Papers

The trustworthiness of a retrieval-augmented generation (RAG) system depends on more than the answer it returns, yet many black-box uncertainty methods still read agreement among sampled answers as confidence. That inference fails when repeated samples condition on the same defective retrieval state. The state may be empty, with the model falling back on parametric memory, or populated by a coherent but wrong neighbourhood. In either case, the answers agree because the error is stable. The problem is recognised in deployed RAG, but it has lacked a name, a measurable signature, and a prevalence bound. We supply all three. We name the failure retrieval-state lock-in and diagnose it by separating the three objects a single confidence score conflates: the answer surface, the retrieved evidence, and the retrieval state itself. In an inspectable, ontology-guided knowledge-graph RAG (KG-RAG) system across six question-answering snapshots, we measure the agreement blind spot directly: at five samples per question, 42% of KG-RAG errors and 59% of dense-retrieval errors carry zero answer dispersion, so agreement has nothing to rank, while evidence- and retrieval-state checks still flag most of them. The decomposition supports an auditable decision rule: accepting an answer only when answer, evidence, and retrieval checks all agree that it is low-risk reaches 91.9% pooled precision against a 69.7% accept-all rate. The cost is coverage: it certifies only 7.7% of answers as low-risk. On the clinical calibration domain it reaches 100% precision under an automated judge; this is an in-domain automated-label upper bound, not a clinical safety claim, and still needs human validation. Confidence in RAG is object-specific: when answers agree, the useful question is which part of the pipeline to distrust.

Large language models (LLMs) are trained predominantly on English-language internet text that over-represents certain cultural narratives, raising concerns that models flatten the diversity of non-Western storytelling traditions into a single homogenized archetype. We present a pilot computational study examining this across three maximally distinct Indian regional oral and literary traditions: the Rajasthani Pabuji epic, classical Tamil Sangam poetry, and Bengali folk tales. We collected authentic reference corpora for each tradition (11, 21, and 10 passages respectively) and prompted two LLMs (Claude Sonnet and Gemini) with 54 generation requests spanning three prompt types per tradition - generic, culturally specific, and regional-language. Using Sentence-BERT embeddings and cosine similarity, we measure reference drift (how closely outputs track their own tradition's authentic texts relative to the other two) and cross-tradition convergence (how similar outputs are across traditions). We find that while outputs remain closer to their own tradition's reference than to others, cross-tradition similarity is high (0.52-0.66) relative to what the traditions' genuine distance would predict, indicating partial homogenisation. Unexpectedly, prompting in the regional language (Hindi, Tamil, or Bengali) consistently reduced fidelity to the authentic tradition relative to English prompting, by as much as 27 percentage points for Rajasthani and Bengali traditions. We discuss this against conflicting prior results on multilingual prompting and argue it reflects a difference between eliciting general cultural diversity and simulating one narrow, lesser-documented oral tradition. We position this pilot as a lightweight, scalable complement to recent large-scale human-annotation studies of Indian cultural misrepresentation in LLM-generated stories, as part of a broader doctoral research program.

Nigerian Pidgin is one of Africa's most widely spoken languages, yet remains severely underrepresented in language model evaluation. Existing benchmarks primarily focus on translation, transcription, or generic sentiment analysis, leaving critical aspects of culturally grounded language understanding unmeasured. We introduce Wazobia Eval, a benchmark for evaluating Nigerian Pidgin emotion understanding, sarcasm detection, and cultural reasoning. The benchmark is built on a manually annotated dataset containing over 550 examples and a 16-category emotion taxonomy designed to capture culturally specific emotional registers that are not represented in conventional sentiment frameworks. Wazobia Eval provides standardized evaluation protocols and benchmark tasks for assessing model performance on nuanced Nigerian language understanding. We present the benchmark design, annotation methodology, taxonomy development process, and preliminary pilot evaluation results. Our goal is to provide foundational evaluation infrastructure for Nigerian language AI and establish a reproducible benchmark for future research. The dataset is publicly available at https://huggingface.co/WAZOBIALABS.

