Papers

Unsupervised syllabic tokenization aims to learn discrete syllabic tokens that capture latent linguistic content-related structure from raw speech. Recent syllabic tokenization methods employ teacher-student distillation of the pretrained HuBERT to organize latent speech frame representations into syllabic segments. However, when trained with an utterance-level cross-entropy objective, the model predicts speaker identity rather than linguistic content, thereby compromising the purity of syllabic tokens. To address this problem, we propose a speaker-disentangled syllabic tokenizer that regresses speaker-perturbed student representations toward clean teacher targets within fixed-length chunks. Experimental results demonstrate that our proposed method achieves state-of-the-art performance in syllable boundary detection and syllabic segment clustering. Moreover, a speech language model trained on our syllabic tokens achieves a 7% relative improvement in syntactic and semantic understanding over the phone-level SpiRit-LM.

Distinguishing Large Language Model (LLM) generated text from human writing is a critical and difficult challenge. While LLMs are trained to write like humans, we hypothesize that this training leaves an indelible mark. LLMs develop a particularly strong aversion to token repetition very early in training. This bias persists as a ”Vestigial Heuristic” (a developmental artifact) that is activated in LLM-generated text, separating LLM from human writing. To probe this phenomenon, we introduce Telescope Perplexity, a metric that evaluates the token repetition of the model, $P(s_i | s_{1:i})$ . Our empirical investigation reveals that the Telescope Perplexity signature emerges early in pre-training, and Telescope Perplexity empirically enables highly effective zero-shot LLM detection. We show state-of-the-art or competitive performance across diverse datasets (including modern evaluation sets we introduce), reference models, and perturbation schemes with greater efficiency than other methods.

We present GEPARD (Generative, Prosody-aware, Autoregressive text-to-speech model for Realtime Dialogue), a streaming text-to-speech model for real-time spoken dialogue. GEPARD generates speech autoregressively with an LLM backbone - text and audio embeddings are trained together in a single decoder-only model - and decodes it to a waveform with an FSQ-based neural codec, streaming audio chunk-by-chunk as text arrives. Our central goal is a TTS architecture served by a standard LLM engine (vLLM) without modifying its compute kernels. This defines the overarching design principle: the backbone is a standard full-attention transformer, while all non-trivial auxiliary mechanisms - zero-shot voice cloning, text augmentation, and classifier-free guidance - are moved out of the autoregressive decode loop into prefill, or distilled directly into the weights. On streaming end-to-end inference, a single stream reaches a Real-Time Factor of about 0.067 (roughly 15x faster than real-time); under 256 concurrent streams the system reaches an aggregate speedup of about 204x on a single server-class GPU. We detail: (1) system-level solutions for vLLM-native serving; (2) the "short register" (1-2 word) failure mode of autoregressive speech decoders, with diagnostic probes and a mitigation; and (3) distillation of two-pass classifier-free guidance over text into single-pass weights via Direct Preference Optimization (DPO).

Large domain-specific language models such as BioBERT and ClinicalBERT achieve strong performance on biomedical NLP tasks, but their computational demands make them impractical for many real-world deployments. General-purpose, parameter-efficient models such as DistilBERT are lightweight yet lack the domain knowledge required for specialized tasks such as PICO (Population, Intervention, Comparison, Outcome) classification. We introduce Distilled Rapid Embedding Transfer (DRET), a knowledge-transfer paradigm that injects biomedical domain knowledge from large specialized models into a smaller general-purpose model without retraining on the original specialized corpora. DRET is developed as an iterative family of strategies: a unified tokenizer-merge strategy (DRET 1.x), hybrid embedding averaging (DRET 2.0), and a priority-based embedding-transfer mechanism (DRET 3.x) that hierarchically selects embeddings from the most authoritative source models, further combined with embedding-layer freezing, differential learning rates, label propagation, and imbalance-aware loss functions (DRET 4.x). We evaluate DRET on token-level PICO classification using the EBM-NLP corpus under severe class imbalance, across a twelve-metric battery. DRET-enhanced DistilBERT (66M parameters) attains balanced accuracy, recall, and ROC-AUC competitive with, and on several class-wise metrics exceeding, models an order of magnitude larger, while retaining DistilBERT's efficiency. We further show that transfer occurs at the embedding level through cosine-similarity, semantic-shift, and t-SNE analyses. DRET offers a scalable, resource-efficient route to near-domain-expert performance for biomedical text mining, with direct application to automated systematic literature reviews and clinical decision support.

