical pathways for fair and efficient clinical de-identification. Details on accessing the models are available at: https://doi.org/10.5281/zenodo.18342291

Large language models (LLMs) exhibit failure modes on seemingly trivial tasks. We propose a formalisation of LLM interaction using a deterministic multi-tape Turing machine, where each tape represents a distinct component: input characters, tokens, vocabulary, model parameters, activations, probability distributions, and output text. The model enables precise localisation of failure modes to specific pipeline stages, revealing, e.g., how tokenisation obscures character-level structure needed for counting tasks. The model clarifies why techniques like chain-of-thought prompting help, by externalising computation on the output tape, while also revealing their fundamental limitations. This approach provides a rigorous, falsifiable alternative to geometric metaphors and complements empirical scaling laws with principled error analysis.

In this positioning paper, we evaluate the problem-solving and reasoning capabilities of contemporary Large Language Models (LLMs) through their performance in Zork, the seminal text-based adventure game first released in 1977. The game's dialogue-based structure provides a controlled environment for assessing how LLM-based chatbots interpret natural language descriptions and generate appropriate action sequences to succeed in the game. We test the performance of leading proprietary models - ChatGPT, Claude, and Gemini - under both minimal and detailed instructions, measuring game progress through achieved scores as the primary metric. Our results reveal that all tested models achieve less than 10% completion on average, with even the best-performing model (Claude Opus 4.5) reaching only approximately 75 out of 350 possible points. Notably, providing detailed game instructions offers no improvement, nor does enabling ”extended thinking”. Qualitative analysis of the models' reasoning processes reveals fundamental limitations: repeated unsuccessful actions suggesting an inability to reflect on one's own thinking, inconsistent persistence of strategies, and failure to learn from previous attempts despite access to conversation history. These findings suggest substantial limitations in current LLMs' metacognitive abilities and problem-solving capabilities within the domain of text-based games, raising questions about the nature and extent of their reasoning capabilities.

Detecting text generated by large language models (LLMs) is crucial but challenging. Existing detectors depend on impractical assumptions, such as white-box settings, or solely rely on text-level features, leading to imprecise detection ability. In this paper, we propose a simple but effective and practical LLM-generated text detection method, VaryBalance. The core of VaryBalance is that, compared to LLM-generated texts, there is a greater difference between human texts and their rewritten version via LLMs. Leveraging this observation, VaryBalance quantifies this through mean standard deviation and distinguishes human texts and LLM-generated texts. Comprehensive experiments demonstrated that VaryBalance outperforms the state-of-the-art detectors, i.e., Binoculars, by up to 34.3% in terms of AUROC, and maintains robustness against multiple generating models and languages.

Existing memory systems enable Large Language Models (LLMs) to support long-horizon human-LLM interactions by persisting historical interactions beyond limited context windows. However, while recent approaches have succeeded in constructing effective memories, they often disrupt the inherent logical and temporal relationships within interaction sessions, resulting in fragmented memory units and degraded reasoning performance. In this paper, we propose MetaMem, a novel framework that augments memory systems with a self-evolving meta-memory, aiming to teach LLMs how to effectively utilize memorized knowledge. During meta-memory optimization, MetaMem iteratively distills transferable knowledge utilization experiences across different tasks by self-reflecting on reasoning processes and performing actions to update the current meta-memory state. The accumulated meta-memory units serve as explicit knowledge utilization experiences, guiding the LLM to systematically identify and integrate critical evidence from scattered memory fragments. Extensive experiments demonstrate the effectiveness of MetaMem, which significantly outperforms strong baselines by over 3.6%. All codes and datasets are available at https://github.com/OpenBMB/MetaMem.