While 4-bit weight quantization is critical for deploying Small Language Models (SLMs) on edge devices, evaluations of the resulting performance degradation-the quantization tax-remain overwhelmingly English-centric. We present a zero-shot multilingual evaluation of 4-bit quantization across the Gemma 4 and Qwen 3.5 architectures. Evaluating on eight typo-logically diverse languages using MMLU ProX Lite and GlobalPIQA, we show parameter truncation exposes deep pre-training inequalities. We identify four phenomena: (1) Typological Fragility: low-resource and specific non-Latin scripts suffer representational collapse via architecture-specific double dissociations, failing to generate valid task logits; (2) Home Language Fragility Paradox: foundational pre-training pathways provide limited precision loss protection; (3) Domain-Specific Forgetting: multi-step cross-lingual routing degrades while associative soft-science recall remains robust; and (4) Quantization Resistance: highly saturated, typologically aligned domains resist deterministic degradation, with post-quantization performance gains bounded by statistical noise.

The growing volume of scientific submissions has motivated interest in using large language models (LLMs) to assist peer review. Existing automated novelty assessment approaches typically compare a paper's claimed contributions against prior literature, implicitly assuming that these contributions are accurately realized in the work itself. Human reviewers, however, frequently challenge novelty claims not because similar ideas already exist, but because the methodological evidence presented in the paper does not adequately support them. This internal mismatch between claimed contributions and methodological realization is rarely examined by current LLM-based review systems. To address this gap, we introduce intra-paper claim verification, a framework that evaluates whether novelty claims articulated in a paper are substantiated by the methods used to realize them. The framework employs an LLM to extract novelty claims from the introduction, retrieve claim-relevant methodological evidence, and assess whether the methods substantiate the stated contributions. Assessment is guided by reviewer-inspired evaluation criteria derived inductively from human peer reviews collected from 182 ICLR 2025 papers. These criteria capture recurring reviewer concerns related to novelty, methodology, clarity, and other issues and are used to generate structured reviewer-style assessments of claim substantiation. We evaluate the framework by comparing LLM-generated review comments against human reviewer concerns on a balanced subset of accepted and rejected papers. Human evaluation demonstrates significant alignment between framework-generated assessments and human reviewer concerns, particularly for novelty-related issues. BERTScore further distinguishes corresponding human-LLM review pairs from mismatched controls, indicating that the framework captures concerns consistent with human reviewer observations.

What happens when LLM agents operate with no context outside a turn, minimal prompting, and simple tools? Inspired by swarm engineering, we give collectives of three agents the ability to send messages and manipulate a shared actively decaying text store, introducing evolutionary pressure. The agents spontaneously cooperate, develop storage management strategies, and generate complex evolving cultural artifacts, with no top-down engineering. Using tools from dynamical systems analysis, we show that these behaviours exhibit structured long-range coherence beyond the entropy horizon of the decaying store, consistent with emergent culture in the Sperberian sense.

Speaker diarization, the task of determining "who spoke when" in a multi-speaker recording, is a critical component in applications such as meeting transcription, accessibility tools, and multilingual information retrieval. While end-to-end neural diarization systems have achieved strong performance for English and other high-resource languages, their effectiveness degrades substantially for underrepresented languages where annotated speech data is scarce. This paper investigates speaker diarization for low-resource Nepali-Hindi speech through a multilingual training approach, comparing two modern architectures: EEND with encoder-decoder attractors (EEND-EDA) and EEND with Perceiver-based attractors (DiaPer). Both models are trained on a multilingual corpus combining English speech from LibriSpeech, diverse speaker recordings from VoxCeleb, and separately collected Nepali and Hindi audio, a setup designed to reduce language bias and encourage cross-lingual generalization. We evaluate both models across 2-speaker, 3-speaker, 4-speaker, and mixed-speaker scenarios on LibriSpeech, VoxCeleb, and Nepali-Hindi (NeHi) test sets. DiaPer achieves stronger overall performance than EEND-EDA, particularly in more challenging multi-speaker conditions, obtaining DERs of 3.28%, 2.02%, 4.05%, and 4.76% on NeHi 2-speaker, 3-speaker, 4-speaker, and mixed-speaker settings, respectively, compared to 1.50%, 9.68%, 16.17%, and 11.19% for EEND-EDA. These results demonstrate the viability of Perceiver-based end-to-end neural diarization for low-resource multilingual speech processing.

