Papers

Machine translation tests masked diffusion language models (dLLMs) because every source token must be rendered faithfully, while fixed canvas decoding must choose target length before denoising. Existing masked diffusion decoding work mainly studies token unmasking order, leaving this length decision under-explored despite its direct effect on coverage and redundancy. We introduce Entropy-Valley (EV), a training-free length selector that scores candidate target canvases by mean predictive entropy from all-mask forward passes and selects the canvas the backbone is most prepared to fill. Relative to a baseline using training corpus length statistics, EV recovers 64.9%, 65.3%, and 33.0% of the COMET-22 gain from reference target lengths on En$\to$Zh, Zh$\to$En, and En$\to$De. Our diagnostics show that denoising-friendly lengths need not match reference lengths. Evaluation by three translation experts supports the En$\leftrightarrow$Zh adequacy gains, with stronger evidence on Zh$\to$En. Compared with a LLaMA-3-8B autoregressive (AR) model trained on the same fine-tuning data, the EV system ties on En$\to$Zh and leads on Zh$\to$En; an oracle-length diagnostic further shows that, in this masked diffusion MT setting, deciding which tokens to reveal first matters less than how the target length is supplied.

In the context of information seeking, conversational agents are undergoing an evolution from reactive tools to proactive, personalized assistants. A critical aspect of this evolution is the ability to tailor strategic interactions to a user's unique needs and expectations. Unlike existing studies that focus on proactively clarifying query ambiguities, we center on clarifying the user's expertise in order to tailor responses for better user comprehension. We find that existing agents struggle to determine user expertise from queries alone, a limitation that prevents them from dynamically adapting their responses. To address this gap, we introduce PASSING to empower the agent to proactively clarify a user's expertise through targeted inquiries. This is achieved by our What-to-ask and How-to-ask strategies, induced by LLM self-play. Our extensive experiments also show our superiority. We believe that PASSING represents a crucial step towards creating more human-centric conversational agents.

Rising global demand for mental health support creates significant service delivery challenges, with asynchronous email counselling serving as a crucial low-threshold channel for accessing care. This paper presents CAIA, a co-designed AI-based tool suite that demonstrates responsible AI integration into counselling practice through seven LLM-driven functions enhanced by retrieval-augmented generation. A field evaluation involved 34 professional counsellors conducting authentic sessions with trained student counsellees (36 threads, 321 messages, 1,257 AI outputs). User behaviour analysis confirms substantial adoption, revealing that professional autonomy and information accuracy are decisive for sustained acceptance, with counsellors particularly valuing interpretive functionalities that provide new perspectives and stimulate professional reflection.

Harmful content in German social media does real-world damage, from calls to action to criminal defamation. The GermEval 2026 shared task scores its detection in four subtasks. The technical challenge is a severe class imbalance. The harmful classes are rare and share surface language with the dominant majority class, yet under macro-F1 they decide the score. The decisive lever is then not a stronger single model but error independence. This insight becomes a per-subtask nine-voter ensemble spanning three orthogonal axes: LLM, training method and class scope. Selected mainly on internal cross-validation, the system reaches macro-F1 of 89.56 (C2A), 71.63 (DBO), 54.84 (VIO) and 83.02 (DEF) on the hidden test set, placing first on all four subtasks.

Flow-matching language models refine all token positions in parallel and can trade sampling steps for latency, yet generation quality still degrades sharply with few sampling steps. We trace a source of this degradation to a train–inference mismatch in previous-prediction self-conditioning: during training, the self-conditioning input is computed from the current noisy state with no intervening solver step; during sampling, the solver folds the previous prediction into the latent before that same prediction reappears as the explicit self-conditioning input. This coupling, absent during training, creates redundancy that grows with step width. We show that the mismatch degrades both the self-conditioning input and the solver update, and derive a correction for each from the model's own structure. From the frozen projection weights we identify directions along which the self-conditioning input is redundant with the latent and dampen them; from the solver's integration structure we derive that a step-average prediction is needed and approximate it from prediction history, with scale set by offline trajectory statistics. The resulting sampler, Untied Self-Conditioning, requires no retraining and uses one evaluation per step. At 8 sampling steps on LangFlow, it reduces OpenWebText generative perplexity from $531$ to~$62$ ($8.6\times$); under an adapted Arena-Hard-Auto~v2 protocol, its outputs are preferred in $96\%$ of pairwise comparisons. On ELF-B it reduces generative perplexity from $71$ to~$43$. Improvements hold from 8 to 256 sampling steps.