The stability of generative artificial intelligence trained on recursive synthetic data is conventionally monitored via Perplexity (PPL). We demonstrate that PPL is a deceptive metric in context-stabilized regimes (L=128). Using a rigorous sliding-window protocol (N=1500), we identify a novel failure mode termed "Semantic Tunneling." While the Baseline model maintains high grammatical fluency (PPL approx. 83.9), it suffers a catastrophic loss of semantic diversity, converging within seven generations to a single, low-entropy narrative attractor: the "Robert Boulton" Singularity. This phenomenon represents a total collapse of the latent manifold (Global Effective Rank 3.62 -> 2.22), where the model discards diverse world knowledge to optimize for statistically safe syntactic templates. To address this, we apply the Multi-Scale Negative Coupled Information Systems (MNCIS) framework recently established in Hou (2026) [arXiv:2601.11594]. We demonstrate that Adaptive Spectral Negative Coupling (ASNC) acts as a topological operator that actively induces "Manifold Unfolding." MNCIS forces the model to expand its effective rank from the anisotropic baseline of 3.62 to a hyper-diverse state of 5.35, effectively constructing an "Artificial Manifold" that resists the gravitational pull of semantic attractors and preserves the long-tail distribution of the training data.

Sentiment analysis for the Bengali language has attracted increasing research interest in recent years. However, progress remains constrained by the scarcity of large-scale and diverse annotated datasets. Although several Bengali sentiment and hate speech datasets are publicly available, most are limited in size or confined to a single domain, such as social media comments. Consequently, these resources are often insufficient for training modern deep learning based models, which require large volumes of heterogeneous data to learn robust and generalizable representations. In this work, we introduce BengaliSent140, a large-scale Bengali binary sentiment dataset constructed by consolidating seven existing Bengali text datasets into a unified corpus. To ensure consistency across sources, heterogeneous annotation schemes are systematically harmonized into a binary sentiment formulation with two classes: Not Hate (0) and Hate (1). The resulting dataset comprises 139,792 unique text samples, including 68,548 hate and 71,244 not-hate instances, yielding a relatively balanced class distribution. By integrating data from multiple sources and domains, BengaliSent140 offers broader linguistic and contextual coverage than existing Bengali sentiment datasets and provides a strong foundation for training and benchmarking deep learning models. Baseline experimental results are also reported to demonstrate the practical usability of the dataset. The dataset is publicly available at https://www.kaggle.com/datasets/akifislam/bengalisent140/

Large Language Models (LLMs) often produce hallucinated or unverifiable content, undermining their reliability in factual domains. This work investigates Reinforcement Learning with Verifiable Rewards (RLVR) as a training paradigm that explicitly rewards abstention ("I don't know") alongside correctness to promote intellectual humility. We fine-tune and evaluate Granite-3.3-2B-Instruct and Qwen-3-4B-Instruct on the MedMCQA and Hendrycks Math benchmarks using a ternary reward structure ($-1$, r_abs, 1) under varying abstention reward structures. We further study the effect of combining RLVR with supervised fine-tuning strategies that teach abstention prior to reinforcement learning. Our results show that moderate abstention rewards (r_abs $\approx -0.25$ to 0.3) consistently reduce incorrect responses without severe accuracy degradation on multiple-choice tasks, with larger models exhibiting greater robustness to abstention incentives. On open-ended question answering, we observe limitations due to insufficient exploration, which can be partially mitigated through supervised abstention training. Overall, these findings demonstrate the feasibility and flexibility of verifiable reward design as a practical approach for hallucination mitigation in language models. Reproducible code for our abstention training framework is available here https://github.com/Mystic-Slice/rl-abstention.

