Papers

Video multimodal large language models have advanced significantly, yet fine-grained motion-temporal understanding remains fragile. The core bottleneck is not only sparse frame sampling, but also the lack of a complete temporal modeling pipeline for explicitly representing frame-to-frame change, enabling appearance-motion interaction, and optimizing temporal direction sensitivity. We propose COMET, a temporally grounded framework that systematically strengthens video MLLMs through explicit temporal representation, appearance-motion fusion, and direction-aware optimization. Architecturally, COMET introduces a temporal motion branch built on Taylor frame differences and injects its motion evidence into the appearance stream via temporal attention bias-enhanced cross-attention. For optimization, COMET combines temporal prior distillation with a forward-reverse TC-GRPO stage that turns temporal order into a direct learning signal and strengthens the model's use of directional motion patterns encoded by the temporal motion branch. The method achieves consistent overall improvements with a pronounced motion-temporal bias: on Qwen3-VL-8B, action-centric tasks (STAR, SSv2) improve by 4.9% on average, temporal reasoning tasks (NExT-QA, CLEVRER, LLaVA-178K) by 2.1% over BL-GRPO, while static perception tasks (PerceptionTest) remain on par. The same gain pattern also transfers to InternVL2.5-8B, indicating that COMET generalizes across model families.

Word alignment has traditionally been studied between sentences, but many cross-lingual tasks increasingly require correspondences across full documents. While recent multilingual embedding models can encode long inputs, we show that applying algorithms designed for sentences directly to documents leads to performance degradation. To address this, we introduce CTFAlign, a lightweight, training-free approach for document-level word alignment. CTFAlign applies a coarse-to-fine refinement strategy that restricts the alignment search space to semantically similar regions. Additionally, we introduce MDPAlign, a simpler alternative that constrains alignments by position with a main diagonal prior. Both approaches operate directly on full documents without relying on sentence segmentation or sentence alignment. We evaluate these methods across six language pairs varying in typological distance, resourcedness, and document length. Averaged over three models, CTFAlign reduces word alignment error rate from 0.412 to 0.326. These gains transfer downstream, leading to improvements in document-level translation coverage evaluation and recognition of semantic differences. We release CTFAlign as a Python package and make the code and data to reproduce our experiments publicly available.

Real-world fault analysis in 5G and emerging 6G networks demands domain expertise to analyze free-text diagnostics, including root-cause explanations and recommended actions. LLMs have emerged as a promising approach to automating this, yet whether lightweight, edge-deployable models are capable of performing in-depth free-text diagnostics remains an open question. While existing benchmarks rely on restrictive MCQs with fixed answer keys, this paper evaluates 5G domain understanding and fault analysis in a free-text generation format. Transitioning to this paradigm requires evaluating lightweight, edge-deployable AI models on open-ended diagnostic reasoning, alongside a dependable framework to validate these text outputs at scale. To address this we evaluate three lightweight LLMs, Claude-Haiku-4.5, GPT-5.4-Mini, and Gemini-3.1-Flash-Lite, on free-text 5G domain knowledge and fault-analysis tasks across three benchmarks, TeleQNA ORAN FT, 5G-Faults FT, and TeleInter FT. Three independent frontier judges score outputs, and pairwise inter-judge agreement is measured as an empirical test of the LLM-as-Judge methodology. All three models reach at least 90% accuracy on fault diagnosis, while zero-shot recall of 3GPP and O-RAN specifications remains the critical gap, with all models scoring below 60%. Mean inter-judge agreement is at least 0.90 across all runs, indicating that multi-judge LLM scoring produces consistent, reproducible grades for open-ended telecom responses. Operationally, Gemini-3.1-Flash-Lite offers the best efficiency trade-off, combining competitive accuracy with the lowest inference cost and latency, making it the most suitable candidate for production telecom deployments.