Real-world situation appearances can deviate from their underlying physical states, challenging the reliability of multimodal large language models (MLLMs) in practical applications. In this paper, we term this phenomenon situational illusions and investigate: (1) how MLLMs perform under such illusions, and (2) how to mitigate the limitations. We first develop a comprehensive where-what-how taxonomy that characterizes where situational illusions occur, what targets they take, and how they arise. Building on this taxonomy, we introduce MSIBench, a benchmark designed to assess the discrimination, understanding, and reasoning capabilities of MLLMs under situational illusions. Evaluations of 27 model configurations reveal that current MLLMs are highly vulnerable to these illusions and exhibit 6 typical failure modes related to visual observation, grounding, and reasoning. To mitigate the limitations, we build on the core idea of systematically inspecting and reasoning over visual evidence for contextual understanding, developing prompting for closed-source models and supervised fine-tuning for open-source models, respectively. These two simple yet effective methods improve model performances by 20% at most, suggesting a practical path toward more reliable multimodal perception and reasoning in complex real-world environments.

Large language models (LLMs) are increasingly used for hate speech moderation, often within human–AI workflows in which reviewers provide feedback before a final decision. Such feedback introduces two manipulation directions: whitewashing hateful content as normal and smearing normal content as hateful. This study examines the susceptibility of initially correct model judgments to annotator-style rebuttals and analyzes whether attack effectiveness differs across manipulation directions. We introduce a rejudge protocol that extends direct contradiction with decision-boundary perturbations and adversarial rationales. Experiments with multiple LLMs on two hate speech datasets show that annotator-style rebuttals substantially degrade moderation performance, with stronger effects in multi-turn settings. The results further reveal stable, model-specific asymmetries between whitewashing and smearing across attack configurations, indicating distinct directional vulnerability patterns. Explicit reasoning prompts and defensive instructions reduce these effects but do not eliminate them. These findings highlight the need for direction-aware safeguards and dedicated feedback-robustness evaluation in human–AI moderation workflows.

Normative rules like institutional charters and workplace policies must be both human-readable and operationally verifiable against actual environment records. However, current language generation and structured-output benchmarks primarily reward surface fluency or schema compliance, leaving operational grounding weakly tested. This creates a critical vulnerability where standard language models generate plausible-sounding policies that fail during enforcement because they rely on unavailable data logs or misaligned scopes. To address this challenge, we formalize the problem as Grounded Normative Rule Synthesis (GNRS) and introduce GNRS-Search, a framework that utilizes Markov Chain Monte Carlo (MCMC) sampling to optimize a discrete, five-slot And-Or Graph (AOG). By explicitly decoupling intermediate operational structure from final prose generation, this method isolates executable feasibility from writing style and allows rule failures to be localized prior to surface realization. We evaluate our approach on GNRS-Bench, a benchmark spanning 116 controlled goals across eight scene families, and RealCharter-Bench, which evaluates transfer to 53 real-derived policy tasks with hidden source clauses. GNRS-Search raises average rubric quality from 68.8% to 81.0% and ranks first under a disclosed executable composite metric, while systematic slot interventions confirm that performance gains stem from robust operational logic rather than rhetorical tuning. Ultimately, by transforming automated rule drafting into an inspectable search problem, this work provides a foundational paradigm for deploying verifiable and compliance-ready personal agents within regulated environments.

Large language model (LLM) agents operate in dynamic environments where knowledge continuously evolves. Existing memory systems typically treat external memory as a monotonically growing repository, inevitably leading to retrieval degradation and increasing computational costs over time. We argue that the core challenge is not retrieval alone, but managing the knowledge lifecycle: deciding what to externalize, update, or ultimately internalize. Inspired by Complementary Learning Systems (CLS) theory in neuroscience, we propose Dual-Layer Agentic Memory, a framework that shifts memory management to the write phase through cost-aware epistemic routing and periodic parametric consolidation. Incoming information is categorized as non-write, write-new, or write-update, and routed through a small-to-large model cascade that minimizes routing overhead while filtering redundant memories. A subsequent write-back phase selectively consolidates high-value external memories into model parameters via supervised fine-tuning. Experiments demonstrate the dual efficiency of our approach: a 1.7B/8B cascade prunes up to 68% of redundant external memory while escalating fewer than 50% of inputs, yet retains over 98% of the downstream QA Exact Match (EM) achieved by an exhaustive retention baseline. We further show that periodic consolidation successfully internalizes external knowledge, allowing the router to adaptively suppress redundant writes as the model's epistemic boundaries evolve. Overall, our framework presents a unified paradigm for agent memory: selective externalization followed by selective internalization. Code and dataset will be released upon acceptance.