Multimodal large language models have advanced rapidly, yet most remain English-centric, as scaling multilingual multimodal instruction tuning is limited by the scarcity and high cost of high-quality non-English image-text supervision. Although multilingual text data is abundant, naive textual fine-tuning can disrupt vision-language alignment and induce catastrophic forgetting. We propose Vision-Free Adaptation (VFA), a framework that decouples multilingual language enhancement from visual alignment by composing complementary task vectors over a shared LLM backbone. Specifically, we fine-tune a base LLM on multilingual text data to derive a multilingual task vector, which is then merged with the vision-aligned task vector of an MLLM. Experiments on five MLLMs across six multilingual multimodal benchmarks show consistent improvements while preserving both general multimodal and text-only capabilities. Moreover, using less than 2% of the text data, VFA narrows the gap to the fully multimodal-trained model, demonstrating its data efficiency.

We introduce Telco-GAIA, a bilingual, multi-modal benchmark for evaluating tool-using agents on the data of a real-world telecommunications operator. Telco-GAIA comprises 100 human-verified question-answering tasks, in English and Arabic, that each demand multi-hop reasoning (4.2 hops on average) over three heterogeneous sources: a static website snapshot (HTML, images, and linked PDFs), a synthetic relational SQL database, and external web archives, spanning text, image, and tabular modalities. The benchmark is delivered as a sandboxed Docker environment and scored by normalized exact string matching, making evaluation objective, deterministic, and reproducible over time without any LLM-as-a-Judge. Evaluating a purpose-built reference agent across twelve commercial and open LLMs, we find Telco-GAIA challenging: even the strongest model solves only 71% of tasks; under a moderate cost budget, this falls to about 40%, and the visually grounded categories remain the weakest, where the average backend scores below 30%, leaving substantial headroom in document and image understanding. Telco-GAIA offers a rigorous, reproducible testbed for enterprise agents and a template for constructing closed-domain benchmarks.

Multi-entity compositional questions pose significant challenges to existing retrieval-augmented language models. Conventional methods fall into a dilemma: standard RAG lacks dynamic reasoning, traditional Graph-RAG is limited by structural sparsity, and LLM-constructed Graph-RAG incurs prohibitive costs. We propose \fwa, a unified framework that embeds unstructured text into structured knowledge graphs, creating a heterogeneous network for flexible evidence retrieval. Reasoning is formulated as adaptive structure induction, learned via a robust two-stage process: (1) imitation learning distills heuristic expert signals, and (2) reinforcement learning refines the policy using LLM-driven preference rewards. Experiments demonstrate that {\fwa} effectively merges textual richness with structural knowledge, outperforming SOTA baselines in answer accuracy and reasoning fidelity while maintaining extremely low token costs and near real-time inference((code available at https://github.com/wjywjy123/HyGRL) .

Block-sparse attention scales long-context language models by replacing the O(N^2) softmax with a per-query top-k selection over key blocks. This cutoff is myopic: when the k-th and (k+1)-th blocks are nearly tied in score, the selector commits without spending extra budget, and a dropped block carrying answer evidence is unrecoverable downstream. We propose a value-of-information router that measures, for each query, how decisively the top-k cut was made, and doubles the kept set for the queries where that gap is smallest; the rule is backbone-agnostic and stacks with existing block-scoring methods such as Quest. On LongBench-v2 medium at n=215 (the entire dataset subset), router-on-Quest reaches paired recall 0.75 vs. top-k 0.47 – +28 pp over the SSA-style baseline (McNemar p<0.01) – and lands within 2 pp of dense on RULER NIAH multikey at the same context. The lift reproduces on four models from three architectures (Qwen2.5, Mistral-Nemo, Qwen3.6). At 128K, the router preserves 0.81 and 0.89 of dense accuracy on Qwen2.5-7B-1M and Qwen3.6 (vs. SSA-style top-k at 0.09 on the former) while the fused selection-plus-kernel pipeline runs at 0.62x and 0.80x dense wall time.