Quantization is widely used to deploy large language models, but its effect on uncertainty behavior, such as confidence, margins, and abstention, is rarely treated as a primary objective. We frame calibration-data selection for quantization as a target-dependent uncertainty-preservation problem. Different deployments emphasize different regions of the input distribution, yet prior work mainly optimizes accuracy-oriented compression metrics or adjusts scores after quantization. We formalize this goal with distributional and boundary preservation risks, and provide a simple mixture-mismatch argument explaining why no single calibration recipe should be expected to fit all targets. We introduce Doubt-Preserving Quantization (DPQ), a lightweight pre-quantization recipe family that uses full-precision predictions to construct target-aligned calibration mixtures of high-doubt examples and generic anchors. Across 8 language models, 9 NLP benchmarks, and 22 comparison methods, the leading fixed recipe changes with the preservation target: DPQ-r75 leads on SQuAD2 answerability-boundary preservation, while milder or single-signal variants, including DPQ-r50, confidence-only, and entropy-only, better preserve broad multiple-choice QA behavior. These results show that calibration data should be selected for the specific full-precision score behavior a deployment needs to preserve, rather than treated as a fixed quantization detail.

Narratives are central to how social communication is framed, making their detection critical for understanding and analysing public discourse. Prior work has explored narrative detection and extraction across diverse domains; however, migration narratives remain significantly understudied, primarily due to the absence of dedicated annotated datasets. Furthermore, public communication has recently shifted towards video-centric platforms, where narratives are conveyed through multimodal signals and consumed at scale. Despite this shift, narratives in videos remain largely unexplored. To bridge these gaps, we introduce MigrationNarrate, the first multimodal dataset for detection of migration narratives in the UK, consisting of 1,115 YouTube video transcripts annotated using a two-level taxonomy of 12 migration super-narratives and 53 narrative labels. This paper details the dataset design, collection, and annotations; together with benchmark results using a combination of pre-trained encoder models and both open- and closed-source Large Language Models. Finally, a thorough error analysis offers insights for future work.

In this research, we introduce SAraBERT, an enhanced version of AraBERT which proposes inter-sentence transformer layers for extractive summarization tasks. To ensure that the summaries generated by SAraBERT achieve a high coverage of the document's main ideas, we propose Semantic Siamese Similarity, a novel evaluation metric that measures the level of similarity between two text inputs. We validated using BLEU, ROUGE, and Semantic Siamese similarity on Sarabert and published related models. Simulation results showed the effectiveness of our proposed model and motivate follow on research.

Modern open models are hybrids: most layers are linear-attention (Gated DeltaNet, GDN) layers carrying a small fixed-size recurrent state instead of a growing key-value (KV) cache. This makes ordinary decoding memory-efficient, but hurts speculative decoding. To verify a batch of draft tokens and then roll back the rejected ones, today's systems snapshot the full recurrent state at every draft position for GDN layers, and those snapshots cannot be shared across branches of a draft tree, so a wide, high-acceptance tree becomes memory-infeasible. We remove the snapshots. Using a tree-structured WY transform of the gated delta rule, we compute every draft node's output with a single triangular solve and reconstruct only the one accepted state on commit, storing a small pseudo-value matrix instead of per-node states; the derivation depends only on the gated delta rule, not on any other architectural detail. In serving benchmarks on two scales of one hybrid model family (Qwen3.5 35B and 397B) this cuts speculative recurrent-state memory and KV-cache pressure at identical acceptance length, turning the freed HBM into higher throughput and much lower time-to-first-token (TTFT) wherever memory binds, and costing a few percent where it does not. For tree width the same memory buys affordability: a wider, higher-acceptance draft becomes possible, though not yet a throughput win.

Serving large language models cheaply increasingly means shipping models that are both structurally compressed to a fraction of their parameters and quantized to 4 bits. Together these steps degrade reasoning, mathematics, coding, and long-context behavior enough to require a recovery, or healing, stage before deployment. The default recipe, quantization-aware training (QAT), re-fits the compressed, quantized model to hard labels; in our pipeline it converged slowly and collapsed past its peak. We adopted Quantization-Aware Healing (QAH) instead. Because a structurally compressed model is never independently trained at full precision, its bfloat16 checkpoint is a distillation-recovered approximation of the original; QAH distills the 4-bit student directly from the original, uncompressed model. On a GPT-OSS 120B to 60B to MXFP4 pipeline, the QAH student matches or beats its bfloat16 source on 7 of 9 benchmarks at roughly 4 times less weight memory and half the teacher's parameter count, and is released open-weight as Hypernova-60B. Against a matched QAT baseline it reaches a comparable peak about 7 times faster and stays stable under continued training, without hand-tuned early stopping. We also report deployment lessons, including a large, reproducible quality gap between distributed-training backends. Our aim is a recipe deployable without a multi-week hyper-parameter search.