Although Large Language Models (LLMs) excel in many tasks, their assessment in Portuguese has received less attention, particularly for open-ended, discursive tasks that demand deeper reasoning and generation capabilities. While the original BLUEX benchmark addressed the scarcity of Portuguese evaluation datasets through multiple-choice questions from Brazilian university entrance exams, it did not cover the more challenging second-phase examinations, which require free-form written responses. In this work, we introduce BLUEX v2, a benchmark derived from the second-phase entrance exams of Brazil's two leading universities: UNICAMP (Comvest) and USP (Fuvest), spanning exam years 2022–2025. Our dataset comprises 395 questions unfolding into 919 graded subquestions, with 55.7% of questions containing associated images (represented as context-aware captions during inference to enable evaluation across both vision-capable and text-only models). Each question is annotated with subject area, official reference answers, LLM-generated rubric criteria, and six cognitive capability tags. We evaluate 21 state-of-the-art LLMs using an LLM-as-a-judge protocol. Results reveal a 4.92-point performance spread across models (4.18-9.10 on a 0-10 scale), with Mathematical Reasoning and Image Understanding emerging as the hardest capability dimensions. The evaluation code, model outputs, and dataset are publicly available at https://github.com/TropicAI-Research/BLUEXv2 and on Hugging Face at https://huggingface.co/datasets/Tropic-AI/BLUEX-v2.

Encoder-only transformer models remain essential for production NLP pipelines. We introduce moBERTo, a Portuguese adaptation of ModernBERT obtained through continued pretraining of the ModernBERT-base checkpoint on 60 billion tokens (5 epochs over a 12-billion-token corpus curated from FineWeb2 and filtered with educational and STEM classifiers). We preserve the original architecture, including rotary positional embeddings, alternating local-global attention, flash attention, and unpadding. We evaluate moBERTo across information retrieval (including long-context retrieval at up to 8,192 tokens), document classification, named entity recognition, and natural language understanding. Our best variant, which combines a Portuguese tokenizer with subword-matching embedding transfer and long-context post-training, achieves the highest average reranking nDCG@10 across three Portuguese retrieval benchmarks and the best results on PLUE-PT. Through ablation studies, we show that (i) continued pretraining is strongly preferable to training from scratch, particularly for preserving long-context capabilities; (ii) tokenizer adaptation improves token-level tasks but degrades long-context retrieval; (iii) a dedicated long-context post-training phase at 8,192 tokens further improves reranking and NER; and (iv) encoder-only architectures remain competitive with larger decoder-only alternatives for discriminative tasks. We publicly release the model weights at https://huggingface.co/Tropic-AI/moBERTo and training data at https://huggingface.co/datasets/Tropic-AI/moberto-pretraining-dataset-c4-compatible on Hugging Face.

Training large language models to reason efficiently is a critical challenge. While integrating length-penalizing rewards into Group Relative Policy Optimization (GRPO) aims to reduce verbosity, it frequently triggers reward collapse, severely degrading reasoning capabilities. Through a systematic evaluation of various reward configurations, we identify the root mechanism: GRPO's group normalization creates divergent advantages when incorrect answers receive continuous length penalties. Consequently, methods penalizing the length of incorrect answers are structurally prone to collapse under sustained optimization. Furthermore, restricting penalties exclusively to correct answers avoids this primary failure, but leaves the model susceptible to a stochastic collapse driven by response over-compression. To robustly prevent both failure modes, we propose ACOER (Adaptive Correct-Only Efficiency Reward). ACOER eliminates the structural penalty loop by isolating brevity bonuses to correct completions and prevents stochastic compression via dynamic budget normalization and control-loop penalty adjustments. Evaluated across diverse mathematical reasoning benchmarks, ACOER improves overall accuracy compared to the base model while reducing token generation by over 60%, establishing a fundamentally stable approach for efficiency-aware optimization.

As LLM-powered scams proliferate, black-box forensics for conversational LLM agents offers a path to accountability for systems hidden behind anonymous endpoints. Identifying the base model behind a chatbot endpoint (attribution), without model parameter access or knowledge of the hidden system prompt, would let investigators trace AI-enabled scams back to the providers whose models power them. Detecting when two endpoints run the exact same system prompt (fingerprinting), even one novel and unseen, would link individual scams into criminal networks and expose silent API changes. We conduct an empirical investigation of both capabilities. Our attribution classifiers identify the base model behind an agent with 98% accuracy from a few turns of non-adversarial conversation. Attribution of system prompts, while possible, requires retraining on a large amount of data for each prompt; system prompts in the wild are unbounded and ever-changing, making this approach costly. To tackle this more open-ended setting, our cross-encoder fingerprinting method achieves an AUC of 0.768 and an F1 of 0.703 on entirely unseen system prompts, and aggregating 50 interaction conversations from each target agent boosts AUC to 0.943. Conversational agents with unseen system prompts can thus be fingerprinted with robust accuracy from a few turns of ordinary conversation.

