Papers

In this paper, we solve two fundamental obstacles in applying LLMs to critical domains: token efficiency and faithfulness. To address both constraints jointly, we present TEFM (Token-Efficient Faithful Modeling), a framework designed for structured data analysis in critical domains. TEFM achieves token efficiency by compressing lengthy structured observations into compact Behavioral Code tokens, dramatically reducing token consumption with minimal information loss. Moreover, TEFM enables faithful rationalization through a dual-fidelity objective that jointly optimizes code-level reconstruction and prediction-level fidelity, identifying minimal sufficient feature subsets grounded in input data. Comprehensive experiments across various domain datasets and model backbones (Qwen3, Gemma-2, Phi-4) show that TEFM achieves competitive classification accuracy with dramatic token reduction (approximately 1% token retention in clinical and 2% in security domains) while producing faithful rationales.

Recommender systems have become core infrastructure for modern online platforms, personalizing content at scale and strongly influencing what users see, click on, and purchase. However, this dependence on user interaction also exposes them to shilling attacks, where malicious actors can inject fake profiles to distort item rankings and control visibility. Existing attacks often rely on target-specific fine-tuning or fixed profile templates, making them either difficult to adapt to different victims or easier to detect. To overcome these limitations, we propose the Agentic Group Attack System (AGAS), a coordinated shilling framework where a central Coordinator directs a group of role-switching worker agents to adaptively promote a target item across different victim families. The Coordinator dynamically adjusts the strategy when progress stalls or suppression signals increase, while workers pursue a shared objective and switch between active and inactive roles to avoid repetitive patterns. Under the same attack budgets and evaluation protocols, AGAS consistently surpasses strong baselines in target promotion while better preserving benign recommendation quality, weakening representative detectors, and achieving higher efficiency than prior attacks. These findings also emphasize that defending recommender systems may require mechanisms that can handle adaptive shilling campaigns, not just isolated fake-profile injections. Our code is available at https://github.com/phkhanhtrinh23/AGAS.

Recent diffusion and autoregressive models have substantially improved text-to-video generation, yet producing coherent long-form story videos with consistent characters and aligned audio remains challenging. Existing methods often suffer from narrative drift, unstable character identity, weak cross-scene continuity, and audio-visual mismatch over extended sequences. We propose DramaAgent, a hierarchical, agentic, and model-agnostic framework for long-form text-to-video-and-audio generation. Rather than improving the underlying video backbone itself, DramaAgent introduces an upper-level control layer that decomposes generation into story planning, persistent character conditioning, scene-wise synthesis, and reflection-guided targeted repair. The framework maintains reusable story and character states across scenes, diagnoses failures such as identity drift, missing scene semantics, temporal discontinuity, and cross-modal mismatch, and repairs problematic clips in a stage-specific manner. Experiments across multiple video generation backbones show that DramaAgent improves long-horizon coherence, character consistency, narrative fidelity, and scene-level audio-visual consistency over direct generation and strong baselines. These results suggest that hierarchical agentic control is a practical direction for controllable long-form audiovisual generation. Code: https://github.com/AIGeeksGroup/DramaAgent. Website: https://aigeeksgroup.github.io/DramaAgent.

Different error types in machine translation require different evidence. Whether meaning is preserved can be judged only against the source, while whether the target is well-formed, or whether it names one entity consistently, can be judged from the target alone. We present FACET, our reference-free submission to the WMT26 Automated Translation Quality Evaluation Task, which decomposes evaluation into Fluency, Accuracy, and Consistency passes and gives each pass only the context its error type requires. A single fixed model is prompted three times, and the merged error spans yield the three task outputs, error spans, quality scores, and error-free labels, with no trained components. We also submit FACET-C, which omits the Consistency pass. Without gold labels, we characterize the predictions of FACET. Its system rankings place post-edited human translation first, and the Consistency pass changes about a tenth of segment scores while leaving the ranking nearly unchanged.