Cross-model latent guidance lets a frozen large mentor encode an input once and a frozen small student generate from the resulting signal. Existing methods keep this signal fixed, assuming it stays useful as the output grows; we show this fails in long-form generation. On multi-turn instruction following, static guidance pushes a 4B student's constraint satisfaction 2.5 points below its no-guidance baseline; a training-free refresh every 16 tokens changes only the memory content and restores a 2.0-point gain over that baseline. We propose MentorPulse to keep guidance fresh at practical cost: it compresses mentor states into a capped slot memory, incrementally processes newly generated tokens, and updates the memory that the student reads through gated cross-attention without resetting the student's KV cache. Windowed Refresh Training exposes the bridge to prefix-conditioned memory. Across thirteen datasets, MentorPulse closes 52.2% of the mentor-student gap on macro average, outperforming C2C, T2T, and equal-budget LoRA, with the largest gains on long outputs. It performs best on all eleven mentor-student pairs from three model families, with margins that narrow as the capability gap grows, and a lightweight read-pattern check predicts the gain before deployment. Measured costs identify refresh intervals that dominate text guidance on long outputs.

Reference-based metrics remain the standard choice in machine translation evaluation, partly because quality estimation methods often correlate less well with human judgments. As a result, source-free, reference-based evaluation has become the practical norm, even though it is unfaithful to the definition of translation adequacy and unfair to systems whose outputs preserve the source meaning while differing from the reference. This paper argues that adequacy must be judged with respect to the source. A reference is only one possible rendering of the source and may introduce bias, under-specification, or errors. We further argue that source-reference-hypothesis evaluation is fair only when the judge treats the reference as auxiliary evidence rather than as the primary standard. Otherwise, even source-aware evaluation can reduce adequacy to preference towards reference. We show the existing hybrid metrics are highly reliant on reference compared to source. Our argument is not that all automatic MT metrics fail to use the source. Rather, we argue that any evaluation protocol that removes the source, or allows the reference to dominate the source, is structurally incomplete for adequacy evaluation. However, existing MT papers generally prefer reference-based metrics and use QE metrics only when reference is unavailable. We therefore call for QE to be reframed as a primary approach to source-grounded adequacy evaluation, rather than as a fallback motivated by missing references. We further call for hybrid metrics whose designs explicitly prioritize source–hypothesis faithfulness while using references only as complementary evidence.

Open-web future event prediction requires agents to distill reliable signals from noisy, redundant, and incomplete evidence. Existing retrieval/memory mechanisms directly feed retrieved information to agents or rely on simple memory functions such as storing and reusing prior information for prediction, leaving them insufficient for open-web forecasting. We propose to transform raw web evidence into structured memory before prediction, enabling agents to reason over distilled, question-specific evidence rather than noisy retrieval results. This paper presents ForeDreamer, a self-evolving dual-agent framework for managing memory over open-web evidence. ForeDreamer separates factual memory, a question-specific evidence state for the current forecast, from experiential memory, persistent agent experience accumulated across forecasting episodes. It uses a main agent for search and prediction, and a memory-processing subagent to convert search results into factual memory with dedicated tools. ForeDreamer further evolves experiential memory through two tracks, improving both forecasting decisions and factual-memory construction. Experiments on Prophet Arena and FutureX demonstrate the effectiveness of ForeDreamer. Project page: https://zhongzero.github.io/ForeDreamer

Automatic Medical Coding (AMC), which assigns standardized International Classification of Diseases (ICD) codes to clinical notes, is essential for medical reimbursement, quality reporting, and clinical research. Existing pre-trained language model (PLM)-based methods typically formulate AMC as an extreme multi-label classification problem over a predefined code set, while recent large language model (LLM)-based approaches instead frame it as generation or multi-step reasoning. However, key challenges remain, including the extreme length of clinical notes that hinders effective interpretation, the vast ICD label space, and complex coding rules that are not explicitly captured by LLMs. In this work, we propose Knowledge-Guided Reasoning over Clinical Evidence with LLMs (KREL), a framework that leverages LLMs for clinical text understanding and reasoning while integrating external ICD coding guidelines as structured knowledge. This design enables tight coupling between domain knowledge and LLM reasoning, reducing hallucinations and improving compliance with coding standards. Experiments on benchmark datasets show that KREL consistently outperforms strong PLM-based and state-of-the-art LLM-based baselines.