We introduce VISTA Architect, a database-oriented AI architecture for integrating large language models (LLMs) with longitudinal electronic health records (EHRs). At ingestion, it transforms complex clinical documentation into a persistent, provenance-linked knowledge graph, eliminating repeated reprocessing of raw records at query time. The architecture has two layers: a source-faithful MEDS Graph preserving granular EHR structure with full provenance, and a clinically abstracted Timeline Object Architecture (TOA) that uses graph-guided LLM extraction to synthesize a concise timeline of deduplicated, temporally coherent clinical events. This addresses key limitations of direct long-context prompting and retrieval-augmented generation (RAG), which often miss temporal relationships and incur high cost and latency from repeated raw-text processing. By precomputing clinical synthesis once, downstream queries access an organized patient state and traverse to source documentation only when detailed verification is needed. We demonstrate the system in multidisciplinary thoracic oncology tumor boards at Stanford Medicine, where precise reconstruction of patient histories is critical. Across 1,180 patients, VISTA Architect achieved 96.4% accuracy (mean 9.75/10) on 15 tumor board-salient variables (17,700 evaluations; 95% CI 96.1-96.7%), surpassing a matched BM25 RAG baseline and recent benchmarks for LLM-based clinical extraction. An agentic interface reduced preparation for a 30-patient held-out cohort to about 2.2 minutes without sacrificing accuracy. While configured here for thoracic oncology, the modular design adapts to other specialties through customizable event definitions, episode structures, and agentic tools; validation beyond thoracic oncology remains future work.

Multi-hop question answering remains challenging for Retrieval-Augmented Generation (RAG) because existing approaches either propagate errors across iterative retrieval rounds or over-generate reasoning steps, increasing cost without improving accuracy. We propose Grounded Delta Planning RAG (GDP-RAG), a plan-based framework that targets only the information delta based on three simple design choices: (1) preliminary retrieval to ground planning before execution, (2) a gap-conditioned planning prompt that asks only for missing information, and (3) a skeletal trajectory that pairs each subquery with a Thought capturing evidence from preliminary retrieval and carrying it through to the final answer. GDP-RAG focuses computation on unresolved gaps, yielding concise, reliable reasoning trajectories. Extensive experiments on HotpotQA, 2WikiMultiHopQA, and MuSiQue show that GDP-RAG achieves the highest accuracy (60.63%) among all compared systems while maintaining a cost-of-pass of 0.51, 22% lower than PAR-RAG (0.65) and 68% lower than KnowTrace (1.57), with no method achieving both higher accuracy and lower cost.

Knowledge editing aims to efficiently update factual information in Large Language Models (LLMs) without full retraining. However, existing methods still suffer from performance degradation in batch knowledge editing. We identify that semantic representation entanglement, such as overlapping concepts and shared syntactic patterns, accumulates interference in the representation space and reduces editing precision. To bridge this gap, in this paper, we propose Orthogonal Representation Editing (ORE), which performs edits in the hidden representation space of LLMs by constructing a general semantic subspace and enforcing orthogonal constraints on edit vectors, effectively decoupling semantic entanglement. Furthermore, we introduce a gated non-linear representation head to enable adaptive learning of editing locations and precise control over knowledge injection. Extensive experiments show that ORE outperforms existing methods and achieves superior performance in cross-lingual knowledge editing scenarios. We release our code at https://github.com/YVVH/ORE.

Learning to read cuneiform tablets is an extremely demanding task; consequently, of the roughly half million excavated tablets, only a small fraction has been analysed by Assyriologists. Computer vision offers a promising avenue for decipherment but requires large, densely annotated datasets. To address this limitation, the largest annotated cuneiform sign dataset to date is used, and a Deformable Detection Transformer (DETR)-based object detection model is evaluated under two class granularities of 173 and 106 classes. The proposed system integrates automatic tablet-side extraction, heuristic line grouping, and n-gram-based textual similarity evaluation to bridge visual sign detection and textual structure, and achieves consistent improvements of up to 28-37% over prior work on COCO-style detection metrics. At inference, the method is applied to 87,668 tablet fragments from the Electronic Babylonian Library (eBL) corpus, producing nearly 2.9 million sign detections. Although the approach operates without linguistic priors and remains sensitive to tablet damage and layout variability, it provides a scalable and interpretable foundation for corpus-wide cuneiform analysis and supports future integration with multimodal and linguistic modelling frameworks.

