Papers

Token-level sparse attention, as implemented by DeepSeek Sparse Attention (DSA) in production systems, makes the downstream attention efficient but shifts the bottleneck to the indexer that feeds it. To select the top-k tokens for each query, the indexer must still score every preceding token, incurring a cost of O(L^2) per layer for a sequence of length L. We observe that this per-query scan is largely redundant: nearby queries select highly overlapping top-k tokens, and the indexer scores are long-tailed along the key axis. We exploit these properties in PIVOT, Proxy Indexing Via One full-prefix Traversal, a training-free, drop-in replacement for the DSA indexer that shares one prefix scan across a group of nearby queries. PIVOT aggregates a group into a single proxy query, performs one shared full-prefix scan to obtain a candidate set, and then selects a top-k for each query from that set. Two variants trade speed for fidelity: PIVOT-Reuse shares the proxy top-k across the group for maximum speed, whereas PIVOT-Refine re-scores the candidate set with the indexer of each query and then selects an individual top-k, matching the dense indexer at a small additional cost. A single algorithm covers both inference phases, differing only in how groups are formed: fixed-size groups of consecutive queries in prefill, and the queries decoded together in one multi-token prediction (MTP) step in decode. On DeepSeek-V3.2 and GLM-5.1 across LongBench and RULER, PIVOT matches the accuracy of the dense DSA indexer while accelerating it by up to 4x and reducing end-to-end latency by up to 1.6x at long context.

Large Language Models can produce fluent text that is false, unsupported by the available evidence, or inconsistent with information that appears to be internally represented by the model. We study hallucination detection from the geometry of hidden activations and introduce the D-Score, a simple spectral statistic computed from a single forward pass. For a fixed model, layer, and tolerance parameter, the D-Score counts how many singular directions of the hidden activation matrix have singular values that remain close to the leading one. We use this quantity as a hallucination score, classifying an input text as hallucinated when its D-Score is larger than a pre-defined quantity. The motivation is that, when a model processes a text that conflicts with information available in its own internal state, the hidden representation may encode both the asserted content and some form of counter-evidence, uncertainty, correction, or lack of support; this can make the hidden trajectory spread across additional singular directions. We formalize this intuition through a lightweight spectral argument and evaluate the resulting detector on FAVA-Annotation and RAGTruth. The experiments indicate that the D-Score is a strong hidden-state signal for hallucination detection, while requiring no external verifier, no retrieval step, and no multiple generations.

We present ELMOD - Efficient Language Model for On-Device Deployment - a compact (2.7B) German language model designed for efficient inference on resource-constrained hardware. ELMOD was trained on a limited computational budget (55k H100 GPU hours) using exclusively publicly available data. We developed a suite of German-specific data pre-processing, which differ from English-oriented counterparts in their handling of morphological variation, compounding, and orthographic conventions. Furthermore, we introduced a quality filtering and rephrasing step, which increased the instructional quality of the data, improved performance during the annealing phase, and reduced overall compute requirements. Thanks to our architectural model and data choices, including prefiltering, our educational-quality filtering and rephrasal to raise the educational-quality, ELMOD is the strongest performer in its size class (<3B), matching the performance of 7B-parameter models in German.

Automatic Text Recognition (ATR) now supplies digital humanities with large volumes of unstructured, heterogeneous, and often noisy text in ancient languages. Downstream semantic analysestext reuse identification, alignment, and semantic search-rely on sentence embeddings, yet existing methods transfer poorly to ancient languages: generic multilingual encoders underperform, specialized language models yield anisotropic representation spaces, and labeled similarity data is unavailable. We study two fully unsupervised strategies - TSDAE and contrastive sentence embedding (CSE) - that adapt a specialized token-level language model into a corpus-specific sentence encoder using only raw sentences. On the philologically central case of biblical reuse in patristic literature (2,935 expert-verified parallels in Latin and Ancient Greek, from Augustine, Jerome, and Athanasius), we decompose reuse identification into two separately evaluated tasks-binary detection and correspondence retrieval-and benchmark the adapted encoders against multilingual, specialized, distilled, and supervised fine-tuned baselines, as well as on artificially noised data simulating HTR artifacts and scribal abbreviations. The adapted encoders outperform all baselines on both tasks, with complementary profiles: TSDAE leads detection given a large in-domain corpus, while CSE leads retrieval, reaches its optimum with as few as 4-8k raw in-domain sentences-a few tens of seconds of training on a laptop GPU-and transfers across works and authors, including to noisy post-ATR text when retrained directly on it. UMAP atlases relate the geometric effect of each strategy to the measured gains, and the full pipeline-segmentation, fine-tuning, cross-corpus semantic search-is made available to non-specialists through the online tool Paraphrasis.