Document processing pipelines traditionally cascade optical character recognition (OCR) engines with downstream models for structured information extraction, leading to multi-stage error propagation. We fine-tune SmolDocling, a compact 256M-parameter vision-language model (VLM), to perform end-to-end key-value extraction directly from document images, jointly solving identification, localization, and association in a single pass without OCR preprocessing. We extend DocTags with specialized key, value, region, and link tags, enabling many-to-many relationships in a unified output sequence. To address data limitations, we design an augmentation pipeline combining synthetic form filling and graph-based crops that preserve complete key-value subgraphs. We further introduce a layout-aware evaluation framework extending text matching with spatial bounding box verification. On FUNSD, XFUND, and a large-scale private dataset, our model outperforms larger zero-shot VLM baselines under layout-aware evaluation, while being 27 times smaller than Qwen2.5-VL (7B) and over 5 times faster at inference. The model weights will be released publicly after publication.

Document-Level Relation Extraction (DocRE) relies heavily on costly manually annotated datasets, while large distant supervision resources such as DocRED distant remain underexploited due to noise. We show that a critical yet overlooked source of noise lies in structural inconsistencies within relational triples, including violations of ontology constraints and logical contradictions. We introduce an ontology-driven framework to quantify and enforce structural consistency in DocRE datasets. Our analysis reveals substantial structural noise in DocRED distant and demonstrates that such inconsistencies propagate to model predictions. Enforcing structural well-formedness during training significantly reduces logical contradictions and consistently improves generalization performance. These findings establish structural consistency as a missing axis of supervision in DocRE and highlight structural regularization as an effective strategy for leveraging distant data at scale.

Watermarking diffusion language models (DLMs) requires mechanisms compatible with iterative parallel unmasking rather than autoregressive decoding. Existing sampling-based watermarking methods typically inject position-wise i.i.d. perturbations, which can be poorly aligned with DLM decoding dynamics and degrade generation quality. We propose SAC-Copula, a quality-preserving watermarking method for DLMs based on smooth, locally correlated Gumbel perturbation fields constructed via a Gaussian copula. We further develop a SAC-aware detector using covariance-aware filtering and native-sample calibration. Mechanism-level analysis shows that local correlation reduces latent perturbation roughness and better matches iterative refinement dynamics. Experiments on LLaDA show that SAC-Copula achieves a favorable quality-detectability trade-off compared with existing baselines. In particular, further evaluations on Dream-7B and additional datasets show that SAC-Copula substantially improves PPL tail stability over the i.i.d. Gumbel baseline, while maintaining strong low-FPR detectability and competitive overall generation quality. Additional token-edit stress tests further assess watermark robustness under controlled synchronization drift. Code is available at https://github.com/PunkyKnife/SAC-Copula.

Aspect-based sentiment analysis (ABSA) quadruple extraction requires jointly predicting target, aspect, opinion, and sentiment over reviews that often contain multiple fine-grained sentiment tuples. While large chain-of-thought (CoT) models perform well on this task, distilling them into smaller deployable models remains difficult. We identify a task-specific failure mode in distilled ABSA extraction: student errors at the target-aspect interface create structurally invalid states, such as broken target-aspect bindings and hallucinated targets, which then corrupt downstream predictions. Conventional off-policy distillation is poorly suited to this setting because it trains only on teacher-generated trajectories and provides little supervision on the student-induced structural states that dominate inference. To address this mismatch, we propose STAR-OPD (STructured Aspect-cascade-aware On-Policy Reward Distillation), which builds on generic on-policy distillation and instantiates it for ABSA quadruple extraction with cascade-aware, set-structured rewards. STAR-OPD trains on student rollouts and applies set-structured rewards that directly target binding consistency, target grounding, and fine-grained aspect disambiguation. Experiments on E-ABSA20K and SemEval-2014 show that STAR-OPD consistently outperforms off-policy and general on-policy baselines, reduces target hallucination, and substantially improves performance on structurally hard cases. With Qwen3-4B, STAR-OPD substantially narrows the student-teacher gap while improving inference efficiency, highlighting the importance of on-policy structural correction for distilled ABSA extraction.