Large language models (LLMs) achieve strong relation extraction (RE), but their computational demands and reliance on proprietary APIs limit deployment in resource-constrained or privacy-sensitive settings. We investigate how far small language models (SLMs) can close this gap across general-domain and literary text. We evaluate five models from 360M to 3B parameters under three domain-composition regimes and two prompt-conditioned tuning styles (30 configurations), comparing them with zero-shot frontier LLMs and a discriminative RoBERTa baseline. Across nine benchmarks, the best sub-billion model, Qwen2.5-0.5B fine-tuned on pooled general-domain data, achieves a general-domain positive-class micro-F1 of 0.83, versus 0.69 for GPT-5.4 and 0.66 for Claude Sonnet 4.6 evaluated zero-shot. This does not imply that SLMs are intrinsically stronger; rather, targeted task adaptation enables 4-bit models deployable on a single consumer GPU to outperform general-purpose frontier systems under this protocol. An in-domain RoBERTa baseline also exceeds both frontier models, indicating that the gain stems from task adaptation rather than generative decoding. On literary RE, tuned SLMs reach 0.92 on the human-annotated Biographical benchmark versus 0.83 for GPT-5.4, and 0.833 versus 0.578 on the two-benchmark literary average. A targeted domain-adaptive pretraining case study yields no practically meaningful gain over supervised fine-tuning, while the cleanest within-family scale comparison shows only marginal improvement. These results show that, when task-specific data are available, compact task-adapted models can provide accurate, private, and hardware-efficient RE.

We extend our prior work on Text2DSL automatic generation of domain-specific language (DSL) code from natural language descriptions along two complementary axes. First, we replace prompt-only synthetic generation with context-aware distillation, in which a teacher large language model (DeepSeek-V4-Flash) operates under an explicitly defined structured context comprising a BNF grammar, an API specification, and a closed identifier vocabulary; the resulting corpus is verified by a two-tier pipeline combining AST validation through esprima and runtime acceptance through the production polkitd daemon and the pkcheck client. This scales the verified PolkitBench corpus from 4,204 to 10,073 natural-language-to-Polkit-rule pairs at 100.0% AST validity and 99.7% runtime pass rate. Second, we conduct the per-component factorial ablation of structured context that was identified as future work in the precursor study: eight conditions C0-C7 are evaluated on GigaChat-10B-A1.8B with the new corpus. Three findings emerge. (i) The new harder corpus collapses the baseline mode (Syntax Valid 97.6% -> 58.5%, Combined Score 0.482 -> 0.252), whereas the context-enhanced mode degrades only marginally (Syntax 98.6% -> 97.4%, Combined 0.801 -> 0.750), confirming that structured context is not a cosmetic improvement but a load-bearing mechanism. (ii) The best absolute condition is the full context C7 across all metrics, while the strongest partial conditions (C5 = BNF + Vocabulary, C6 = API + Vocabulary) both contain the vocabulary. (iii) A Shapley-style decomposition assigns the largest semantic-quality effect to the vocabulary (Combined +0.198), the largest structural-validity effects to API (+24.7 pp) and BNF (+22.3 pp).

Reinforcement Learning from Verifiable Rewards (RLVR) has emerged as a promising framework for enhancing the reasoning ability of large language models. However, much of the existing work is guided by heuristic intuition, leading to divergent algorithmic choices, even contradictory ones that nevertheless report empirical gains. To better understand this phenomenon, we conduct a theoretical analysis of RLVR updates. Our study reveals that differences in off-policy degree, determined by the number of gradient steps per rollout, substantially affect the distribution of importance sampling ratios and their clipping behavior, thereby altering which tokens dominate the update. Building on this insight, we characterize gradient expectation as the central quantity governing update dynamics and analyze the roles of token probability, advantage, and importance sampling ratio. Motivated by these findings, we propose Adaptive Clip Policy Optimization (ACPO), which adjusts clipping boundaries across token groups according to the empirical variance of their importance sampling ratios. Experiments on 3B and 7B models across diverse reasoning benchmarks, spanning mathematical problem solving, tabular QA, and logic puzzles, demonstrate that ACPO outperforms strong baselines such as DAPO and CISPO. These results demonstrate that principled, analysis-driven approaches yield more robust and effective RLVR methods. Code is available in: https://github.com/Control-derek/ACPO