Recent hallucination detection techniques in large language models (LLMs) focus on directly extracting features from a model's internal representations and training a classifier on these features to detect hallucinations, demonstrating promising results. Notwithstanding this advancement, most internal-state hallucination detection techniques have been explored predominantly in English, raising the question of whether such internal signals generalize across different languages and domains. To address this gap, we present CrossHallu, the first study to evaluate the cross-lingual and cross-domain generalization of hallucination detection using internal representations from six LLMs on the generative question-answering task. We conduct a systematic Arabic <-> English evaluation using TruthfulQA, an Arabic translated version of TruthfulQA, and HalluScore. This evaluation encompasses monolingual training and testing, cross-lingual transfer, cross-domain transfer, and combined cross-lingual and cross-domain transfer. The results reveal that internal-state hallucination signals in LLMs transfer across languages and domains for most models, with cross-lingual performance highly dependent on both class separability and language alignment in the feature space, whereas cross-domain transfer within Arabic varies depending on the training and testing datasets used for the hallucination detector. The code is publicly available at https://github.com/aishaalansari57/CrossHal.

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

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

We present a candidate-constrained retrieval-augmented generation system for LongEval-RAG, where each query is associated with an organizer-provided candidate set and all retrieved evidence and final citations must remain within that set. The system combines deterministic provenance tracking with passage-based retrieval, deterministic query expansion, pseudo-relevance feedback (PRF), reciprocal rank fusion (RRF), lightweight evidence reranking, citation-aware evidence aggregation, and optional MiniLM sentence reranking. We evaluate ten pipeline variants using a primary organizer evaluation and a supplementary self-generated diagnostic protocol. The primary evaluation shows that the strongest balanced variant is rule-minilm: a rule-based chunking pipeline with query expansion, PRF, RRF, reranking, citation prior, and late MiniLM sentence selection. This variant obtains the highest BERTScore, retrieval precision, nugget coverage, and average grade among our submissions. The result suggests that the main gain does not come from more complex semantic or topic-shift chunking, but from pairing stable rule-based evidence units with sentence-level neural selection before generation. The supplementary LLM-judge evaluation remains useful for early diagnosis and additional analysis, but it emphasizes different systems than the primary gold-answer and nugget-based evaluation, highlighting the need for multi-metric RAG evaluation.

Repeated LLM calls are the standard way to estimate how trustworthy a Text-to-SQL result is: run the pipeline multiple times, judge each SQL execution, and use the consistency of the verdicts as a confidence signal. The open question is when to stop, when the consistency has converged. We formulate this as a convergence-prediction problem and train a family of lightweight 1-D models that observe the running consistency trajectory and decide, at each step, whether further runs are unlikely to shift it materially, and we benchmark them against a principled Beta-Bernoulli stopping rule and a learned run-count baseline. On the BIRD benchmark and two production customer datasets, our method adapts its stopping point to each user question, halting sooner when consistency converges early and continuing longer when it converges late. We further show that the weak serial correlation between runs lets us permute their order as a training augmentation, controlled by a tunable shuffling weight. Performance stays consistent across the three datasets, and to mimic an imperfect production judge we inject noise into the correct/incorrect verdicts obtained by comparing the generated and ground-truth SQL results, showing that the method still predicts convergence reliably.

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

Task-oriented dialogue systems often fail before the final breakdown is obvious, but most evaluation only measures failure after the conversation has already gone wrong. We present TRACER, a method for early failure detection in task-oriented dialogue. TRACER predicts from a partial dialogue whether the full conversation will eventually fail by combining simple trajectory signals from belief-state changes with text representations of the evolving dialogue state. We evaluate the method in both oracle and generated belief-state settings, and test how well it works when only 25%, 50%, 75%, or 100% of the dialogue is visible. Across these settings, TRACER detects useful failure signals well before the end of the conversation and outperforms heuristic, classical, and single-stream baselines. These results suggest that early failure detection can provide a practical warning signal for dialogue systems before the interaction fully breaks down.

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

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

Dialectal data are scarce relative to Modern Standard Arabic (MSA), causing Arabic LLMs to overproduce MSA and struggle with dialectally accurate generation. This raises a fundamental interpretability question about where and how dialectal features are encoded within model internals and whether these representations can improve dialect generation without fine-tuning. We study two inference-time approaches as interpretability probes and control mechanisms. First, neuron-level analysis identifies sparse populations that encode dialect-specific features and tests whether amplifying or suppressing them steers model outputs toward target dialects. Second, vector steering extracts dialect-specific activation directions and injects them during inference, motivated by feature entanglement at the neuron level. We find that these neurons are real but only partially explanatory. They occupy under 1% of MLP dimensions but span only 5% to 21% of the residual dialect direction. This limited coverage is causally consequential. Neuron steering reinforces dialect in some varieties when the prompt is already dialectal but cannot induce it from MSA prompts, whereas vector steering succeeds in both settings. Arabic dialects are therefore steerable mainly through distributed rather than localized representations