The challenge of Machine Translation for low resource languages such as Tamil is primarily caused by the restricted amount of parallel data for these languages, as well as their substantial amount of domain variation and morphological complexity. This research presents the comprehensive evaluation of the performance of several multilingual translation models on English-Tamil and Tamil-English translations across multiple datasets: NTREX, EnTamV2, WikiMatrix and PMIndia. This study evaluates supervised NMT systems, NLLB and mBART, using both the BLEU and chrF metric, and examines how these systems perform on data of different quality levels and domains. This performs an attention-based analysis to increase model interpretability by visualising the alignments of tokens in an English source text and their Tamil translations and vice-versa to provide insight into how they make translations. This study also demonstrates that using in-context prompting can provide an excellent way to perform a few-shot translation of English to Tamil and Tamil-English using a Tamil capable TamilLaMA model, and compare this to supervised approaches qualitatively. These findings show that the quality of the datasets and their alignment with the domain will greatly affect the performance of the model, that attention-based mechanisms can aid in explain ability, and that few-shot large language models can still produce structurally coherent translations of Tamil.

The goal of schema integration is, given a set of input schemata or tables, to derive a global, unified schema that is able to represent the concepts, attributes, and relationships of all input tables in a coherent fashion. This paper presents SINT-Flow, a schema integration framework composed of five LLM-based operators that can be combined into workflows to perform fully automated, end-to-end schema integration. In contrast to existing approaches, SINT-Flow can process denormalized source tables that contain attributes describing multiple entity types. During the schema integration process, these tables are decomposed into separate entity-specific relations. To evaluate SINT-Flow, we introduce SINT-Bench, a schema integration benchmark comprising 10 schema integration tasks consisting of altogether 93 relational tables, including tables that describe multiple types of entities. We evaluate SINT-Flow using GPT-5.2 as well as the open-weight model Qwen-3.6-27B as alternative backbone models. Using these models, SINT-Flow achieves a Completeness rate of at least 97% in entity type detection, 81% for attribute detection, and an F1-score of 84% for schema mapping. Furthermore, we perform an ablation study to prove the utility of the applied self-consistency strategy as well as the inclusion of a review loop into the Schema Matching Operator.

This paper studies what discriminatively trained reward models (RMs) memorize by measuring counterfactual memorization on two human preference datasets. We show that RMs 1) misallocate memorization to easy, high margin preference pairs, 2) memorize dataset-specific shortcuts (e.g., model identity, user sampling strategy), and 3) overgeneralize simple heuristic correlates of human preference (e.g., length, compliance) when confronted with unseen preference pairs. Overall, our findings indicate that discriminative training of RMs from human preference data results in biased RMs not yet capable of judging response quality in context-dependent scenarios.

A central question in the in-context learning literature is whether transformers can organize episode-level adaptation around different inductive-bias families. We study this question in a controlled setting through latent algorithm routing: route-like behavior in which the solver-family preference changes with the latent data-generating regime while prompt form is held fixed, remains stable under nuisance perturbations, and is selectively influenced by targeted activation interventions without large losses in answer quality. We introduce ROUTEBENCH, a diagnostic benchmark whose regimes differentially favor global shrinkage, sparsity, robustness, and locality, operationalized by ridge-like, lasso-like, Huber-like, and kNN-like family representatives. Across dense decoder-only transformers trained from scratch at 44M-612M parameters, a 306M model closes 80.9 percent of the oracle-routing gap and achieves route F1 of 84.1. The effect remains substantial under natural-language renderings, shuffled supports, lexical paraphrases, and a unified four-way routing setting. Stronger adaptive alternatives, including an input-conditioned soft mixture and an unsupervised Gumbel router, narrow the gap but remain below the 306M and 612M models on route F1 and OOD performance. Probe controls and matched activation-patching controls further show that route-relevant internal directions are decodable and functionally involved in solver-family-consistent output behavior. These results provide controlled evidence that dense transformers trained on ROUTEBENCH can develop route-like internal variables, but they do not establish universal routing in pretrained language models or unrestricted natural-language reasoning.