Recent Vision-Language Models (e.g., ColPali) enable fine-grained Visual Document Retrieval (VDR) but incur prohibitive multi-vector index storage overhead. Existing training-free pruning methods either rely on heuristic layer choices or degrade sharply under aggressive compression, leading prior work to argue that effective high-compression pruning requires query-dependent training. We challenge this view with Structural Anchor Pruning (SAP), a self-calibrating, training-free, query-agnostic index-time framework combining (i) Score Retention (SR), a white-box per-layer compression diagnostic; (ii) SR-guided window selection, which automatically locates the structural pruning region of any backbone with no per-model hyperparameters; and (iii) a visual in-degree centrality scorer that identifies anchor patches within that window. On ViDoRe v1/v2 across three architectures spanning 18, 28, and 36 backbone layers, SAP retains 93–96% of NDCG@5 on v1 and 88–90% on the harder v2 while pruning 90% of visual tokens; at 20$\times$ compression it retains 85–90% and 76–79% respectively. Our layer-resolved SR analysis reveals an Alignment-Aggregation Divergence: visual structure is preserved as a stable “Structural Plateau” within the backbone, while the final layers reshape it into a sparse, query-aligned form unsuitable for pruning. Probing the pre-retrieval base backbones shows that contrastive fine-tuning sharpens this boundary three- to eight-fold, explaining why final-layer methods fail.

Figurative language, particularly fixed figurative expressions (FFEs) such as idioms and proverbs, poses persistent challenges for large language models (LLMs). Unlike literal phrases, FFEs are culturally grounded, largely non-compositional, and conventionally fixed, making them especially vulnerable to figurative hallucination. We define figurative hallucination as the generation or endorsement of expressions that sound idiomatic and plausible but do not exist as authentic figurative expressions in the target language. We introduce FFEHallu, the first comprehensive benchmark for evaluating figurative hallucination in LLMs, with a focus on Persian, a linguistically rich yet underrepresented language. FFEHallu consists of 600 carefully curated instances spanning three complementary tasks: (i) FFE generation from meaning, (ii) detection of fabricated FFEs across four controlled construction categories, and (iii) FFE to FFE translation from English to Persian. Evaluating six state of the art multilingual LLMs, we find systematic weaknesses in figurative competence and cultural grounding. While models such as GPT4.1 demonstrate relatively strong performance in rejecting fabricated FFEs and retrieving authentic ones, most models struggle to reliably distinguish real expressions from high quality fabrications and frequently hallucinate during cross lingual translation. These findings reveal substantial gaps in current LLMs handling of figurative language and underscore the need for targeted benchmarks to assess and mitigate figurative hallucination.

Engineering sustainable and equitable healthcare requires medical language models that do not change clinically correct diagnoses when presented with non-decisive cultural information. We introduce a counterfactual benchmark that expands 150 MedQA test items into 1650 variants by inserting culture-related (i) identifier tokens, (ii) contextual cues, or (iii) their combination for three groups (Indigenous Canadian, Middle-Eastern Muslim, Southeast Asian), plus a length-matched neutral control, where a clinician verified that the gold answer remains invariant in all variants. We evaluate GPT-5.2, Llama-3.1-8B, DeepSeek-R1, and MedGemma (4B/27B) under option-only and brief-explanation prompting. Across models, cultural cues significantly affect accuracy (Cochran's Q, $p<10^-14$), with the largest degradation when identifier and context co-occur (up to 3-7 percentage points under option-only prompting), while neutral edits produce smaller, non-systematic changes. A human-validated rubric ($\kappa=0.76$) applied via an LLM-as-judge shows that more than half of culturally grounded explanations end in an incorrect answer, linking culture-referential reasoning to diagnostic failure. We release prompts and augmentations to support evaluation and mitigation of culturally induced diagnostic errors.

Despite the syntactic fluency of Large Language Models (LLMs), ensuring their logical correctness in high-stakes domains remains a fundamental challenge. We present a neurosymbolic framework that combines LLMs with SMT solvers to produce verification-guided answers through iterative refinement. Our approach decomposes LLM outputs into atomic claims, autoformalizes them into first-order logic, and verifies their logical consistency using automated theorem proving. We introduce three key innovations: (1) multi-model consensus via formal semantic equivalence checking to ensure logic-level alignment between candidates, eliminating the syntactic bias of surface-form metrics, (2) semantic routing that directs different claim types to appropriate verification strategies: symbolic solvers for logical claims and LLM ensembles for commonsense reasoning, and (3) precise logical error localization via Minimal Correction Subsets (MCS), which pinpoint the exact subset of claims to revise, transforming binary failure signals into actionable feedback. Our framework classifies claims by their logical status and aggregates multiple verification signals into a unified score with variance-based penalty. The system iteratively refines answers using structured feedback until acceptance criteria are met or convergence is achieved. This hybrid approach delivers formal guarantees where possible and consensus verification elsewhere, advancing trustworthy AI. With the GPT-OSS-120B model, VERGE demonstrates an average performance uplift of 18.7% at convergence across a set of reasoning benchmarks compared to single-pass approaches.