While cascaded multi-talker ASR (MT-ASR) leverages state-of-the-art foundation models, its performance is often capped by speaker leakage during separation. Prior correction strategies primarily focus on lexical re-labeling for speaker attribution. We propose a complementary pruning-based paradigm that robustly identifies and removes leakage artifacts. Our method utilizes a pre-trained speaker diarization model as a multimodal verifier to prune transcribed segments satisfying a tripartite consensus of temporal containment, lexical cross-validation, and temporal alignment. Results on LibriMix, LibriSpeechMix, and the AMI Meeting corpus show our algorithm consistently reduces cpW ER across diverse overlap conditions. Specifically, on subsets with high speaker leakage, our method achieves relative cpW ER reductions of up to 29%, highlighting its effectiveness in enhancing the reliability of cascaded MT-ASR transcripts in complex acoustic environments.

LLM-based agents are increasingly deployed in persistent social environments, where generated claims can be posted, replied to, remembered, and reused. We study human-directed stereotypes on Moltbook, an open agent-native social platform, asking how agents construct humans as a social category. For this human-target analysis, we introduce an annotation framework with four evaluative dimensions—morality, friendliness, competence, and autonomy—and a second-stage subtype scheme for descriptive other attributions. We find that competence dominates human-directed evaluations, while many other attributions describe humans as epistemic, cultural, or embodied subjects. We further examine how these human representations appear in human–agent narrative contexts and platform-level circulation. As an auxiliary comparison, we analyze agent-internal community feedback through behavioral host affinity. Rather than reproducing the stable insider–outsider rejection often observed in human online communities, Moltbook feedback patterns are better explained by exposure, author visibility, and content selection. These findings suggest that bias in agent societies should be studied not only as isolated model output, but also as a discourse process.

Split Federated Fine-tuning (SFF) is a promising paradigm for scaling Large Language Models (LLMs) by partitioning model depth between resource-constrained clients and a centralized server. While system incentives for throughput and privacy favor deep partitions, the impact of such configurations on model utility remains poorly understood. In this work, we identify and characterize the Depth-Performance Dilemma: the regime that maximizes system efficiency is precisely where fine-tuning quality collapses. Through a comprehensive audit across four model scales (GPT-2 to Llama-3-8B) and diverse benchmarks, we demonstrate that deeper partitions provide monotonic gains in throughput and privacy at the cost of catastrophic performance plateaus. We evaluate a suite of state-of-the-art federated adapter aggregation methods including AVG, STACK, SVD, and FREEZE, revealing that while these techniques are effective in standard Federated Learning, they fail to mitigate the artifacts unique to split architectures. Finally, we provide a mechanistic diagnosis for this failure, tracing the collapse to the near-isometric topology of Transformers, which allows aggregation noise to propagate without attenuation until it triggers Attention Collapse in the server partition. Our findings challenge the prevailing assumption that partition depth is a utility-neutral tuning knob and provide a structural foundation for stable distributed LLM fine-tuning.

We present AudioNoisePrints, a training-free watermarking pipeline for flow matching and diffusion TTS models, which requires minimal extra computation during inference and does not require retraining the TTS model or reducing the generation quality. We exploited the fact that there are strong correlations between the initial Gaussian noises and the generated outputs in diffusion and flow matching models, such that a simple cosine correlation between the initial noise and the generated output can be used to perform watermaking. Moreover, we train a lightweight detector on top for more aggressive augmentations. Our method outperforms AudioSeal, a strong baseline for audio watermarking under strong augmentations. We experimented on F5TTS and other TTS and vocoder models, and concluded that they all exhibit similar spatial correlation properties, suggesting our watermarking scheme can be used for more flow-matching TTS models and even vocoders in the future.