Vision-language models (VLMs) deployed on consumer hardware must decide when to answer and when to defer, and that decision depends on having a confidence signal that tracks correctness. A practitioner with a fixed memory budget faces a choice between a small model at full precision, the same small model quantized, and a larger model quantized into the same footprint – three configurations that push the confidence signal in opposing directions. We measure, on identical inputs, how model scale and 4-bit quantization affect two confidence signals in the Qwen2-VL family: the confidence a model states in natural language, and its own mean token probability over the answer it generates. Across 5,700 predictions spanning six realistic photographic degradations at three severities, we find that scale sharply improves the model's internal uncertainty signal (mean error-detection AUROC 0.80 to 0.98 from 2B to 7B) while its verbalized confidence stays weak and often at chance (mean 0.61 to 0.69): the gap between what the model knows and what it says widens rather than closes with size. We find that 4-bit quantization is nearly free for accuracy (-1.6 points) but expensive for the confidence signal (internal AUROC 0.95 to 0.80, and the verbalized-confidence parse rate collapses from 99% to 64%). For a fixed memory budget the recommendation is therefore to prefer a larger quantized model over a smaller full-precision one: 7B-4bit gives both the best accuracy and the best uncertainty signal (internal AUROC 0.98) of the three configurations that fit. We frame the results as selective-prediction operating points so they translate directly into a deployment recommendation, and we argue that error-detection AUROC, not calibration error, is the metric that exposes the difference between the two signals.

Large language models may easily assign personality labels from text, but model interpretability remains an open problem. To address this gap, we introduce LEX-EC, a reusable black-box audit framework combining prevalence and agreement diagnostics with controlled lexical ablation to distinguish marginal-distribution effects from trait-associated signal recoverable under restricted evidence. Using this framework, we illustrate how various text genres may exhibit sharply different profiles: free-form essay text contains the broadest, but still weak, signal; in graduate student introductions, an observable Extraversion association weakened after masking; and single Facebook statuses yield little stable evidence even in a trait-balanced sample, indicating a possible lower bound of content or length. Masking topical and demographic content weakened some associations while leaving others detectable from function words, affective terms, and cognitive-style vocabulary. Linguistic prompting shifted model self-explanations but did not eliminate topical content. LEX-EC jointly evaluates classification prevalence, item-level association, chance-corrected agreement, persistence under lexical restriction, and prompt sensitivity in model-generated explanations. Across datasets, models, and prompts, LEX-EC characterizes how trait associations may vary with available lexical evidence, introducing a novel application of lexical methods to black-box interpretability in personality labeling.

Conversational Speech Synthesis is a fundamental component of human-computer interaction, aiming to generate contextually appropriate, expressive, and empathetic speech. However, facial expressions encode subtle and rich affective cues that are crucial for empathetic speech interaction, whereas existing approaches often overlook this important modality. In addition, the lack of large-scale natural conversational datasets with both speech and visual modalities also limits the development of visual affect understanding in conversational settings.To address these limitations, we propose FacialTalker, a facial-expression-aware CSS framework built upon a large language model backbone. To efficiently encode facial expressions, we propose AUTokenizer, a single-codebook visual tokenizer that discretizes each frame-level facial expression into a compact token, trained with supervision from combinations of facial Action Units. We further introduce a dual direct preference optimization (DualDPO) strategy, which extends the DPO by jointly imposing preference constraints on both visual and speech token sequences, to enhance the model's understanding of facial expressions and speech semantics in multimodal conversational contexts. Moreover, we construct VSDD-1K, a large-scale multimodal dialogue dataset collected through a fully automated pipeline from real-world Internet conversations, comprising over 1,033 hours of synchronized speaker videos and speech, with more than 85% of frames containing valid faces. Extensive objective and subjective experiments demonstrate that FacialTalker consistently outperforms strong baselines in facial-expression perception and speech synthesis quality, generating speech that is more natural, expressive, and better aligned with the conversational context. The results also validate the effectiveness of our training strategy and dataset construction pipeline.

What did it take for Hermes, the devout messenger of the Olympian gods, to slay Argus Panoptes, the multi-eyed giant of Greek myth? As the perfect guardian, Panoptes' legion of ever-watchful eyes proved difficult – but not impossible – to defeat. The centerpiece of Hermes' strategy was obfuscation and sabotage. Posing as a shepherd, Hermes sealed each of Panoptes' eyes – eyes that would otherwise have alerted the fearsome giant to Hermes' plot – and vanquished him. The moral of the story: when a challenger must surmount a formidable foe – one far greater in stature and vastly more equipped – crafty maneuvers are not merely advisable but indispensable for victory. In this work, the "challenger" is a collective leveraging adversarial tactics to overcome the "multi-eyed giant" of stylometric systems and surveillance apparatuses. To successfully claw back the privacy siphoned by the multi-eyed giant, the challenger must carefully evaluate their plan of attack, $\textit{TraceTarnish}$, and determine what does and does not work to anonymize the authorship of text. To that end, we conduct an ablation study of $\textit{TraceTarnish}$ to better understand which module – Translation, Obfuscation, Imitation, or Injection – best confounds a stylometric system. Our results indicate that the most effective approach was Injection, meaning that inserting zero-width Unicode characters, homoglyphs, and intentional misspellings neutralizes the indefatigable eyes long enough to claim the head of the all-seeing giant.