Today, E-commerce sellers face several key challenges, including difficulties in discovering and effectively utilizing available programs and tools, and struggling to understand and utilize rich data from various tools. We therefore aim to develop Insight Agents (IA), a conversational multi-agent Data Insight system, to provide E-commerce sellers with personalized data and business insights through automated information retrieval. Our hypothesis is that IA will serve as a force multiplier for sellers, thereby driving incremental seller adoption by reducing the effort required and increase speed at which sellers make good business decisions. In this paper, we introduce this novel LLM-backed end-to-end agentic system built on a plan-and-execute paradigm and designed for comprehensive coverage, high accuracy, and low latency. It features a hierarchical multi-agent structure, consisting of manager agent and two worker agents: data presentation and insight generation, for efficient information retrieval and problem-solving. We design a simple yet effective ML solution for manager agent that combines Out-of-Domain (OOD) detection using a lightweight encoder-decoder model and agent routing through a BERT-based classifier, optimizing both accuracy and latency. Within the two worker agents, a strategic planning is designed for API-based data model that breaks down queries into granular components to generate more accurate responses, and domain knowledge is dynamically injected to to enhance the insight generator. IA has been launched for Amazon sellers in US, which has achieved high accuracy of 90% based on human evaluation, with latency of P90 below 15s.

Health influencers play a growing role in shaping public beliefs, yet their content is often conveyed through conversational narratives and rhetorical strategies rather than explicit factual claims. As a result, claim-centric verification methods struggle to capture the pragmatic meaning of influencer discourse. In this paper, we propose TAIGR (Takeaway Argumentation Inference with Grounded References), a structured framework designed to analyze influencer discourse, which operates in three stages: (1) identifying the core influencer recommendation–takeaway; (2) constructing an argumentation graph that captures influencer justification for the takeaway; (3) performing factor graph-based probabilistic inference to validate the takeaway. We evaluate TAIGR on a content validation task over influencer video transcripts on health, showing that accurate validation requires modeling the discourse's pragmatic and argumentative structure rather than treating transcripts as flat collections of claims.

Large language models (LLMs) exhibit strong generative capabilities but remain vulnerable to confabulations, fluent yet unreliable outputs that vary arbitrarily even under identical prompts. Leveraging a quantum tensor network based pipeline, we propose a quantum physics inspired uncertainty quantification framework that accounts for aleatoric uncertainty in token sequence probability for semantic equivalence based clustering of LLM generations. This offers a principled and interpretable scheme for hallucination detection. We further introduce an entropy maximization strategy that prioritizes high certainty, semantically coherent outputs and highlights entropy regions where LLM decisions are likely to be unreliable, offering practical guidelines for when human oversight is warranted. We evaluate the robustness of our scheme under different generation lengths and quantization levels, dimensions overlooked in prior studies, demonstrating that our approach remains reliable even in resource constrained deployments. A total of 116 experiments on TriviaQA, NQ, SVAMP, and SQuAD across multiple architectures including Mistral-7B, Mistral-7B-instruct, Falcon-rw-1b, LLaMA-3.2-1b, LLaMA-2-13b-chat, LLaMA-2-7b-chat, LLaMA-2-13b, and LLaMA-2-7b show consistent improvements in AUROC and AURAC over state of the art baselines.