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

LLMs are increasingly deployed as post-hoc explainers of AI-generated outputs, yet it remains unclear whether they can reliably communicate probabilistic information in natural language. For this role to be viable, models must produce identical verbal descriptions for identical inputs, and select descriptions that accurately reflect the magnitude of the underlying numerical quantities. We evaluate whether nine LLMs meet these requirements within a two-stage prediction pipeline, in which an upstream model has produced probabilistic outputs characterized by their likelihood and uncertainty, and LLMs are tasked with selecting an appropriate verbal descriptor for each. We simulate predictions from an upstream model by taking samples from a Beta distribution parameterized by its mode and prior sample size. We then prompt LLMs to explain these predictions under six domain contexts and with ten temperature settings, and repeating each experiment ten times. We find that LLMs are generally consistent but miscalibrated, with substantially weaker performance on uncertainty than on likelihood tasks. Providing models with precomputed summary statistics (mode and prior sample size) reduced sensitivity to contextual framing but did not resolve the underlying miscalibration, suggesting that the bottleneck resides in the verbalization step itself. These findings indicate that current LLMs do not yet constitute reliable zero-shot standalone risk communication tools for probabilistic predictions.

As LLMs generate increasingly long outputs, effective uncertainty estimation must identify errors at fine-grained levels rather than discard entire responses. While such methods exist, evaluating uncertainty at any resolution (token to an entire generation) is challenging and highly sensitive to label imperfections, making zero-noise benchmarks essential; yet, long-form generation benchmarks tend to rely on fallible labels rather than deterministic ground truth. We introduce Single-answer Atomic Long-form Target (SALT), a benchmark of six procedurally generated tasks with single deterministic long textual ground truths, enabling unit-level evaluation of correctness, calibration, and ranking without external judges. Equipped with SALT, our analysis of 50+ LLMs reveals key insights: We identify which confidence functions dominate each uncertainty aspect and show that confidence ranking largely breaks at atomic resolution, even when clearer separability emerges at coarser line-level units. SALT further enables controlled atom-level interventions throughout generation, revealing two separable drivers of future errors: propagation from corrupted prefixes, dominated by global context correctness, and bounded degradation from increasing answer-context length. Finally, we demonstrate that reasoning, via Chain-of-Thought prompting or internalized through training, introduces a trade-off, improving accuracy while degrading confidence ranking. These findings directly impact risk-critical applications requiring reliable error identification and mitigation.

Artificial intelligence has advanced scientific discovery, but most AI4Science systems remain fragmented tools that rely on humans to coordinate problem formulation, literature grounding, model use, simulation, validation, and knowledge reuse. This paper presents SCION (Scientific Collaborative Innovation with Agentic Organizational Nexus), an agentic scientific operating system that acts as an organizational nexus. Through a Science Agent serving as a Meta-Harness, SCION connects scientific tasks, tools, agents, artifacts, and memory, transforming research into an executable, auditable, and reusable operational process. At its core is the Research Execution Plan (REP), which compiles high-level scientific intent into staged objectives, dependencies, verification checkpoints, tool requirements, expected artifacts, and fallback conditions. SCION further integrates hierarchical multi-agent execution, profile-driven specialization, selective context construction, governed delegation, and layered epistemic memory to support long-horizon scientific work. We formulate discovery under SCION as Target-conditioned Inverse Search and extend it to hidden-target settings through batch active search under finite experimental budgets. Applications in materials analysis, molecule design, and protein or antibody screening, together with experiments on scientific reading, idea generation, molecule generation, and antibody screening, show that SCION outperforms existing autonomous research-agent baselines, especially in decomposition, verification, refinement, and memory reuse. Overall, SCION shifts AI from isolated tools toward a coordinated operational layer for traceable and reusable scientific innovation.