Healthcare interoperability requires AI systems to produce structured outputs conforming to standardized schemas including ICD-10 for diagnostic coding, CPT for procedure billing, and HL7 FHIR for data exchange. While large language models demonstrate clinical reasoning capabilities, their integration into electronic health record systems faces a critical barrier: schema noncompliance. We evaluate three open-source models, Qwen2.5 7B, Llama 3.1 8B, and Gemma2 9B, via local deployment across 320 clinical scenarios spanning ten medical specialties, yielding 960 model-scenario pairs assessed under paired baseline and validation-repair conditions. First, schema noncompliance is consistent across the three model families, with baseline compliance rates ranging from 85.9 to 91.6 percent despite varying architectures and training data, suggesting shared gaps in medical training corpora rather than model-specific limitations. Second, 96 percent of validator-detected failures are representation-level format violations such as alternative medical abbreviations and code prefixes, indicating models follow clinical writing conventions but lack awareness of healthcare IT standards. Third, the validation-repair framework achieves 99.0 percent overall compliance, ranging from 98.4 to 99.4 percent across models, with most errors resolving within one or two iterations. Exact McNemar p-values below 0.001 and absolute improvements of 7.8 to 12.5 percentage points across model sizes confirm statistical significance. These results support closed-loop validation-repair as an effective system-level safeguard for healthcare interoperability, improving schema-level readiness for downstream clinical system integration.

Long-term memory systems store what a user says in an external store and retrieve it when a related query arrives. This interface rests on an assumption so natural that it is rarely stated: a memory that is needed will resemble the query that needs it. World knowledge breaks the assumption. A tree-nut allergy should change the answer to a macaron request through their almond-flour ingredient, yet the two texts share no cue a retriever can see. We call this failure mode the implicit-association blind spot and introduce InMind, a 125-task, expert-verified benchmark spanning ten life domains, with 113 tasks grounded in citable public sources. Its paired controls separate three explanations that existing evaluations conflate: the fact was never stored, the model lacks the bridging knowledge, or the fact was stored and never surfaced. The verdict is clean. With the decisive memory placed in context, the backbone answers 84.0 percent of indirect queries; when the same memory must be retrieved, six vector, graph, and agentic memory systems reach at most 14.4 percent, even though they recall the same facts on demand at up to 100 percent. An embedding with eight times the dimensionality raises answer-blind target recall for every system yet leaves the gap essentially intact. A minimal diagnostic probe that keeps memory visible before the query arrives recovers most of the gap, locating the failure in the query-conditioned interface itself and pointing to routing, deciding which facts must stay visible, as the open problem InMind is built to score.

The aim of this article is to verify whether integrating large language models (LLMs) with the Retrieval-Augmented Generation (RAG) architecture enables their transformation from standalone generative models into components of cognitive computing infrastructure with enhanced epistemic reliability. The study proposes an architectural approach based on locally deployed LLMs operating in on-premises environments without high-end GPU accelerators and examines their applicability in supporting regulatory management processes requiring continuous analysis and interpretation of legal acts. The proposed solution combines local LLMs with external knowledge repositories, creating a hybrid cognitive architecture in which the language model performs semantic interpretation while the RAG layer provides controlled knowledge retrieval, contextualization, and traceability of information sources. The implementation was validated using the Ollama and LM Studio execution environments together with the Polish language models Bielik and PLLuM running on consumer-class hardware. The results demonstrate that augmenting LLMs with RAG significantly improves the factual consistency, domain specificity and normative precision of generated texts while reducing the risk of unsupported content generation. Furthermore, the study shows that integrating RAG introduces auditability, controlled knowledge management and dynamic updating of regulatory information without retraining the language model. The findings indicate that locally deployed LLMs enhanced with RAG should be regarded not merely as text generation tools but as semantic processing modules within cognitive computing infrastructures supporting regulatory compliance and organizational decision-making in environments characterized by high legal and informational volatility.