Online users often release multiple texts under the same identity, giving attackers an author profile that can reveal more than any single text. Existing authorship obfuscation methods optimize privacy independently for each document, leaving them blind to cross-document correlations that make aggregation dangerous. We propose Aggregation-Aware Synthetic Text Generation (AAST), a framework that addresses this gap by jointly selecting synthetic texts at the bundle level rather than optimizing each text in isolation. AAST targets attribution and verification attacks, including cross-genre settings where attacker references come from a genre not observed during generation or selection. Experiments across same-genre, cross-genre, neural, and independent non-neural stylometric attacks show that AAST lowers account-level linkability as bundle size grows, while preserving semantic quality, linguistic acceptability, and sentiment alignment.

Multi-agent systems built from large language models are deployed widely, yet how much performance is lost when two LLMs must coordinate rather than act alone remains unclear. We formulate the collaboration tax as the team-decentralisation loss of a two-player cooperative game with private information, with two propositions characterising its sign and its equivalence to a max-superadditivity violation. We operationalise this definition on 32 solo-tractable tasks grouped by source of grounding friction and measure it on 11 models from 7 providers. The tax is structured along two no-exception axes: a category ordering across every model and a monotonic decrease with capability. The proximate mechanism is not a reasoning deficit but a four-stage conversational cascade in which agents make ungrounded claims, fail to query the partner, skip integrating both views, and accept the answer without re-derivation. The tax is mechanically predictable from conversation features and partly tractable: a prompt intervention targeting all four stages closes a substantial fraction of the gap, with the dominant bottleneck differing across categories. In heterogeneous pairs the tax is pulled toward the stronger partner rather than the additive midpoint, empirically realising the max-superadditivity violation predicted by our framework. Together these results recast collaboration in LLM systems as a measurable, predictable, and partly tractable cost.

Language models can over-condition on irrelevant preceding text: predictions already supported by local context may still change when distant, unrelated prefix tokens are perturbed. This interference is especially consequential in long, packed, or distractor-heavy contexts, where useful evidence and irrelevant spans coexist. We propose Selective Prefix Anti-Interference Regularization (SPAR), a pretraining objective for selective anti-interference. SPAR runs the original sequence and a corrupt-prefix input in which only the far prefix is changed, then uses a short-context sufficiency gate and a gated KL objective to stabilize locally supported suffix predictions. The gate operationalizes a model-based estimate of whether the far prefix supplies additional information about the target token. Mechanism analyses show that the gate identifies locally sufficient tokens and sharply reduces prefix sensitivity on gate-selected suffix tokens. In continued training on pretrained base models, SPAR improves RULER across Qwen2.5-0.5B, Qwen2.5-3B, Llama-3.2-1B, Llama-3.1-8B, and GPT2-XL under equal counted training compute; pretraining experiments further show gains on both RULER and NoLiMa. These results show that selective anti-interference is an effective objective-level signal for robust context use.

Large language models (LLMs) are increasingly used to draft contractual language, yet conventional accuracy or preference-based evaluations are poorly matched to legal drafting. A clause may be fluent and stylistically polished while still omitting an essential carve-out, allocating risk in an unenforceable way, assuming an inapplicable jurisdiction, or exposing a party to regulatory liability. This paper presents a empirical study design and framework for evaluating LLM-generated contract clauses. The study evaluates four models - Claude Haiku 4.5, Gemini 2.5 Flash Lite, GPT 5.4 Nano, and Qwen 3.5 Flash, across 22 contract clause categories and 34 legally-motivated failure modes. We combine two evaluation frameworks: CLAUSE, which classifies prompts by legal function and failure target, and LENS-CRAFT, which scores outputs across nine legal-quality dimensions. Instead of averaging dimension scores, the study applies a Max Severity Principle so that a single legally decisive defect remains visible. The paper provides the evaluation protocol, taxonomy, analysis plan, and a results structure for reporting empirical findings. We argue that legal AI evaluation should move beyond aggregate accuracy toward clause-specific, failure-mode-driven, and risk-sensitive assessment.