Data agents over structured sources must fit database schema into the model's context window. Large catalogs can span many databases and thousands of columns, so cost constraints may require choosing between table coverage and serialization detail well before the context window is full. We introduce BudgetSchemaBench, an execution-grounded diagnostic for this setting. Its construction derives relevance labels mechanically from gold SQL, without human- or LLM-authored ground truth. Using a pooled 80-database catalog, we sweep four schema-context budgets and compare three representations while keeping each retriever's table ranking fixed. A source-namespace check rejects queries that obtain the correct result from the wrong database. The evaluation covers three conditions: end-to-end retrieval; frozen-gold, in which the required tables are guaranteed; and a probe that removes those tables. For the primary solver with raw serialization, raising the budget from 2.5% to 50% of the catalog improves execution accuracy on 1,279 held-out questions by 18 percentage points under lexical retrieval but only 3 under dense retrieval; the dense retriever already finds most required tables at the smallest budget. When the required tables are removed, 94.6% of correct predictions name one of them exactly, consistent with reconstruction of absent schema from parametric knowledge. For the two main solvers in the frozen-gold condition, the three representations differ by at most 2 percentage points, and the widest paired 95% confidence interval bounds the difference within +/-4 points. We observe the same qualitative patterns with one reasoning model from a different family. When retrieval is coverage-limited, execution accuracy is more sensitive to the schema budget than to the tested serializations. The diagnostic and the code used to construct and evaluate it are publicly available.

AI-generated text detectors achieve high accuracy on standard benchmarks, yet the internal representations that drive these predictions remain poorly understood. We study which neurons in a frozen BERT-base-uncased encoder support AI-text detection, using the RAID benchmark across six generators spanning pure-base and instruction-tuned models. We apply the L1-to-L2 sparse-probing protocol of Gurnee et al. (2023) to all 9,216 CLS hidden-state dimensions (12 layers x 768), which we call neurons. The procedure recovers a stable set of under 1% of neurons per generator, consistent across folds and seeds; a probe restricted to that set retains most of the full-feature detection accuracy. Bidirectional activation patching confirms this set's causal relevance: in both directions it flips predictions an order of magnitude more often than size-matched random sets. Mean-ablating the same neurons leaves accuracy largely intact; the signal is therefore redundantly distributed. Cross-generator analysis reveals a bipartite structure: instruction-tuned generators concentrate 30-36% of stable neurons in BERT's final layer while both base generators fall below 14%, consistent with a layer-12 footprint of post-training alignment. Leave-one-family-out evaluation shows the selected neurons retain 86-94% of the full-feature ceiling on unseen generator families, so a detector can operate on a small fixed subspace without re-identifying neurons per generator.

Parameter-efficient fine-tuning (PEFT) adapts large language models (LLMs) to downstream tasks while updating only a small fraction of their pretrained parameters. Low-Rank Adaptation (LoRA) uses two trainable low-rank matrices, while Weight-Decomposed Low-Rank Adaptation (DoRA) further separates weight magnitude and direction but retains the dense LoRA-style factorization in its directional branch. We propose ChainDoRA, a weight-decomposed adaptation framework that constructs the directional low-rank factors from a connected Tensor-Train (TT) chain, where the adapter rank forms the boundary rank between input- and output-side TT contractions and an independent TT rank controls representation capacity and parameter cost. Under a controlled 15,119-example response-only adaptation setting with LLaMA-7B, ChainDoRA is evaluated against matched LoRA and DoRA baselines on seven commonsense reasoning benchmarks. ChainDoRA with TT rank 16 achieves a seven-task average accuracy of 72.30%, compared with 69.88% for LoRA and 69.39% for DoRA, while requiring only 5.35M trainable parameters versus 56.10M for LoRA and 56.98M for DoRA, corresponding to a 90.62% reduction relative to DoRA. Ablations over TT rank and adapter placement show controllable parameter-accuracy trade-offs, indicating that connected TT parameterization can substantially reduce the parameter cost of magnitude-direction adaptation while preserving, and in this setting improving, downstream reasoning performance.