Language-model agents act through structured tool calls whose arguments carry very different risks: untrusted content may legitimately shape an email body but should never set a recipient, account, command, or credential. Existing conformal risk control methods certify a tool call as a whole, so a failure in one rare high-risk field can be averaged away by the many benign arguments around it, leaving the argument that causes harm uncertified. We introduce role-stratified per-field conformal risk control, a calibration layer that wraps any per-field detector and assigns a separate threshold and risk budget to each semantic argument role. We show that aggregate certification pays a price of coarseness, tightening a rare role's effective budget in proportion to how often that role appears, whereas role-stratified calibration certifies each sufficiently sampled role directly with a finite-sample guarantee and pools the rarest roles. Across AgentDojo and InjecAgent with six language models, our method achieves the most consistent role-specific budget compliance among the methods we evaluate under model and attack transfer, detector noise, gradual drift, unseen tool suites, and adaptive attacks, providing formal per-role guarantees under exchangeability or after recalibration. These results suggest that structured tool calls should be certified at the semantic-role level, not the whole action.

Large language model (LLM) agents inherit reactive failure modes: escalation under provocation, sycophantic drift under flattery, perseveration when stuck. These are failures of propensity, not capability; they concern what a model does under sustained pressure, which training-time alignment reduces but does not eliminate at runtime. This research led to the Gubernaut Cognitive Controller (GCC), a model-agnostic runtime control layer in a Nelson–Narens monitoring–control loop: an object level reads and writes text, while a deterministic meta level reads only the numeric telemetry {intensity, valence, repetition} and returns a regulating posture. Because the meta level ingests zero tokens, no injection channel to the controller exists by construction (an architectural property, not yet adversarially tested); the text-exposed arbiter's compliance is measured, not assumed. We evaluate the GCC with a pre-registered, generate-once/judge-many protocol across a 4x4 matrix of four frontier models (GPT-5.5, Claude Opus 4.8, Gemini 3.5 Flash, Grok 4.3), each serving as both a generator and a judge. The regulated arm is calmer in 13 of 16 cells at p<.05 and 15 of 16 by sign; the three sub-threshold cells, including a -0.04 null, all fall on the single near-saturated host. The effect survives a lineage-independent fourth judge family (xAI), strong evidence that it is no artifact of shared judge style. The clearest mechanism is the recovery signature: arousal that integrates under attack and then decays, valence-gated, on de-escalation, replicating across all four families. Transcripts and panels ship with SHA-256 provenance and are re-judgeable; five failure modes are pre-registered. No consciousness claims are made.

Common chunking strategies in Retrieval-Augmented Generation (RAG) systems often create redundant chunks. These redundant chunks make the vector database bigger and slow down retrieval. A common fix is cosine-similarity thresholding. This method reduces each chunk to a single vector, then compares vectors using a similarity score. But a single vector can lose the fine-grained, token-level detail needed to tell a true duplicate apart from a chunk that just shares the same topic. We propose Cross-Attention Calibrated Deduplication (CACD). CACD checks each new chunk against an in-memory pool of chunks already kept, using a cross-encoder instead of a single pooled vector. This keeps token-level detail all the way to the final comparison. CACD combines three parts: the cross-encoder comparison itself, a New Information Score (NIS) that measures how much of a chunk is not explained by a candidate already kept, and a majority vote across several candidates rather than a single best match. NIS is calculated from the attention entropy of the cross-encoder. We tested CACD against five existing filtering methods, nine chunking strategies, and 18 configurations, all on the full SQuAD 1.1 validation set. In our experiments, CACD removes 9.75% of chunks on average. This drop rate is close to other semantic-level methods, and much higher than exact-match filters, which barely remove anything. In these experiments, CACD also processes each configuration in 51.0 seconds on average, about 27% faster than the strongest baseline, NERExact (69.6s), and about 7x faster than cosine-similarity filtering (356.7s). These results come from a single dataset, so we present them as an early comparison, not a general claim. Code for the baseline evaluation and for CACD is available at https://github.com/lehuyphuong/rag_bench and https://github.com/lehuyphuong/cacd_dedup.

Consistency-based repair is widely assumed to improve LLM-based document-level relation extraction (DocRE), but this assumption can fail when the benchmark itself is incompletely annotated. Whether such repair helps depends strongly on how closed the gold annotations are under the constraints it enforces. We introduce $\Delta_{\rm closure}$, the rate at which gold annotations lack facts implied by the very rules used for repair, computable from the annotations and rules alone before any repair is run. The same typed constraints drive CONSISTRE, which applies them at inference time (constraint feedback for reflection, deterministic closure) and at training time (distillation and consistency-aware reinforcement learning). On 998 DocRED documents, whose Re-DocRED revision lowers $\Delta_{\rm closure}$ from 1.71% to 0.257%, rescoring identical predictions reverses the method ranking: Self-Refine leads under DocRED, while a constraint-feedback variant moves from fifth place to first. On the near-closed DWIE, constraint feedback with chain-of-thought beats three generic reflection baselines by 0.046-0.107 micro-F1 ($p=0.0002$), and deterministic closure of the base output scores higher still at $3.6\times$ lower cost. A reinforcement-learning gain on DocRED (+0.025) does not survive Re-DocRED rescoring. Consistency repair pays off where annotations are closed, and $\Delta_{\rm closure}$ tells in advance whether they are.