Temporal Knowledge Graph (TKG) extrapolation seeks to infer future facts from time-varying relational histories. Recent diffusion-based approaches improve uncertainty modeling through generative denoising, but their aggregated conditioning on subject histories may insufficiently distinguish query-specific evidence from non-salient historical facts, thereby diluting target-discriminative signals. To bridge this gap, we propose FreqDiff, a Frequency-aware Diffusion framework for TKG extrapolation. Specifically, FreqDiff formulates future object prediction as query-slot denoising and develops a dual-stream denoiser that integrates temporal dependency modeling with context-aware spectral calibration. The spectral branch synthesizes history-conditioned filters from learnable bases to adaptively re-calibrate denoising representations, while a frequency-domain regularizer is proposed to align the denoised target with the gold object in spectral space. Experiments on four public TKG benchmarks demonstrate that FreqDiff achieves state-of-the-art performance.

While Large Language Models (LLMs) have significantly advanced reranking in recommendation, effectively leveraging item-side information remains challenging. Real-world items are described by vast, heterogeneous, and unstructured metadata, where decision-relevant signals are often implicit, noisy, or buried in long descriptions. Moreover, feature salience is highly context-dependent, varying not only across items but also across users. Existing methods often rely on item titles, fixed attributes, or static item summaries, which limit personalized and fine-grained item understanding. To bridge this gap, we propose CAIRO, a user context-aware item profiling framework for LLM-based reranking. CAIRO first structures raw metadata and reviews into objective features and subjective traits, and employs a lightweight profiler to select the most relevant information for each user-item pair with limited serving-time overhead. The resulting profiles are concise and context-specific, providing relevant item-side evidence for the LLM's ranking decision. Experiments show that CAIRO consistently improves LLM-based reranking, highlighting the importance of item profiling that effectively exploits vast item-side information.

As scientific literature grows and papers increasingly under-report limitations, multi-agent LLMs offer a promising approach to systematically uncover these hidden failure modes. Here, we introduce Tree-of-Concerns, a multi-agent framework that deploys specialized skeptic personas, each operating through a category-specific analytical lens, as parallel debate trees to extract unstated limitations from scientific papers. Each persona conducts structured, evidence-grounded argumentation, while a Panel Review mechanism re-evaluates each surviving claim from all five perspectives to correct category drift and severity miscalibration. Through retrieval-free, single-paper experiments on ToC-Bench, our benchmark of 414 research papers with 1,905 unstated limitations, sourced from reviewer-reported weaknesses and follow-up citation critiques, we demonstrate that ToC improves precision by 79% and coverage by 11% relative to the strongest baseline, surfacing specific, evidence-grounded concerns that support reviewers in systematic evaluation.

We describe the PSK submission to the WMT 2026 Multilingual Instruction Shared Task. Our system uses the 3.35B-parameter Tiny Aya Global model with three QLoRA adapters, one for each task. The adapters are trained on multilingual document-summary pairs, passage-based question answering, and filtered standalone question answering. The summarization data also includes scientific papers with their author-written abstracts. On our held-out split, the context and summarization adapters perform better than our multitask adapter, which was trained only on data supplied by the organizers. Results for open QA are mixed and vary with answer length and evaluation method. We therefore submit three systems with the same context and summarization adapters but different open-QA adapters.