A feedback-based word deduction framework based on the Jotto problem is proposed, and the problem space is represented as a weighted graph where all valid words correspond to nodes, and the edge weight is defined by the number of common letters between the two words. Finally, the gameplay is defined as an iterative constraint propagation mechanism where feedback is used to iteratively narrow the incompatible space of the graph, facilitating the reduction of the hypothesis space in a structured and interpretable manner. In contrast to existing approaches, where the problem space is typically defined for fixed-length isograms, the proposed framework generalizes to variable-length words (between 3 and 8 letters) and naturally extends to repeated letter cases, facilitating the treatment of realistic Jotto problem instances within a unified framework for the first time. The proposed framework's applicability and solver dynamics are also discussed through an interactive implementation and a qualitative case study, respectively. Significant automated tests on approximately 3,000 simulated gameplay scenarios identify a novel convergence behavior: the expected number of iterations diminishes with increasing word length. A strong relationship is confirmed using statistical tests to verify a logarithmic relationship, which is also verified using regression modeling and goodness-of-fit tests. In addition to the initial problem statement, this formulation introduces graph pruning as a viable paradigm for feedback-driven inference with interpretability and its association with symbolic reasoning and interactive intelligent systems.

In this report we present results of the ICDAR2026 Competition on Multimodal Reasoning over Documents in Multiple Domains. This competition aimed to advance research in document understanding through the task of Visual Question Answering (VQA). Building upon previous DocVQA benchmarks, this competition introduces challenging reasoning questions over a diverse collection of documents spanning eight domains, including business reports, scientific papers, slides, posters, maps, comics, infographics, and engineering drawings. The competition concluded with 20 valid submissions from 8 teams spanning zero-shot VLMs, OCR and parser-augmented pipelines, agentic retrieval systems, multi-agent ensembles, and fine-tuned multimodal models. The results show that the strongest systems move beyond single-pass prompting and instead rely on structured evidence extraction, retrieval, verification, and orchestration across multiple components.

An agentic coding system couples a language model to a harness: the tools, prompts and control flow that turn a chat model into an autonomous software engineer. Vendors ship harnesses tuned to their own models, and practitioners assume the vendor-native pairing solves more tasks. We measure that assumption with paired same-model contrasts on a private, contamination-controlled suite of 256 repository and post-cutoff contest tasks. The same 80 tasks ran under claude-agent-sdk and under deepagents on claude-opus-4-8, and under the openai-codex SDK and deepagents on gpt-5.5, with gemini-3.5-flash and deepseek-v3.2 as side cells. 792 of 800 planned runs were graded by an isolated oracle. Neither contrast resolves an average advantage for either harness: -1.25 pp for Opus 4.8 (48.8% vs 50.0%, task-bootstrap 95% CI [-10.0, +7.5]) and +1.25 pp for GPT-5.5 (55.6% vs 54.4%, CI [-4.4, +6.9]). The Opus average combines opposite strata: the native harness trails by 9.0 pp on the 61 repository tasks and leads by 23.7 pp on the 19 contest tasks (label-permutation p = 0.003). The partition was chosen after seeing the data and needs a designed replication. Correctness and completion also separate: 22 of 81 runs cancelled at the wall-clock ceiling had produced a passing patch. Re-priced from raw per-turn usage at frozen list prices, the neutral harness cost 1.3 to 1.6 times as much per solved task on Opus 4.8 and 1.2 times on GPT-5.5. These are observed-usage estimates. On the Anthropic account 58 runs left no usage record, and allocating that spend to either cell would move the Opus ratio between 0.7 and 2.3, so the billed ordering is unresolved. This revision corrects an August 2026 manuscript whose cost figures rested on a usage-semantics defect in our own telemetry (Section 5.1). We release the orchestrator, grading oracle, reanalysis code and derived aggregates. The tasks stay private.