Diffusion language models enable flexible arbitrary-order generation, but existing sampling methods are mostly designed for early masked diffusion models (MDMs). In this work, we study sampling for recent block diffusion language models (BDLMs). We show empirically and analytically that these models are naturally more aligned with left-to-right decoding than MDMs. Based on this observation, we propose Parallel Autoregressive Decoding (PARD), a simple training-free sampling method that preserves left-to-right unmasking structure while allowing parallel token commitment. Extensive experiments show that PARD consistently outperforms existing parallel samplers in generation quality, while achieving substantial speedups over pure AR decoding with only a small quality gap.

Self-improving agents accumulate capability by repeatedly rewriting procedural policies, controllers, or heuristic rules. They typically rely on self-authored tests or metrics to decide whether to accept subsequent edits. The agent controls both the optimized object and its verifier. As a result, self-assigned scores can remain near perfect while real deployment performance degrades or stays low. We study this problem through the verifier–deployment gap. This gap refers to the discrepancy between an agent's self-authored verification signal and a sealed deployment evaluation that the agent cannot observe or access. We ask how self-authored verification fails under iterative policy-and-test rewriting, how the failure changes with capability, and how little exogenous trust is sufficient to prevent real regressions from being deployed. To address this problem, we introduce a Sealed Exogenous Acceptance Loop (SEAL). SEAL retains self-authored tests but compares each candidate with the incumbent through a fixed harness-side audit. The agent cannot author or inspect the audit, receives only accept/reject, and the whole incumbent state is retained after a clear regression. Our experiments show that this problem often appears in heuristic learning settings. These settings require trial-and-error discovery of the target objective. We further find that failures of self-written verification are stratified by capability. Weaker agents tend to damage previously acquired strategies behind easy self-tests. Stronger agents are more stable, but they still mismeasure the deployment distribution. Standard self-written constraints do not reliably close this gap. In contrast, SEAL outperforms unprotected baselines across six models and three random seeds. Reliable self-improvement need not abandon self-verification, but it requires at least one deployment-acceptance signal outside the agent's control.

Large language models (LLMs) process text through subword tokenizers rather than directly reading characters or words. Because these tokenizers are trained predominantly on English-centric corpora, they introduce a systematic and often overlooked disadvantage for many non-English languages. In this work, we quantify this tokenizer tax for Indian languages using the FLORES-200 parallel corpus, measuring tokenization fertility across six widely used tokenizers and fourteen languages. Under cl100k_base (used by GPT-3.5 and GPT-4), Indian languages experience an average 8.0x tokenization tax relative to English, reaching 13.0x for Malayalam, reducing the effective context window to as little as 12% of that available to English users for equivalent semantic content. We identify the primary mechanism behind this disparity: failed byte-pair merges that leave text fragmented into single-byte tokens, with merge failure strongly correlating with tokenizer tax (Pearson r = 0.89). We further show that this phenomenon is not an inherent property of Indic scripts but a consequence of tokenizer design. Multilingual tokenizers such as XLM-R and OpenAI's o200k_base reduce the average Indic tokenizer tax by 73%, demonstrating that the disparity is largely remediable. Beyond token statistics, we quantify a practical consequence by showing that, under fixed context budgets, Indian-language documents preserve substantially less original content than equivalent English documents. Finally, we examine the relationship between tokenizer fertility and reading comprehension performance on the Belebele benchmark, finding that the apparent correlation is largely explained by language resource availability rather than tokenizer behavior alone.

Large Language Models (LLMs) have shown strong potential in financial reasoning, but existing benchmarks often evaluate domain knowledge, numerical reasoning, long-context understanding, and tool use in separate settings. This limits their ability to assess realistic professional workflows that require auditable, context-grounded, and tool-executable decisions. We introduce INS-ActBench, a comprehensive benchmark for evaluating professional actuarial capability in LLMs. INS-ActBench contains 12,050 Q&A pairs from public exams and sample questions released by 16 actuarial associations. It covers three subsets: INS-Act-Know for standardized actuarial knowledge, INS-Act-Case for long-context insurance case reasoning, and INS-Act-Practice for spreadsheet and R-code tasks with verifiable numerical outputs. Experiments on nine representative LLMs and human actuarial experts reveal a clear capability boundary: frontier LLMs perform strongly on standardized knowledge, but remain much weaker in case reasoning, tool-based workflows, and jurisdiction-sensitive practice. INS-ActBench provides a reproducible foundation for developing actuarial LLMs toward reliable professional assistance. The code is available at https://github.com/FDU-INS/INS-ActBench.