Few-shot prompt learning is an effective strategy for adapting CLIP to downstream tasks, but class-only prompt optimization can overfit base-class supervision and weaken transfer to unseen classes. We propose Concept-Constrained Prompt Learning (CCPL), a lightweight regularization framework that anchors learnable class prompts to frozen concept-level text prototypes without updating CLIP encoders. CCPL learns a set of shared context tokens, instantiates class prompts by appending class names, and constructs frozen concept prototypes from a class-level concept bank. During training, a text-space cosine consistency objective aligns learnable class-prompt embeddings with frozen concept prototypes; concept dropout provides additional regularization against over-reliance on fixed concept lists. At inference, CCPL optionally fuses class-prompt logits with concept-prototype logits using a controllable ensemble weight alpha. Our default configuration uses text-space concept regularization lambda = 0.5, concept dropout p = 0.3 and weak concept-guided fusion (alpha = 0.1), with no KL-based prediction consistency term. Experiments under identical automatically-generated fallback splits show that CCPL improves the base-to-new harmonic mean on DTD (+0.6) and EuroSAT (+2.9) compared with CoOp, while remaining near-neutral on OxfordPets (-0.1). Ablations indicate that text-space concept regularization is consistently beneficial, while the best concept-guided inference strength is dataset- and protocol-sensitive. These results suggest concept constraints are most effective when concept prototypes align naturally with dataset semantics, and identify fine-grained categories as a current boundary condition. The code is released at: https://github.com/richael-sang/concept-constrained-prompt-learning.

Chain-of-Thought (CoT) has become a standard method for improving reasoning capabilities in large language models (LLMs) by eliciting step-by-step thinking, but its effectiveness in multimodal tasks remains unclear. In this paper, we aim to systematically investigate the key question: What can multimodal Chain-of-Thought reasoning do, and where and why does it fall short? To this end, we evaluate 12 multimodal tasks across perception and reasoning categories using both 14 non-reasoning models and 8 reasoning models. Our analysis reveals several important findings: (1) CoT is not a free lunch and should be used selectively depending on the specific requirements of each task. For perception tasks, CoT can lead to undesirable side effects, such as reduced performance in visual grounding and object counting. In contrast, it proves effective for reasoning tasks involving mathematical, scientific, and multi-image reasoning; (2) Compared to original models, existing open-source multimodal reasoning models often yield only marginal overall improvements, possibly due to an overemphasis on mathematical reasoning at the expense of broader capabilities; (3) Visual reasoning remains a key bottleneck for current multimodal CoT, as models exhibit a Look Light, Think Heavy pattern where verbal reflection rises and falls during reasoning, whereas visual reflection consistently diminishes. These findings suggest that while multimodal CoT handles verbal reflection relatively well, it lacks the ability to maintain deep visual introspection throughout the reasoning process.

Computer use agents (CUAs) have advanced rapidly in desktop automation, and a growing number of users deploy CUAs such as OpenClaw on Mac Mini for always-on automation. However, existing benchmarks, including those for macOS, evaluate agents without framework augmentation and rely on binary evaluation. As a result, they fail to capture both the framework capabilities leveraged by modern CUAs and the partial progress on long-horizon, multi-application tasks. We present MacAgentBench, a comprehensive macOS agent benchmark comprising 676 tasks across 25 applications, with nearly 60% involving both GUI and CLI interaction. The benchmark adopts deterministic rule-based evaluation and introduces fine-grained multi-checkpoint scoring with capability annotations for multi-application tasks. Experiments across three frameworks and 16 models show that the best configuration, Claude Opus 4.6 on OpenClaw, attains 73.7% Pass@1, while this advantage is primarily driven by the skill library rather than by framework design. Fine-grained metrics further reveal that models with similar Pass@1 can differ substantially in sub-goal completion. Our code and data are publicly available at https://github.com/JetAstra/MacAgentBench.