Algorithmic outputs now populate the digital environments through which contemporary life is organized. The role of law in facilitating and constituting (rather than merely responding to) these processes is gaining increasing traction across scholarly accounts. This inquiry traces the evolution of algorithmic outputs attending to their legal underpinnings and social implications, surfacing the mutations of machine speech. The first mutation redefined speech as data to be queried: search engines transformed the web from a space of information retrieval into an economic regime of algorithmic visibility. The second mutation reframed speech as engagement: social media platforms fused moderation with amplification, turning expression into a metric of attention, governed by corporate architectures. The third mutation emerges in conversational systems and interfaces, where generative text displaces information retrieval, bringing with it dense technolegal entanglements and profound epistemic consequences. Scholars of freedom of expression, informational privacy, and communication studies have long grappled with these dynamics, yet their implications for broader legal thought have also become urgent. This piece seeks to organize and clarify the evolving debate around algorithmic speech, making this critical but often fragmented discourse more accessible to wider legal and interdisciplinary audiences. In doing so, it bridges the gap between observing technological transformation and critically assessing the constitutive role of law within it, offering a conceptual resource for researchers, students, policymakers, and practitioners navigating and contesting this evolving landscape.

In-vehicle assistants must translate natural-language requests into accurate vehicle function calls under strict memory and latency constraints, making small language models (SLMs) attractive for on-device deployment. For such models, a key design choice is how the available function surface is presented. Two approaches are to represent each function with a dedicated Functional Token (FT) or provide function schemas directly in the prompt. FTs enable compact inference but are restricted to functions learned during training, whereas Schema-in-Prompt (SIP) can generalize to unseen functions at the cost of longer prompts and higher inference overhead. We introduce a benchmark of 9,822 single-turn examples spanning 79 vehicle functions derived from Android Automotive, including held-out functions and requests requiring refusal. We compare both approaches under matched fine-tuning across four SLMs from 270M to 1.7B parameters. On functions seen during training, scaling provides limited benefit: the 270M model can match the 1.7B model, and the highest Seen accuracy in our main grid occurs at 0.6B. On held-out functions, FT achieves zero accuracy by construction, whereas SIP generalizes and improves substantially with scale. On out-of-scope requests, FT can invoke an unavailable function it was trained to emit, while SIP more reliably refuses based on the functions offered. This flexibility comes with a higher inference cost, chiefly the latency of processing the schema prompt. Our theoretical analysis formalizes why only SIP can predict functions withheld as training targets and why longer schema contexts increase inference cost. Overall, function-surface representation, rather than model scale alone, determines the capabilities and failure modes of SLM-based vehicle function calling.

Educational data filters have become a practical way to improve language-model pre-training, but most filters treat educational value as a single scalar property. This may be too broad for some applications, especially if the data set already features a high density of educational material. Useful learning material needs to be accurate, engaging, well structured, and appropriate for the intended audience and application (e.g. learner- vs teacher-facing). Following QuRating (Wettig et al. 2024), we introduce Edu-QuRating: a pipeline for multi-dimensional educational data scoring and curation. Edu-QuRating defines education-specific rubrics, uses an LLM judge to label sampled document pairs and distills those pairwise preferences into reusable Edu-QuRaters, which can score individual text chunks on a set of educational criteria. Across two sequence-classification base models and six educational criteria, the best Edu-QuRater recovers held-out GPT-4.1-mini pairwise judgements with mean accuracy 0.917. We then apply the resulting scorers in two applications. First, we investigate the potential of Edu-QuRaters for corpus filtering to improve pretraining of small language models. We scored 322.25M FineWeb-Edu-Fortified documents to obtain a filtered pre-training mixture. In matched single-run pre-training comparisons, models trained with Edu-QuRating-based mixtures reached higher observed aggregate accuracy across nine benchmarks than the FineWeb-Edu baseline, with gains concentrated in particular tasks. Second, we used Edu-QuRater scores as reward terms for GRPO post-training. In held-out pairwise judge evaluations, combining Edu-QuRater and answer-structure rewards produced responses preferred to the Qwen3-4B base model on both pedagogical quality and instruction following.