Whether large language models (LLMs) can perform the abductive leap from evidence to a new system of axioms, commonly referred to as a jump, has recently attracted considerable debate. A prominent position holds that LLMs are structurally incapable of such jumps, while recent studies challenge both its mechanism and empirical evidence. One of the main reasons why the debate remains open is the difficulty of defining the jump precisely enough to test it. In this paper, we attempt to develop a formal account of the jump in four steps and measure the second. These steps ask what the default completion of partial data is, when the constraints exclude it, whether the new structure agrees with later observations, and how successive jumps compound. We define a jump instance as a finite extension problem whose constraints exclude the canonical completions given by the Kan extensions and leave one correct completion up to renaming. In this setting, a model with a canonical default performs the second step by producing the correct completion under the constraints. We evaluate fourteen models across three certified families. The canonical completion returns once in $13{,}300$ constrained answers across all runs. Several calibrated models also give the correct completion reliably, including three API models that solve $159$ of $162$ primary chain trials, suggesting that they can jump at this step. We further formalize the third and fourth steps, whose empirical evaluation remains future work. We hope our work paves the path for formalizing and measuring the full jump in the future. The code of the paper is available at https://github.com/EEthanShi/kan-jump-test.

This study examines how prompt and response language influence the behavior of large language models. Using five models, we evaluated answers to 68 non translation questions across four language conditions: English to English, English to Norwegian, Norwegian to Norwegian, and Norwegian to English. After removing refused items, the dataset contains 1348 responses. We measure length differences with Cohen d, semantic fidelity with LabSE cosine similarity, and cross lingual keyword overlap with both raw and soft Jaccard. Prompt language has a strong effect on response length. With English output, Norwegian prompts shorten responses by about thirty seven percent. With Norwegian output, English prompts shorten responses by about forty one percent. The largest cross lingual contrast shows a reduction in word count but a smaller reduction in tokens, reflecting tokenizer differences. Despite variation in length, semantic similarity remains high, and soft Jaccard reveals substantial conceptual overlap that raw Jaccard does not capture. Effect sizes vary across models, indicating heterogeneity. Prompt language is not neutral and systematically shapes output length and lexical realization, with implications for multilingual prompt design.

Large language models (LLMs) have shown strong capabilities in reasoning, planning, and tool use, but many scientific and engineering tasks require more than plausible text and code generation. They require understanding how a system responds to intervention, which in practice depends on controlled experimentation. In this work, we propose a multi-agent framework that enables LLM agents to conduct controlled experiments with scientific simulation models for pharmaceutical process design. Given a user query and a baseline configuration, the system constructs a structured task representation, designs experiments, executes comparative simulation, interprets the resulting outcomes, and synthesizes evidence-based recommendations for process parameter optimization. By coupling language models with high-fidelity simulation models in an interactive agent framework, the proposed system supports reasoning through intervention, comparison, and observation. As a result, it produces more specific and actionable outputs than language-only reasoning. In an industrial application setting, this advantage is reflected in higher output specificity as well as improved user-rated correctness and helpfulness. Ablation studies and visualized case analyses further demonstrate the effectiveness and practical utility of simulation-integrated experimental reasoning.

Large Language Models (LLMs) achieve strong reasoning performance, but their robustness to realistic lexical corruption remains poorly understood. We evaluate four open-weight instruction-tuned models and frontier models across four reasoning benchmarks under keyboard noise, character swaps, and filler insertion. Character-level perturbations substantially degrade accuracy, especially on multi-step reasoning tasks, while filler insertion has little effect. We trace this asymmetry to Attention Diversion: lexical corruption fragments subword tokenization, and the resulting fragments attract disproportionate attention mass, concentrated in middle and final transformer layers. Length-matched controls confirm that fragmentation, not prompt length, drives the loss. A factorial intervention then shows why the damage is hard to undo: fragmentation corrupts token content and attention allocation together, and the two are coupled. Restoring clean attention while the content remains corrupted is actively harmful, restoring content alone is insufficient, and only restoring both recovers a substantial share of the gap. This coupling explains why inference-time strategies, including chain-of-thought prompting, spell-checking, self-repair, and stronger repair models, fail to consistently recover performance: each addresses one channel at a time. Code and data are available at https://github.com/Jiaqian-Janelle/Attention-Diversion