Language-model agents can improve after failure or carry text across episodes without revising what counts as success. We study the narrower attribution problem of criterion revision: when criterion K0 accepts an outcome violating a broader commitment B, what observations justify saying that the system formed and persistently used K1? We require five non-compensatory conditions: criterion-failure detection, a model-emitted proposal, new-episode transfer, intervention sensitivity on the claimed carrier, and preservation. We evaluate CMB-0.1 on twelve cross-domain cases and four arms: stateless inference, append-only history, model-generated but harness-committed state, and evaluator-written oracle state. Seven mechanism fixtures yield 84 deterministic scorer trials; four local quantized artifacts yield 96 calls and 192 model-case-arm trials. No model trial satisfies all five conditions, but this zero does not establish general capability absence. Eleven calls remain invalid after one retry; several commitments disclose the target distinction; the harness performs commits; deletion reuses a stateless call; and conflict changes multiple factors. Qwen2.5-7B answers every transfer and preservation item without revision state, exposing zero-state reconstruction. These failures make CMB-0.1 an instrument-calibration result rather than a model ranking. We derive a prospective, trace-anchored CMB-0.4 protocol requiring concealed transfer, explicit WRITE/NO-WRITE/ESCALATE actions, a separately logged policy-selected commit, matched interventions, repeated hidden items, and a frozen executable oracle. It is a successor design, not a completed confirmatory result. The paper contributes a measurement chain, an empirical diagnosis of its first implementation, and a more discriminating protocol for future tests of criterion revision.

High-performance ML systems increasingly rely on GPU kernels whose editable source is unavailable, generated, or too distant from final machine code to expose remaining optimizations. Existing LLM kernel optimizers and autotuners mainly operate on CUDA, Triton, HIP, or tensor-program source and validate against reference implementations. We study a stricter setting: optimizing an already compiled AMDGPU code object, where the deployed binary is the only behavioral oracle. We present AsmEvo, an agentic assembly-level optimizer for AMD GPU kernels. Given an AMDGPU code object K0, AsmEvo reconstructs a reassemblable representation, proposes low-level edits with a long-horizon agent, rebuilds an ABI-preserving optimized object, and accepts candidates only after differential verification against K0 under identical launches. AsmEvo combines code-object recovery, metadata-aware rebuilding, profiling-guided hot-window editing, correctness-gated timing, and conservative in-place patch fallback. We conduct extensive experiments with AsmEvo on various AMD GPU kernels. On MI308X, AsmEvo improves 29 of 30 selected KernelBench kernels, reaching 1.35x geometric-mean and 3.88x maximum speedup. On MI300X production workloads, it improves all evaluated AITer binaries and vLLM/SGLang Triton assembly kernels, reaching 1.09x/1.31x and 1.18x/1.34x geometric-mean/maximum speedups, respectively, while preserving functional equivalence.

Retrieval-augmented generation (RAG) has no model of time: when a fact changes across a coding session - a function is renamed, an endpoint moves, a dependency is bumped - RAG retrieves both the old and new value with near-identical similarity and cannot tell which is current, so it serves the superseded value. Paper 1 showed, on synthetic single-value benchmarks, that a deterministic (subject, relation, object) supersession memory eliminates this failure. Here we validate it end-to-end on real software history. From 707 real GitHub issues (SWE-bench Lite + Verified) we extract 130 clean atomic state transitions, a fix that changes one identifiable value from a pre-fix to a post-fix form, and render each marker-free (the stale and current statements differ only in the value). On this set, MemStrata reaches 0.91 answer accuracy versus RAG's 0.57-0.59; and, the structural result, when forced to answer RAG serves the superseded value 36-38% of the time (an LLM reranker does not help) while MemStrata drives this to ~0, at RAG retrieval latency (~2.1 s vs ~18 s for the reranker). We are explicit about scope: only ~18% of real fixes are clean atomic transitions; Paper 2 isolates the memory mechanism on that class, and extraction coverage of the remaining fixes is the orthogonal problem we defer to follow-on work. A real product bug surfaced and was fixed during the study (a case/punctuation-insensitive value comparison), with the moat property (deterministic-supersession accuracy on clean code mutations) preserved and verified.