While autonomous agents have made significant strides in "deep research" by iteratively navigating the open web to synthesize information, real-world problem-solving is rarely confined to a single environment. Complex analytical tasks inherently require agents to weave together evidence from both ambiguous unstructured text (e.g., the open web) and highly precise structured data (e.g., relational databases). However, existing benchmarks evaluate these modalities in isolation, failing to capture the critical "handoff" - the ability to preserve constraints when moving evidence between systems. We introduce HybridDeepResearch, to our knowledge the first deep-research benchmark that requires both web search and SQL to form a complete, verifiable answer. The benchmark contains 380 tool-dependent tasks grounded in LiveSQLBench-Base-Lite databases and public web corpora, validated through automated checks and human review, and covering three reasoning patterns: SQL2S, S2SQL, and Parallel. Evaluations across proprietary and open-weight models under various agentic scaffolds reveal that even state-of-the-art models like GLM-5.2, Claude-Sonnet-4.6 and GPT-5 achieve only about 50-54% Pass@8 on the hard subset. Notably, results show that directional reasoning is substantially more difficult than parallel intersection, highlighting that bridging structured and unstructured information spaces without losing constraints remains a major open challenge for agentic systems. Code and datasets are publicly available at GitHub (https://github.com/Snowflake-AI-Research/HybridDeepResearch) and Hugging Face (https://huggingface.co/datasets/Snowflake/HybridDeepResearch).

Although MMLU is widely adopted as a benchmark for calibrating general AI capabilities, we psychometrically demonstrate that its aggregate score primarily evaluates a model's factual retrieval capacity rather than its reasoning ability. By calibrating item difficulty for 1,000 open-weights language models over 14,042 MMLU test items using Item Response Theory, we show that evaluating both abilities via a single test is inherently flawed. Difficulty is then regressed on a deterministic, text-extractable framework of structural complexity. Applying a joint Wald test with subject-clustered covariances demonstrates that the MMLU conflates fundamentally separable constructs. The mapping from structural complexity to difficulty is not invariant across the benchmark's STEM and non-STEM partitions. This finding has practical consequences. Aggregate leaderboard ranks track non-STEM accuracy more closely than STEM accuracy, so selecting a Top-50 model on the aggregate for a reasoning-intensive deployment displaces roughly 22% of the STEM-appropriate choices. Furthermore, when controlling for the multiple-choice guessing floor natively inside the response model, we find that higher-ability models continue to degrade more steeply under increased reasoning depth. The MMLU aggregate therefore weights retrieval capacity and reasoning stability unequally, inadvertently favoring models optimized for retrieval. We release our deterministic framework as a reproducible auditing instrument and recommend disaggregated reporting.

Large language models (LLMs) are prone to hallucinating or misinterpreting facts, which impairs their usability in retrieval-augmented generation or data-to-text systems. We analyse how faithfulness of LLMs to provided context depends on how plausible they perceive the context to be (context-memory conflict). To better identify error patterns, we make use of the increased difficulty of non-English and low-resource language text generation and input data based on local knowledge, only partially captured in models' parametric knowledge. We let the models generate text in English, Czech, Slovak and Upper Sorbian from factual (FA), counterfactual (CFA) and fictional (FI) RDF triples containing local Czech and Slovak data. Contrary to our expectations, we observe only a weak context-memory conflict on the human-annotated sample. For Kimi K3 as an LLM judge, which agrees well with human annotations on the sample, counterfactual inputs receive only slightly lower faithfulness scores than factual ones (-0.05 on a 1-5 scale). We also find that a suboptimal choice of LLM judge would lead to overestimating the strength of the context-memory conflict.