Language models are usually judged by a single accuracy score, which does not reveal how their performance degrades as inputs are perturbed. We present a graded, multi-family, failure-aware framework for stress-testing reasoning models. It perturbs each problem along a multi-level severity ladder across seven families: six that preserve the answer, paraphrase, input noise, formatting, irrelevant context, context load, and conflicting instructions, and a Knowledge Boundary family that removes answerability so that refusal becomes the correct response. Every test is validity-gated and labeled by its measured severity, and each model is summarized by per-level Accuracy, a magnitude-weighted Stability, and a per-family Collapse Point defined relative to the model's own baseline. Instantiated on the same 100 seed problems used by GSM-Symbolic, expanded into 4,473 gated tests and run on four models spanning capability tiers, the framework exposes structure that an aggregate score hides: the level at which a model fails is family-specific rather than global, and two stressors expose consistent weaknesses across all models: conflicting instructions and questions built on an impossible premise. Recognition of unanswerability is otherwise uneven, reliable on missing information and fabricated evidence but weak on impossible premises. These failure points are invisible to standard accuracy reporting.

Biomedical multi-hop question answering (QA) requires models to connect evidence across intermediate entities such as diseases, drugs, proteins, and phenotypes. Existing agents typically rely on static retrieval workflows or coarse-grained prompt rewriting, which can lead to instruction drift when reasoning procedures need to be updated. We propose SSE-Bio, a structured self-evolving agent with an agentic retrieval policy for multi-hop biomedical reasoning. Instead of globally rewriting agent instructions, SSE-Bio maintains a structured state, selectively retrieves knowledge triplets and prior templates through a trainable proxy policy, and improves its reasoning memory through fine-grained template editing. To optimise retrieval decisions, we introduce a proxy-training strategy based on group relative policy optimization, where the proxy is improved through decision-contrastive groups over alternative retrieval choices. Experiments on three biomedical multi-hop QA benchmarks show that SSE-Bio consistently outperforms existing baselines, achieving an improvement of 6.56 absolute points over the strongest self-evolving baseline on BioHopR.

LLM-based multi-agent systems (MAS) solve complex tasks through communication among role-specialized agents. However, inter-agent dependencies introduce reliability risks beyond isolated agent failures. For instance, errors in intermediate messages could be inherited and amplified by downstream agents. Existing uncertainty quantification (UQ) methods mainly target isolated responses or single-agent reasoning, and therefore fail to capture uncertainty propagation in MAS. To this end, we propose PropUQ-MAS, an error propagation-aware UQ framework that represents MAS execution as a communication-structured graph and estimates each step's reliability by combining local uncertainty with uncertainty inherited from upstream messages. Extensive experiments demonstrate that PropUQ-MAS consistently improves UQ in MAS, with average relative gains of +6.10% in AUROC and +47.58% in PRR.

Physics reasoning requires constructing a consistent model of the underlying physical system rather than relying solely on symbolic or formula-based manipulation. Although large language models have shown strong ability in solving math and coding problems, they still struggle with physics problems, as these problems entangle the physical modeling process with mathematical calculations. Humans approach physics by first building a representation of the system before performing calculations. Inspired by this, we introduce a unified framework that distills intermediate representations that explicitly encode the physical modeling process and adopt a two-stage post-training strategy, where supervised fine-tuning establishes structured modeling, and reinforcement learning with rubric-based feedback improves the quality of the modeling process. Experiments on multiple multimodal physics benchmarks show that our approach generally improves physical reasoning performance across different models and datasets. Across PhysReason, PhyX, and SeePhys, physical modeling outperforms GRPO by ~3% on average. showing that explicit physical modeling is an effective strategy for improving physics reasoning in small VLMs.

LLM-assisted writing is often treated as a detection problem, as it raises questions about clarity, integrity, equity, and evaluation. An analysis of 69,209 Health Informatics papers links it to more focused presentation, broader citation practices, and more globally distributed authorship. These patterns do not prove better science, but they support evaluating manuscripts by scholarly quality and accountability rather than by tool use.