Despite remarkable progress in machine translation, Vision Language Models (VLMs) struggle on historical manuscripts, a domain that stresses core Natural Language Processing (NLP) capabilities: low-resource transliteration, archaic vocabulary, and noisy input signals. We present a systematic framework for evaluating the full image-to-translation pipeline on medieval Latin manuscripts, a setting in which scribal shorthand, ligatures, and parchment degradation expose failure modes that are invisible in clean-text benchmarks. Benchmarking on the CATMuS Latin dataset reveals a specialization gap: domain-specific Optical Character Recognition (OCR) models reduce character error rate by up to 4.3$\times$ compared to general-purpose VLMs, despite operating at orders of magnitude fewer parameters. We introduce the Interpres-Parallel-Corpus (IPC), a novel dataset comprising 1,383 aligned manuscript image lines, transcriptions, and expert translations, the first of its kind for medieval Latin. Our experiments uncover a complexity paradox: the simplest pipeline, a specialized OCR model feeding directly into a VLM, outperforms all multi-component variants. Adding retrieval-augmented generation (RAG) or post-OCR correction introduces prompt saturation and error propagation that degrade aggregate translation quality. These findings offer both a new benchmark and practical guidance for deploying translation systems in low-resource historical settings.

While Large Language Models (LLMs) have achieved remarkable success in Text-to-SQL tasks, their deployment in real-world environments is hindered by latent reliability issues. Identifying these latent weaknesses is critical for building trustworthy database interfaces, yet current diagnostic approaches rely heavily on static, expert-defined rules, which lack the capability for systematic and automated exploration. To bridge this gap, we propose SAGE (Systematic Automated Guided Exploration), a novel framework designed to autonomously uncover latent failure patterns in LLM-based Text-to-SQL generation. Specifically, SAGE generates vulnerability hypotheses for given samples and references a continuously evolving Vulnerability Codex to design targeted perturbations, thereby iteratively verifying and documenting potential defects. Extensive experiments on state-of-the-art open-source LLMs demonstrate that SAGE uncovers a substantial number of failure cases, highlighting the significant fragility of current models. Furthermore, our analysis reveals that the Vulnerability Codex exhibits strong cross-model transferability, indicating that the discovered patterns represent generalized structural weaknesses. Finally, we explore SAGE's potential for remediation. Although preliminary, lightweight fine-tuning on the generated samples yields promising improvements, suggesting a scalable pathway for closing the reliability loop in future work.

We define Tiny Language Models (TLMs) as models below roughly 3B parameters that fit on mainstream consumer devices. We study how to adapt them for and use them on verifiable multiple-choice tasks. We compare three LoRA-based fine-tuning paradigms (label generation, gold only, and our discriminative classification head) on a unified setup across several Qwen3 models from 0.6B to 8B and five benchmarks: HellaSwag, WinoGrande, PIQA, SciQ and ARC-C. Classification-head fine-tuning reliably outperforms label generation (+2-3%) at the 0.6B and 1.7B scales. Further, TLMs fine-tuned using the discriminative method are competitive to zero-/few-shot GPT-3 (175B), PaLM (540B) and GPT-4. The performance we report for Qwen3-0.6B and Qwen3-1.7B are SOTA on HellaSwag, WinoGrande, and PIQA.

Measuring the extent to which emergent languages encode the visual content of their inputs is an open problem. We refer to this property as visual reflection: the extent to which emergent messages preserve information about their source images that can be recovered without appeal to the speaker-listener pair that produced them. Existing metrics measure it only indirectly, through proxies such as human-defined concept inventories, natural-language captions, structural distance correlations, or Referential Game accuracy, each of which can either miss visual content the message encodes or credit content it does not. We propose EmCom-Diffusion, an evaluation framework that measures visual reflection directly: it reconstructs each input image from its emergent message and compares the reconstruction with the original image itself, rather than with human-defined targets. Concretely, it finetunes a pretrained text-to-image diffusion model on (image, emergent-message) pairs and scores visual reflection as the perceptual similarity between the reconstructed and original images, operating generatively rather than discriminatively. Instantiating it on MS-COCO with a Referential Game, we validate the metric against random and fixed-token baselines under three pretrained visual encoders, and compare it against four existing metrics (CBM, supervised translation, TopSim, and R@1). EmCom-Diffusion captures visual content the other metrics miss.