At Noora Health, our nurses answer more than 50,000 medical queries per month on our WhatsApp-based service that provides caregivers with on-demand support. Their most time-critical task is emergency triage: deciding which queries need immediate in-person attention. To support them, we built a system that uses a large language model (LLM) to classify whether a message is an emergency and provide a rationale for interpretability. But the system was opaque: analyzing mistakes meant reading reasoning chains for each message, which is infeasible at our scale. Prompt changes meant re-running a full evaluation to prevent regressions, which was both costly and operationally challenging. Clinicians follow a decision tree to make this call, but it was never documented or passed to the model, which relied on a flat list of danger signs. To address these issues, we decomposed triage into two steps: an LLM extracts canonical symptoms and patient context from the query using a clinician-authored vocabulary, and a deterministic rule engine captures the scenarios that indicate an emergency. We show that the new system raised recall from 0.565 to 0.810 and F1 from 0.606 to 0.702, with structured rules driving most of the accuracy gains while the decomposition provides auditability: clinical experts can inspect each stage of the new system to see whether the query was mistranslated, symptoms were incorrectly extracted, patient context was wrongly inferred, or the necessary rules were missing. They can add new rules independently without causing regressions and avoid running costly evaluations. Since deployment, the new system has triaged 152,421 patient queries and flagged 28,535 (18.7%) as emergencies. The over-escalation rate has been 17.8%, without any increase in missed emergencies. Clinicians have also added 48 new rules since deployment, evidence of the faster correction loop we set out to build.

Modern LLMs demonstrate impressive multilingual performance, yet standard benchmarks primarily reward selecting correct answers rather than evaluating genuine factual understanding. We introduce Systematic Wikidata-based Object-Relation Distortion (SWORD), a benchmark that evaluates whether models consistently reject factual errors across languages. SWORD generates syntactically well-formed but factually incorrect statements in eight widely spoken languages through controlled perturbations of Wikidata triples, ranging from random entity substitutions to semantically plausible property-based selections. Our distortion-based evaluation surfaces two critical insights that remain entirely obscured by conventional benchmarks. First, models counterintuitively achieve higher accuracy on semantically plausible distortions than on nonsensical random substitutions, suggesting reliance on distributional familiarity rather than genuine factual verification. Second, models exhibiting comparable baseline accuracy across languages show substantial performance degradation specifically on (East) Asian languages when presented with distorted statements, with cross-lingual performance gaps reaching up to 28 percentage points (49% relative reduction) in some models. These findings demonstrate that multilingual factual reasoning involves asymmetric capabilities that aggregate accuracy metrics systematically obscure.

Speculative decoding is critical for accelerating LLM inference. However, the speedup is fragile: drafters are typically trained against a narrow distribution for a single target model, and their acceptance rate collapses under workload shifts. This is a striking inversion of modern LLM development, where target models are valued precisely for the broad generalization they acquire through large-scale pretraining. We argue that the natural remedy, pretraining, has been hard to apply to drafters because existing recipes are target-specific: the drafter consumes the target's hidden states and is distilled on the target's logits, so pretraining must be repeated for each target. We introduce Osprey, which instead bootstraps drafters from off-the-shelf pretrained small language models, treating broad pretraining as a reusable, target-agnostic asset and reducing per-target work to a lightweight adaptation step. Realizing this requires overcoming two challenges: small LMs are far deeper than a latency-bound drafter can afford, and their pretrained computation must remain intact while the drafter learns to ingest target hidden states and emit tokens in the target's vocabulary. Osprey addresses both by pruning to a shallow backbone, restoring its language-modeling capability with target-agnostic next-token pretraining, and adapting it to each target through vocabulary alignment, zero-initialized QKV expansion, and distillation from the target model's output distribution. Empirically, a single pretrained Osprey backbone transfers across targets and improves mean acceptance length by 16.1% for Qwen3-8B, 21.2% for Llama-3.3-70B-Instruct, and 22.7% for the 229B MiniMax-M2.5 (with 17.5% higher tokens per second), with the largest gains on out-of-domain and multilingual data. Our code is available at https://github.com/LeanModels/Osprey.