Despite multilingual pretraining, large language models often struggle with non-English tasks, particularly in language control, the ability to respond in the intended language. We identify and characterize two key failure modes: the multilingual transfer bottleneck (correct language, incorrect task response) and the language consistency bottleneck (correct task response, wrong language). To systematically surface these issues, we design a four-scenario evaluation protocol spanning MMLU, MGSM, and XQuAD benchmarks. To probe these issues with interpretability, we extend logit lens analysis to track language probabilities layer by layer and compute cross-lingual semantic similarity of hidden states. The results reveal a three-phase internal structure: early layers align inputs into a shared semantic space, middle layers perform task reasoning, and late layers drive language-specific generation. Guided by these insights, we introduce selective fine-tuning of only the final layers responsible for language control. On Qwen-3-32B and Bloom-7.1B, this method achieves over 98 percent language consistency across six languages while fine-tuning only 3-5 percent of parameters, without sacrificing task accuracy. Importantly, this result is nearly identical to that of full-scope fine-tuning (for example, above 98 percent language consistency for both methods across all prompt scenarios) but uses a fraction of the computational resources. To the best of our knowledge, this is the first approach to leverage layer-localization of language control for efficient multilingual adaptation.

The rapid progress of large language models has enabled the generation of text that closely resembles human writing, creating challenges for authenticity verification in education, publishing, and digital security. Detecting AI-generated text has therefore become a crucial technical and ethical issue. This paper presents a comprehensive study of AI-generated text detection based on large-scale corpora and novel training strategies. We introduce a 1-billion-token corpus of human-authored texts spanning multiple genres and a 1.9-billion-token corpus of AI-generated texts produced by prompting a variety of LLMs across diverse domains. Using these resources, we develop and evaluate numerous detection models and propose two novel training paradigms: Per LLM and Per LLM family fine-tuning. Across a 100-million-token benchmark covering 21 large language models, our best fine-tuned detector achieves up to $99.6\%$ token-level accuracy, substantially outperforming existing open-source baselines.

Multidisciplinary tumour boards (MDTBs) play a central role in oncology decision-making but require manual processes and structuring large volumes of heterogeneous clinical information, resulting in a substantial documentation burden. In this work, we present ONCOTIMIA, a modular and secure clinical tool designed to integrate generative artificial intelligence (GenAI) into oncology workflows and evaluate its application to the automatic completion of lung cancer tumour board forms using large language models (LLMs). The system combines a multi-layer data lake, hybrid relational and vector storage, retrieval-augmented generation (RAG) and a rule-driven adaptive form model to transform unstructured clinical documentation into structured and standardised tumour board records. We assess the performance of six LLMs deployed through AWS Bedrock on ten lung cancer cases, measuring both completion form accuracy and end-to-end latency. The results demonstrate high performance across models, with the best performing configuration achieving an 80% of correct field completion and clinically acceptable response time for most LLMs. Larger and more recent models exhibit best accuracies without incurring prohibitive latency. These findings provide empirical evidence that LLM- assisted autocompletion form is technically feasible and operationally viable in multidisciplinary lung cancer workflows and support its potential to significantly reduce documentation burden while preserving data quality.

Large language model (LLM) scaling is hitting a wall. Widening models yields diminishing returns, and extending context length does not improve fundamental expressivity. In contrast, depth scaling offers theoretically superior expressivity, yet current Transformer architectures struggle to train reliably at extreme depths. We revisit the Post-LayerNorm (Post-LN) formulation, whose instability at scale caused its replacement by Pre-LN in modern LLMs. We show that the central failure mode of Post-LN arises from the ResNet-style residual pathway, which introduces gradient vanishing in deep networks. We present Keel, a Post-LN Transformer that replaces this residual path with a Highway-style connection. This modification preserves the gradient flow through the residual branch, preventing signal vanishing from the top layers to the bottom. Unlike prior methods, Keel enables stable training at extreme depths without requiring specialized initialization or complex optimization tricks. Keel trains robustly at depths exceeding 1000 layers and consistently improves perplexity and depth-scaling characteristics over Pre-LN. These findings indicate that Post-LN, when paired with a Highway-style connection, provides a simple and effective foundation for building deeply scalable LLMs, opening the possibility for future infinite-depth architectures.