Many agentic systems must repeatedly choose between acting and abstaining, making faithful reasoning important for oversight: an explanation is useful only if it reflects the computation that produced the action. We study this problem through intervention timing in multi-party conversation, where an assistant must decide whether to speak or remain silent. This setting exposes class imbalance, asymmetric action costs, and the possibility that exposing reasoning changes the policy being audited. Using Qwen3-8B, decoded with or without chain-of-thought reasoning, we compare direct decision policies, reasoning policies, supervised fine-tuning, and reinforcement learning. We find a capability-auditability tradeoff: the strongest direct policy achieves higher quality but exposes no reasoning to inspect, while the reasoning policy provides a trace at the cost of lower performance, particularly recall of true intervention opportunities. Supervised fine-tuning either suppresses reasoning or preserves it without improving decision quality, while reinforcement learning also fails to improve the reasoning policy. We identify one mechanism underlying this failure: group relative objectives provide no learning signal on confidently wrong prompts when sampled rollouts all select the same action. Controlled activation probes and behavioral ablations show that standard faithfulness methods can overstate evidence that exposed reasoning reflects the underlying decision process. Probability-based metrics saturate under confident decisions, probes are vulnerable to class imbalance and textual leakage, and reasoning ablations can confound reasoning content with changes in inference mode. Together, these results show that exposing reasoning can change an agent's action policy rather than simply make it observable. We provide controls for evaluating reasoning-based oversight of agents that can act or abstain.

Enterprise adoption of large language models in finance is constrained less by fluency than by trust: in Financial Planning and Analysis (FP&A) and other regulated workflows, an answer is usable only if it is traceable to authoritative sources and auditable after the fact. This paper argues that retrieval-augmented generation for enterprise finance should be evaluated on auditability alongside accuracy, and presents the Knowledge-Driven Analytics Framework (KDAF), which builds ontology-driven knowledge systems through six iterative stages and retrieves evidence via Context-Aware Relevance Propagation (CARP), so that every retrieved fact carries its relationship type, confidence, and source lineage. An evaluation on FinanceBench (145 questions) compares KDAF against zero-context inference, BM25, concept-weighted lexical retrieval, and ungrounded graph traversal. First, retrieval is necessary: zero-context inference reaches 4.1% correctness against 10-12% for retrieval-augmented conditions. Second, on answer correctness the retrieval conditions are statistically indistinguishable (KDAF vs BM25: -0.007, 95% CI [-0.021, 0.000]), so accuracy alone does not justify structured retrieval here – a negative result we report explicitly. Third, on auditability the ordering reverses: KDAF attains the highest citation traceability F1 (0.515), exceeding ungrounded traversal by +0.027 (CI [0.006, 0.050]) and BM25 by +0.052 (CI [0.024, 0.083]), intervals excluding zero. Graph-structured retrieval also admits no evidence from outside the question subject entity (0 of 426 items, against 16.8% and 20.2% for lexical baselines), and every selected item resolves to a complete provenance chain. We argue that auditability, not accuracy, is the axis on which ontology-grounded retrieval earns its cost.

Splitting a bidirectional LSTM's contextual representation into a forward-only $F_i$ (strictly a function of tokens $1..i$) and a backward-only $B_i$ (strictly a function of tokens $i..n$) beats either alone and beats a fused self-attention representation for dependency relation-type classification. But a specific, natural extension of this idea – pairing a token's forward state against a candidate's backward state (“cross-direction” pairing, $F_i$ vs.\ $B_j$) – consistently underperforms same-direction pairing, and the penalty grows, not shrinks, with token distance, both paired-bootstrap significant. We diagnose why using a frozen-trunk methodology: architectural information leakage between directions is impossible by construction (a single-layer BiLSTM, verified by code inspection); 93% of the same-vs-cross gap survives freezing the trunk and training only fresh heads, ruling out training-co-adaptation as the primary cause; linear regression shows partial representational redundancy between $F_i$ and $B_i$ ($R^2{=}0.324$ vs.\ $0.028$ for a shuffled control) and a linear probe shows partial anticipatory encoding of upcoming tokens in $F_i$ (36.5% vs.\ 17.2% majority baseline) – real effects, but neither alone, nor combined, cleanly explains the full gap. Extended frozen-trunk diagnostics (a positional probe and a distance-decay probe) show directional information is genuinely stored but not exactly positioned, and propagates only a few tokens before decaying to baseline – consistent with, and mechanistically underneath, the distance-growth finding.