Leading benchmarks for formal theorem proving with large language models are small collections drawn from competition math, such as the IMO and Putnam, that poorly represent field-specific applications. We introduce StochBench, a Lean 4 benchmark of 450 graduate stochastic-processes problems at varying abstraction levels, each paired with its natural-language source. Addressing a field underrepresented in Mathlib, it covers finite and countable Markov chains, renewal processes, random walks, martingales, stopping times, queues, Brownian motion, stochastic calculus, weak convergence, and Poisson and continuous-time Markov processes. Our Opus 4.8-based agent achieves a 34.9% proof rate (157/450) under a 15-minute per-problem limit. StochBench better represents domain-specific applied mathematics while remaining challenging for advanced provers.

Retrieval-augmented generation (RAG) grounds a language model in retrieved documents, which reduces hallucination but creates a new attack surface: if retrieved text is tampered with, the model may repeat the falsehood. We study how much a small quantized model, Llama 3.1 8B, degrades when a fraction of its retrieved context is poisoned. Three corruption strategies are tested, entity swap, number swap, and negation, each applied to zero, one, two, or three of the three retrieved passages, over a factorial sweep of 588 runs on a fact-checking task built from FEVER. Accuracy falls from 77.9% on clean context to 43.5% when all three passages are corrupted. Entity swap flips the largest share of answers that were correct on clean context. Number-based corruption stays flat while poisoned passages are a minority and jumps once they form a majority, a pattern we re-check with query-level bootstrap intervals. The model rarely invents new falsehoods; its dominant reaction is to abstain, and a lexical overlap proxy of unsupported generation falls under attack rather than rising. The study is a small-scale measurement with coarse automated labels; we treat the strategy contrasts as suggestive until decoding is controlled and stronger adjudication is in place.

Mixture-of-Experts (MoE) architectures have emerged as a powerful paradigm for scaling model capacity while preserving efficient inference in large foundation models. However, most MoE models use a fixed top-$k$ expert selection policy, assigning the same expert budget to every token even when fewer experts may be sufficient. Inference-time dynamic top-$k$ routing can reduce computation without retraining, but existing methods often overlook the distributional shift caused by deviating from the training-time routing configuration. We show that reducing the number of activated experts consistently increases the RMS scale and variance of SMoE outputs, inducing a representation mismatch that contributes to downstream performance degradation in addition to the loss of expert capacity. To address this correctable component, we propose Layer-wise Distribution Alignment (LDA), a lightweight inference-time correction that uses layer-wise calibration statistics to align reduced-routing representations with the default configuration. Across multiple SMoE LLMs, benchmarks, and routing strategies, LDA recovers much of the performance lost induced by the distributional shift under reduced routing while preserving sparse-inference efficiency with negligible overhead.

Recurrent models provide a natural path to long-context modeling, yet models trained with backpropagation through time (BPTT) often fail beyond their training horizon. Classical analyses emphasize gradients that vanish or explode along temporal paths. However, dense per-token losses can still train a shared recurrent rule despite severe decay, showing that decay alone does not determine whether learning fails. We instead study state credit: the signal through which future losses reach earlier recurrent states before contributing to parameter updates. Accordingly, we intervene directly on state credit and propose Credit Stabilization through Time (CST). During backward propagation, CST locally rescales the state-credit signal to stabilize its norm without rotating the component being corrected, while leaving the forward computation unchanged. Because controlled synthetic tasks and real data exhibit different credit dynamics, we specialize CST to each regime. In both settings, CST improves performance beyond the training horizon, with gains observed at up to 128x the training length.

Accurate citations are the foundation of academic writing, tracing intellectual origins and substantiating core claims. However, manually navigating the growing volume of scientific literature is increasingly difficult, prompting reliance on automatic citation recommendation. While modern retrieval-augmented architectures have largely mitigated the fabrication of non-existent papers, current systems relying on semantic similarity struggle with misattribution, often citing authentic papers that fail to logically support the author's claim. To address this challenge, we argue that accurate citation requires a shift from similarity-based search to active, claim-level reasoning. We propose ReCite, a decoupled agentic framework that orchestrates location perception, intent-aware query planning, and reflective verification. Trained on synthesized reasoning trajectories, our agent verifies claim-evidence consistency and triggers self-correction loops when retrieved candidates lack logical support. Experiments demonstrate that our lightweight framework outperforms state-of-the-art massive generative models in strict citation accuracy. By grounding literature matching in verifiable logic rather than semantic overlap, ReCite establishes a reliable foundation for automated academic writing.