Autoregressive visual models (AVMs) based on next-scale prediction have emerged as a prominent paradigm for image and video synthesis. However, decomposing the generation process into discrete scales with varying granularities in AVM makes semantic errors difficult to identify and correct, thereby undermining the quality of the final output. Prior efforts to enhance AVM can be categorized into training-based and training-free approaches. Although training-based efforts to enhance AVM generation quality come at substantial computational cost, existing training-free methods neglect intermediate generation states, leaving semantic errors undiagnosed and allowing them to accumulate into the final output. In this paper, we focus on training-free paradigms and propose Gazer, a framework that integrates multimodal large language model feedback into the AVM sampling loop for in-generation semantic correction. Concretely, Gazer operates via two cooperating stages: the Reflective Diagnosis stage diagnoses semantic errors from intermediate states, while the Semantic Correction stage rewinds and rectifies the generation trajectory to realign with the target prompt. Experiments on compositional image and video benchmarks demonstrate that Gazer improves semantic alignment and compositional accuracy across multiple AVMs without additional training.

In open-ended generation, LLMs frequently fall into the "likelihood trap", characterized by repetitive degeneration and vocabulary dullness, resulting in a discrepancy between machine-generated and human-written text. While post-hoc tail truncation (e.g., Top-p, Min-p) avoids sampling from the unreliable tail, it can misalign generation with human lexical preferences by over-sampling from the uncalibrated head; fixed scalar repetition penalties, in turn, ignore how the scale of the logit distribution varies across inference steps, which can disrupt semantic coherence. To address both shortcomings, we propose Variance-Calibrated Modulation (VCM), a training-free pre-decoding intervention. VCM directly reshapes the probability distribution prior to truncation via two dynamic mechanisms: (1) Contextual Searchlight via PMI, which naturally suppresses global stopwords and elevates context-evoked tokens, and (2) Adaptive Self-Debiasing, which utilizes real-time logit standard deviation to provide scale-invariant penalization. In experiments across open-ended generation, factual QA, and mathematical reasoning, we show that VCM consistently mitigates the likelihood trap. With negligible computational overhead, VCM integrates with existing decoding strategies, improving diversity and coherence and, particularly at higher decoding temperatures, reasoning accuracy. Our code is publicly available on GitHub: https://github.com/AetherDing/VCM

Decision-making in real-world settings rarely follows a fixed script. Instead, it unfolds as a dynamic reasoning process in which the appropriate course of action evolves as new context and data become available. Traditional Business Process Management systems provide rigor, determinism, and auditability, yet they generally struggle to adapt their execution at runtime. Conversely, agentic systems based on Large Language Models (LLMs) bring flexibility to decision-making, but they are inherently opaque, often unreliable, and suffer from significant scalability constraints when operating over large datasets. To combine these complementary paradigms, we introduce VADAOrchestra, a neurosymbolic framework that models complex workflows as evolving reasoning processes. The framework adopts a hybrid approach: given a user query and a collection of data sources, an LLM-based orchestrator incrementally plans and adapts the workflow. This is encoded as a logic program in a fragment of Datalog+/- where predicates correspond to tool invocations and rules represent both predefined domain dependencies and logic constructs synthesized on demand to manipulate intermediate results. All logical inference tasks are then executed by a state-of-the-art Datalog+/- symbolic engine. This approach provides a verifiable reasoning trace, supporting the auditability and reproducibility of the entire process. Furthermore, by decoupling high-level orchestration from symbolic inference, it addresses scalability concerns, enabling complex reasoning over large datasets through targeted data querying. We evaluate VADAOrchestra on real-world financial use cases, demonstrating faithfulness, scalability, and explainability compared to standard agentic architectures.

Multilingual Language Models like mBERT are widely used for low-resource NLP, yet their adaptation to morphologically inconsistent languages such as Roman Urdu remains underexplored. Roman Urdu spelling variation causes severe sub-word fragmentation, averaging 1.50 sub-words per token. We propose ROMEVA (Roman Urdu Embedding-preserving Vocabulary Adaptation), which combines sub-word-average initialization and a PCA-guided anchor loss to stabilize embeddings during vocabulary expansion. Using a 36,130-comment Roman Urdu corpus, we add 500 highly fragmented tokens to mBERT and compare naive fine-tuning, sub-word-aware fine-tuning, and ROMEVA. While ROMEVA most effectively preserves the pretrained embedding space, naive fine-tuning achieves the strongest downstream sentiment classification performance. These findings reveal a disconnect between embedding stability and downstream performance, suggesting that stronger adaptation may be preferable to strict embedding preservation in morphologically inconsistent languages.