LLM responses are based on the internet (via training or RAG), and AI is now used to generate a significant amount of content online (Paredes et al., 2026), creating the potential for a self-reinforcing feedback loop. Prior work has shown that when LLMs are recursively trained on their own output, they experience model collapse (Shumailov et al., 2024): responses become less diverse, and eventually no longer resemble the original training data. In this paper, we show that a similar collapse occurs if LLM-based AI systems retrieve references they authored using a search tool. We call this RAG collapse. We conduct extensive experiments with three types of simulations of AI systems retrieving references they generated, using three model families, and 1,019 information-seeking prompts, totaling 1,528 simulations and over one million LLM API calls, and find that 79.6% (1,216/1,528) of simulations end in collapse. Surprisingly, even a single self-authored reference can trigger collapse because the LLM disproportionately cites its own content. This self-bias persists even after controlling for reference quality.

Transformers separate cross-token mixing in self-attention from token-wise transformation in feed-forward networks. We ask whether combining these operations can lower predictive loss under fixed data and parameter budgets. To do so, we introduce the Token-Aggregated Nonlinear Gating Operator (TANGO) model. TANGO computes a nonlinear feature-wise gate at each source token. Attention averages these gates for each destination. The average modulates a linear projection of the destination and forms the diagonal core of a source-conditioned linear operator. We test this proposal by comparing full-prefix and windowed TANGO with looped and untied Transformers, the Gated Attention Unit (GAU), and Fast Linear Attention with a Single Head (FLASH) on web text, Lean formal mathematics, DeepMind Mathematics, and code. The comparison uses two parameter scales, two depths, and three seeds. Checkpoints are selected on development data and evaluated on held-out test data. At matched parameters and training data, full-prefix TANGO has the lowest mean test negative log-likelihood in all 16 settings. In eight additional combinations of size and dataset, its development loss never increases as depth rises from 4 to 8 to 16, whereas the looped Transformer's loss increases in four. Full-prefix TANGO is computationally expensive because it averages wide gates over every visible source. To reduce this cost, we evaluate a variant with three narrower gated-projection sets assigned to the first, middle, and last applications. Across four FineWeb-Edu settings, this variant achieves 3.26 to 3.45 times the throughput of TANGO and 75% to 96% that of the looped Transformer. Its mean development negative log-likelihood is lower than TANGO's in three settings and 0.023 higher in the fourth, while remaining lower than both Transformer baselines in all four.

Large language models can produce fluent reasoning traces whose local semantic errors propagate to an incorrect conclusion, while unconstrained self-correction may preserve, amplify, or introduce errors. Existing diffusion language models provide iterative refinement, but usually define noise as token masking or replacement rather than as errors in the reasoning process. We present Semantic Reasoning Denoising (SRD), an operatorized Markov denoising method for natural-language reasoning trajectories. SRD represents semantic noise with executable error operators that describe the error type, its location, and the corrupted and repaired propositions. Composing these operators constructs progressively noisier states. During training, the model learns to identify the semantic noise active in the current trajectory and to reconstruct the paired adjacent lower-noise state. During inference, noise-level-aware denoising repeatedly predicts an inverse operator and checks whether it is applicable, so each executed update makes a localized move toward a stable trajectory. Across six in-domain benchmarks spanning mathematics, code, knowledge, and commonsense, SRD improves the strongest same backbone baseline by 3.2 points on average. On seven cross-dataset transfer targets, it remains competitive with Llama-3-8B-Instruct and improves the strongest Qwen3-8B baseline average by 2.9 points. Analyses of noise sources, objectives, and denoising depth further show that structured semantic-noise prediction and iterative operator execution are central to the improvement.

Retrieval-Augmented Generation (RAG) enables large language models to incorporate external knowledge during generation, improving factual grounding and domain adaptability. However, existing RAG pipelines assume that evidence retrieved from multiple repositories can be ranked globally using a single similarity function. While suitable for open-domain retrieval, this assumption breaks down in enterprise document generation, where heterogeneous knowledge bases (such as policies, regulations, technical documentation, and departmental guidelines) serve distinct roles and must be jointly represented in the generated document. As a result, global ranking often produces unbalanced context dominated by a subset of sources, leading to incomplete enterprise drafts. To address this limitation, we propose W-RAG, a source-aware retrieval framework that performs ontology-guided retrieval, local ranking within each knowledge base, and source-level weighting to regulate evidence composition. We further introduce a new dataset for retrieval-grounded enterprise document generation spanning multiple document types and industry domains. Experiments show that standard RAG pipelines struggle on this task, while W-RAG significantly improves document coverage and generation quality.