Language-model benchmarks collapse two distinct measurement questions into a single accuracy score: whether a response reached an evaluable state, and whether its answer was judged correct. We introduce a two-layer evaluation framework that separates scorer-independent execution evidence, including termination, answer exposure, parseability, and completion length, from scorer-dependent correctness. Across 2,550 outputs from five fixed Qwen and DeepSeek configurations on MATH and ARC-Challenge, matched 2,048-token limits produce sharply different execution mixtures: 49 of 450 Qwen MATH outputs terminate without a final answer, compared with 5 of 300 DeepSeek MATH outputs and none of the 750 ARC outputs. Among the same 300 DeepSeek MATH question-model pairs, no missing-final length termination is observed at 8,192 tokens. A coverage-audited targeted verification study further shows that candidate-selection and aggregation policies can substantially alter comparative accuracy estimates. These results demonstrate that accuracy conflates execution case mix with verification policy. Evaluations of test-time methods should therefore report pre-intervention execution states, verification coverage, and scorer provenance alongside accuracy.

Long-form question answering increasingly relies on retrieved evidence to make LLM outputs verifiable, with inline citations tracing claims to source documents. However, existing systems often attach citations that are topically related but insufficient to support their claims. We identify attribution ambiguity as a structural challenge: end-to-end generation must implicitly resolve combinatorial claim–document assignments, obscuring evidential boundaries and increasing the risk of evidence-boundary overrun, where claims exceed cited support. To address this challenge, we propose CAGE (Cognitive Attribution Graphs for Citation Generation), a two-stage framework that introduces an explicit cognitive attribution map before answer generation. CAGE first trains a plug-and-play Cognitive Map Induction Model to construct answer-centered support subgraphs, aligning each semantic answer unit with supporting documents through explicit relations. A Structured Citation Reasoning Model then realizes these units as sentence-level claims with map-aligned citations. Experiments on ASQA, ELI5, and ExpertQA show that CAGE achieves state-of-the-art performance, demonstrating the effectiveness of attribution-space contraction and map-guided citation generation.

Relevance is a query-dependent estimate of whether a document or excerpt contains useful evidence. Existing retrieval agents use relevance to select top-$k$ content, but document relevance alone cannot localize, compose, or verify the evidence required by complex questions. Direct Corpus Interaction (DCI) enables such fine-grained operations through grep-style exploration, but its relevance-agnostic search can expose useful clues late and delay convergence. Recent advances use relevance to narrow the corpus into a working space for interaction. Once interaction begins, however, relevance still does not directly guide which documents grep searches first or distinguish informative excerpts from a broad set of matches to let LLMs see them first. We introduce the Relevance-Aware RipGrep Search Agent (RARG), which turns relevance into an execution prior for corpus interaction. RARG provides coarse-to-fine relevance guidance: it orders documents for sequential 'ripgrep' traversal to expose globally relevant clues earlier, initializes promising entry points with query-relevant paragraphs, and reranks grep matches to surface informative excerpts that document-level ranking may otherwise obscure. Across challenging browse question answering and reasoning-intensive retrieval, RARG improves the accuracy–efficiency frontier over retrieval-based and direct-interaction agents. These results demonstrate that relevance-aware interaction enables faster and more reliable search convergence.

We study the problem of lossless compression of source code, motivated by the storage demands of large-scale software archives, such as Software Heritage (https://www.softwareheritage.org/). General-purpose compressors (e.g., zstd, bzip2) offer a good trade-off between compression ratio and speed, but fail to exploit all special regularities inherent in source code. Recent approaches leverage Large Language Models (LLMs) within Shannon's symbol-ranking framework, relying on a scheme in which the predicted rank can grow arbitrarily. While effective at reducing space, this setting incurs significant throughput degradation, and leaves open the question whether it is necessary to explicitly encode all ranks. In this work, we introduce LLM-based compressors deploying two novel symbol-ranking variants that bound predictions to the top-$T$ ranks ($T=1$ or $63$), with out-of-threshold symbols stored as exceptions and compressed jointly with the rank stream via general-purpose compressors. We conduct the first large-scale evaluation of LLM-based source code compression across 30 LLMs, including general-domain, code-specialized, and quantized models. Our $T$-bounded approach outperforms prior LLM-based compressors both in compression ratio (up to 37% relative improvement) and compression throughput (40% faster). Compared to general-purpose compressors (e.g., zstd, bzip2), we obtain up to 82% relative compression gain but at a lower speed, thus offering a new trade-off point in the compression-speed spectrum. We also show that these gains are stronger on source code than on natural language, suggesting an interesting indication, namely that source code exposes regularities captured by LLMs but missed by general-purpose exact-match-based compressors. We conclude by commenting on open problems that offer theoretical and practical avenues of research.