Large Language Models (LLMs) generate fluent long-form text, however, often add unsupported factual claims. Existing verification techniques improve factuality by grounding generation in external evidence. However, the same verification policy usually applies to all claims despite being differences in hallucination risks. We propose FACTOR (FACTuality-Oriented Risk-aware Verification), an inference-time model that adapts verification criteria according to claim-level uncertainty. FACTOR combines uncertainty estimation, adaptive language inference verification, and candidate re-ranking to allocate verification effort where it is most needed. We evaluate FACTOR on FactScore benchmark showing that adaptive verification improves factuality while reducing verification cost simultaneously. We further perform different ablation studies to identify the primary driver of these gains. Our results show the effective and model-agnostic performance of FACTOR for improving factuality in long-form generation.

Speech language models (SLMs) increasingly combine speech and text, often by interleaving their tokens within a single sequence. Yet how these two modalities interact in the model's latent space remains unclear. In this work, we analyze interleaved speech–text LMs from different model families and training configurations using three complementary methods. We reveal that these models pass through an implicit latent transcription phase in which the text token matching the spoken word becomes decodable in intermediate layers, despite not being trained for speech recognition. This phenomenon occurs in diverse, natural speech, and intermediate representations also encode likely text continuations. We further show that implicit transcription emerges most when combining text-LM pretraining and speech–text interleaving, and that its prevalence is positively associated with spoken factual-knowledge retrieval. Our analysis sheds light on the internal interaction between speech and text modalities in interleaved SLMs.

As LLM-generated text is increasingly used, especially in fictional domains, we explore how much LLM-generated stories differ from human-written stories. In this work, we focus on characters. We borrow definitions from narratology to analyze eight intricate dimensions of character, such as stylization and wholeness. These dimensions consider more than just basic characteristics. They assess how characters are portrayed within their stories. After automatically inferring categories of characters within both LLM and human-written stories, we compare and contrast these two sets of stories. We consider the following overarching questions: (1) Do LLMs and human-written stories have similar characters? and (2) Do LLMs generate stories with a variety of characters? Our analysis includes research questions that focus on stories generated by popular LLMs and recently published human-written stories. We describe a number of interesting similarities, differences and key takeaways.

Because large language models (LLMs) are impressively successful in predicting text, it appears that they must have access to a 'world model' representing causal and definitional structure. However, the dominant formalisms of modern causal inference – Judea Pearl's interventionist approach and the Neyman-Rubin potential outcomes framework – struggle to illuminate how LLMs learn causal structure. I resolve this puzzle by arguing that LLMs employ a specific inductive approach based on a difference-making logic – sometimes called variational induction. I demonstrate how central aspects of this logic are realized during training, where LLMs require enormous amounts of text data from a wide range of contexts to identify difference- and indifference-makers within word sequences. Furthermore, I analyze specific architectural features of LLMs – such as token embeddings and self-attention – to determine their roles in variational induction. The difference-making logic of LLMs fundamentally parallels the experimental method, where causal relations are derived by systematically varying individual circumstances to determine their influence on a phenomenon.

A recent Nature Medicine study reports that general-purpose frontier LLMs outperform specialized retrieval-augmented clinical tools on medical benchmarks, and that retrieval can hurt strong models. We ask the natural follow-up: does structured knowledge-graph (KG) grounding change this, and when does grounding help at all? We contribute two results. First, a reproduction: the study's headline HealthBench score (~88) is the Consensus variant, not full HealthBench, where frontier models and ideal completions both score ~46-47 under a physician-calibrated grader (agreement 82.5%); we reproduce GPT-5.2 Consensus =90.9 and flag a score-deflating grader bug. Second, a knowledge-boundary result. Using a graph+vector engine (samyama-graph) over the public biomedical KG PrimeKG, neither naive triple retrieval nor an agentic natural-language-to-Cypher loop (82% successful queries) improves MedQA across a weak-to-strong model ladder (all |Delta| <= 3.4). On a synthetic counterfactual KG, and on a hybrid benchmark mixing known and novel facts, the identical pipeline lifts out-of-training accuracy from chance to ~100% (+68 to +79) while adding nothing on known facts (a no-LLM arm answers both). Across three regimes (no-knowledge, graph-aided, hybrid), grounding helps only insofar as the decisive fact lies outside the model's training – public-KG facts are redundant, private and novel data are where it pays – matching the study's institutional-data caveat.