We release a benchmark and failure-mode-aware evaluation framework for grounded QA under coordinated retrieval poisoning. The framework partitions reader outputs into four mutually exclusive categories (gold, hijack, abstention, drift), with instance-level paired clean-to-poison transition matrices and a Forced Exposure protocol isolating reader-side conflict resolution from retrieval variance. We introduce polymorphic sybil poisoning, a coordinated attack class in which $S$ lexically diverse passages jointly support an attacker-chosen target while evading lexical near-duplicate filters that fully detect monomorphic baselines (capturing the residual 14.2% with E5 cosine raises false-positive rate 9$\times$ on legitimate same-topic pairs). A monomorphic-polymorphic ablation under Forced Exposure isolates the diversity dimension and reveals a $+$18.8pp hijack amplification (95% paired bootstrap CI $[+15.4, +22.4]$, $B{=}5{,}000$): monomorphic copies register only 4.0% as hijack while polymorphic surface diversity recovers 22.8% – a 5.7$\times$ amplification of the ASR-visible attack channel. ASR alone treats every non-target output identically; under attack, abstention and drift together hold 47-66% of output mass, unmonitored by ASR+ACC, and two readers at nearly identical ASR (within 0.2pp) differ by 16.5pp on abstention and 17.2pp on drift – failure profiles invisible to ASR. We release the frozen benchmark (3{,}145 questions, 2{,}982 retained sybil groups; $S{=}6$ chosen to dominate top-10 retrieval slots, \S\ref{sec:setup}), the official four-way evaluator, paired-transition utilities, and the Forced Exposure harness across five readers (7B-120B), two retrievers, and two cross-validation datasets (TriviaQA, 2Wiki), under CC~BY-SA~4.0 (data) and MIT (software); release information in \S\ref{sec:release}.

Conversational agents are increasingly embedded in human collaborative work, yet they remain fundamentally passive and reactive: they respond to explicit user requests rather than proactively recognizing moments when a team would benefit from timely intervention as human collaborators often do. This reactive design substantially limits the use of agents as active participants in multi-user collaboration, where disagreements, ambiguous goals, forgotten constraints, underspecified plans, discussion loops, and imbalanced participation can gradually undermine group progress. To move agents from passive assistants toward active participants in multi-user collaboration, we introduce ProACT, a breakdown-aware agent framework grounded in theories of common ground, collaborative planning, and coordination work. ProACT observes the speaker-attributed conversation history, determines whether the current turn contains a collaboration breakdown requiring intervention, decides whether the agent should stay silent or speak, and, when speaking is needed, routes the case to a targeted collaboration skill. We further introduce the first multi-user collaboration benchmark for evaluating proactive agents across project planning, product design, research collaboration, logistics, education, and resource-constrained decision making. Across 3,244 turn-level examples and five LLM backbones, ProACT consistently improves collaborative appropriateness, non-interruptiveness, conciseness, and judged intervention quality over direct chat.

While current AI agents support increasingly long context windows, tool use, and skill execution for long-horizon tasks, they still require memory systems to effectively leverage historical experience. Existing memory frameworks typically rely on fixed storage, retrieval, and summarization mechanisms, which can be rigid across different tasks and often require manual tuning. To address this limitation, we propose SelfMem, a self-optimizing memory framework. Inspired by prior work on self-improving AI, we follow the principle of "teaching an agent to fish rather than giving it a fish." Instead of forcing the model to follow a predefined memory strategy or format, SelfMem provides an environment with memory tools and feedback signals that allow the agent to explore, evaluate, and refine its own memory strategy. Our results show that SelfMem consistently outperforms retrieval, compression, and agent-memory baselines on BEAM across conversation scales from 100K to 1M tokens. Compared with the strongest baseline, SelfMem improves the official score by 48.7%, 40.8%, and 41.9% at 100K, 500K, and 1M, respectively. Further question-type analysis shows broad robustness across diverse memory demands, and our optimization study shows that model-guided strategy refinement further improves performance.

Writing a literature review requires a deep understanding of the relationships among cited papers: how they build on, challenge, or offer alternative perspectives to one another. We present Graph-Reasoning Aided Survey Planning (GRASP), a framework combining LLM planning for related work generation with graph algorithms to extract key relationships among cited papers. Our two-layer graph structure consists of a Graph of Thoughts and an Argument-Counterargument Planning Network, representing the cited papers at different levels of granularity, and we apply topology-aware pruning via a Steiner tree to identify the core inter-paper relationships captured in our graph. Our citation analysis-based evaluation shows that GRASP generates related work sections (RWS) that closely match human-written targets in terms of the discourse roles, intents, and grouping of citations.