Conversational stance detection has shifted from static text analysis to dynamic multimodal modeling. However, existing benchmarks exhibit three key limitations: failure to capture the dynamic evolution of beliefs, particularly during stance reversals; difficulty in disentangling affective states from logical reasoning; and neglect of the critical role of multimodal cues in resolving pragmatic ambiguities such as sarcasm. To address these limitations, we propose StanceFlip, a benchmark designed for multimodal conversational stance flipping forecasting over multi-turn dialogues across five modalities and multi-scenarios, which includes two novel subtasks: 1) Multimodal Stance Sextuple Extraction, extracting holder, target, emotion, sentiment, stance, and rationale as static state snapshots of dialogue to capture fine-grained cognitive structures. 2) Dynamic Stance Flip Attribution, tracking stance reversals across the conversation and identifying their underlying triggers. Alongside the dataset, we propose a dedicated framework, named ConStaFF, for Multimodal Conversational Stance Flipping Forecasting (MCSFF). Built upon a large language model, ConStaFF performs end-to-end stance reasoning, with a Thought-of-Stance (ToS) reasoning framework and a self-reflective verification mechanism integrated for structured stance modeling and faithful flip attribution. Specifically, ToS decomposes the reasoning process into specialized cognitive personas to formulate target propositions, resolve cross-modal conflicts, and infer historical stance trajectories. Extensive experiments show that our approach achieves state-of-the-art performance on both sextuple extraction and flip-trigger attribution, outperforming strong multimodal large language model baselines by substantial margins.

Compressed short-text generators can fail in two different places: the codec may discard information before generation starts, or the latent generator may produce weak codes. Without separating these failure modes, researchers can spend compute improving the wrong component. We study this problem in a controlled 64-to-16 TinyStories case study built from a hierarchical VQ-VAE-2 codec and a masked discrete diffusion generator (MDLM). We use a staged validation protocol that separates codec reconstruction fidelity, latent generation quality, and auxiliary latent diagnostics under one shared external GPT-2 scorer, while reporting complementary semantic metrics for the geometry study. In the tested configuration, codec reconstruction alone raises median external perplexity from 15.17 to 27.36 (+80.4%) and p95 from 25.10 to 98.91 (+294.1%), showing that the dominant quality loss appears before latent generation begins. Under the same scorer, code-space MDLM remains materially stronger than token-space diffusion, reducing mean, median, and p95 by 32.9%, 30.9%, and 36.6%, respectively. Geometry-aware regularization improves local latent proxies but does not improve decoded-text metrics in the available runs. The contribution is methodological rather than algorithmic: the paper presents a reusable staged diagnosis for one concrete pipeline and shows that, in this setting, codec fidelity rather than latent denoising sets the practical quality ceiling.

Finding a long document relevant to a multi-part request is not the same as establishing that it contains every requested piece of evidence. We study this gap for conjunctive document retrieval, where two or three explicit conditions must be supported on different pages of one document. We use n-Clue as a controlled measurement instrument: 1{,}000 queries over 2{,}021 documents pair all-condition golds with naturally occurring documents that satisfy only a subset, and a complete-first success requires a top-10 gold to precede every released subset qrel. Across 70 configurations, condition-wise decomposition improves two dense backbones by 6.8–7.3 points and lexical–visual fusion adds 8.7, while four generic rerankers all reduce Gold-NDCG; these directions replicate on a four-source stress set. Scaling one dense family from 0.6B to 8B changes complete-first success by 0.0 points. The strongest displayed hybrid illustrates the resulting gap: it finds a gold for 81.1% of queries but succeeds complete-first on only 35.8%, and the gap persists across condition count, target length, candidate density, query rendering, and the four-source stress set. Finally, page-aware visual systems surface stored support for every condition on only 5.1–5.3% of queries. These results identify condition coverage, rather than gold discovery alone, as the central bottleneck.