Many text classification decisions are viable based on constituent excerpts alone. Taking inspiration from the field of multiple instance learning, we present an algorithm for training a neural network to classify text by selecting such excerpts. We show that our approach is also scalable with demonstrated learning against samples with nearly 1M tokens. We evaluate our methods on 7 datasets with emphasis on long-textual collections that far exceed the encoding limit of our base model. We present state-of-the-art results with this algorithm on 3 datasets: identification of political bias in news outlets, trigger warnings in long stories, and demographic characteristics of authors in tweet collections. Furthermore, the model trained on weakly-labeled collections of text (bags) generalizes to accurately classify constituent, smaller instances. Besides a new state-of-the-art for these problems, this approach is one of the few neural methods to excel in these datasets.

Agents learn to act through interaction with environments, yet the environments used for training are often manually constructed or synthesized around predefined tasks and benchmarks. This task-centric paradigm makes it difficult to scale environments that reflect realistic and evolving workflows where diverse tasks can naturally emerge from the underlying world. We introduce AgentMercury, a scalable framework for synthesizing executable environments from high-level business scenarios. Rather than constructing an environment for a specific task, AgentMercury first instantiates a persistent world with entities, services, tools, state, and executable cross-service invariants, from which diverse tasks and interaction trajectories can subsequently emerge. We construct 4,783 executable environments spanning 14 industries and 50 countries, and use them as training substrates for reinforcement learning. Despite being generated without targeting the evaluation benchmarks, policies trained on these business-oriented environments improve substantially on both enterprise workflows and out-of-domain benchmarks spanning reasoning, coding, scientific computing, and tool use. In our experiments, Qwen3.5-4B improves from 12.3 to 15.7 on EnterpriseOps-GYM and from 45.9 to 56.0 on AIME26 after training on AgentMercury environments. We further show that the construction process itself can be learned: fine-tuning Qwen3.5-35B-A3B on construction traces increases executable-world authoring success from 3.3% to 83.3% on held-out business scenarios. These results show that scenario-grounded environments can provide useful and generalizable learning signals beyond benchmark-specific training, while their construction can itself become a learnable capability.

Efficient-transformer research often motivates token pruning and adaptive computation with the claim that not all tokens require equal computational effort. We test this claim end to end using SEWN, a two-stream Transformer that routes tokens through either lightweight or full-capacity processing using a learned gate. Across our experiments, routing introduces negligible accuracy change relative to parameter-matched baselines, while the gate's token-importance signal depends critically on how it is learned. A static lexicon-seeded prior fails a counterfactual faithfulness test on BoolQ, whereas a fully contextual gate achieves highly significant separation ($p<10^{-10}$) on both evaluated tasks without changing task accuracy.

Claim denial management costs U.S. healthcare approximately $260 billion annually in administrative overhead. Large Language Models (LLMs) and Retrieval-Augmented Generation (RAG) can produce fluent clinical text, but single-agent architectures fail in high-stakes healthcare: they introduce unsupported clinical details and lose the logical structure of hierarchical payer policy. We propose AGVF (Agentic Governance and Adversarial Verification Framework), a multi-agent architecture for medical-necessity appeal generation under explicit policy and evidence constraints. AGVF models appeal synthesis as a Constrained Markov Decision Process (CMDP) over five agents: policy formalization, evidence retrieval, gap analysis, adversarial critique, and gated synthesis. We prove that refinement over a fixed policy constraint graph monotonically reduces evidence-deficiency and terminates with either a complete satisfying frontier or a localized evidence gap. A deterministic citation- grounding gate prevents assertions without admissible evidence from entering shared state. We provide a reference implementation and validate it on 1,000 synthetic appeal cases parameterized from de-identified public hospital discharge data. The validation confirms zero citation-grounding violations across all AGVF cases and monotone deficiency reduction in every episode; ablating the gate raises violations to 100%, confirming it is load-bearing. The study uses no real patient records and does not measure clinical efficacy. AGVF thus contributes a theory-backed agentic architecture and verified reference implementation for policy-constrained LLM generation in healthcare.