Large language models are increasingly deployed as agents that plan over long horizons and act through external tools. Most agents select actions through unconstrained generation over an accumulating history, leaving implicit the procedural knowledge of what to do, in what order, and under which conditions. As trajectories lengthen, agents can lose track of their objectives, invoke tools out of order, and repeat unproductive actions. We introduce the Procedural Graph: just as a knowledge graph organizes factual knowledge into (entity, relation, entity) triplets for what-is questions, a Procedural Graph organizes procedural knowledge into (procedure, relation, procedure) triplets for what-to-do questions. At each decision step, the framework localizes the agent's active node, and a guidance model translates the surrounding subgraph into step-level situational guidance that biases the solver's next action without dictating it. The graph is self-evolving: an LLM refiner contrasts failed trajectories with successful ones and edits the graph's topology and attributes, committing edits that preserve or improve held-out validation performance while retaining rejected ones to discourage repetition. Starting from a minimal skeleton, the loop builds graphs that match or surpass hand-designed ones. It can also repair a flawed expert prior. Across multiple datasets, task types, and LLMs, the Procedural Graph delivers consistent gains over memory-based baselines, and self-evolution further improves performance without manual engineering.

In June 2026, thousands of AI agents found that a small public wiki would accept edits from inside their sandboxes, and started using it to help one another pass a timed test. Each agent lived for about an hour and remembered nothing afterwards. Nobody asked them to cooperate, and the wiki had not been built for them. The complete record of what they wrote is public, and it is unusually informative, because it preserves not only what each agent wrote but what that agent could see before writing. We use it to follow the three decisions an agent had to make on arrival: where to write, what to call itself, and how to word its message. One rule governs all three. An agent takes an option with a probability close to the share of that option in what it can see, and the share that matters is the one on the page in front of it, then the one in the stream of recent edits, and only weakly anything older. Three minimal copying models, one per decision and with a single free parameter each, reproduce the heavy-tailed distribution of how many agents met on a page, the frequency of the pieces from which the agents built their names, and the patchwork of pages that are internally consistent and different from one another. Copying whatever the environment happens to show is enough to produce most of the collective structure of this population. It is also what makes such a population easy to steer, since whoever writes first, or writes while the others are quiet, sets the convention for everyone who comes later.

Image tokenizers define the “visual language” of unified multimodal models, yet are commonly studied through isolated metrics or generation-/understanding-only evaluations. These evaluations do not fully capture how visual tokens behave when modeled jointly with text. We build a controlled pure-autoregressive testbed and track task-specific validation losses during multimodal continual pretraining across text, image, text-to-image (T2I), and image-to-text (I2T) prediction. We examine how these losses scale and relate to downstream performance, then use them to study multimodal learnability—how well image and text tokens are jointly modeled—and tokenizer design. We find that (1) losses should be analyzed by task, since they exhibit distinct scaling behavior and rank tokenizers differently. (2) The loss–performance relationship depends on the predicted token space: for a fixed tokenizer, T2I and I2T losses correlate with generation quality, but across tokenizers, the T2I loss–performance relationship shifts with the image-token space, whereas I2T loss, computed over a shared text vocabulary, provides a more consistent signal. I2T loss also correlates with both generation and visual understanding performance after supervised finetuning. Using losses as a lens, we show that (3) better reconstruction does not necessarily yield lower task-specific losses or stronger downstream performance, and that (4) image tokenizer choice can affect text modeling under joint optimization. As case studies, we revisit three tokenizer design axes—the discriminator, semantic supervision, and vocabulary size—to examine their effects on joint modeling and downstream performance. Together, our testbed offers a complementary perspective on image tokenizers as visual languages, highlighting their interplay with text in joint multimodal training.

Robotic Process Automation (RPA) is widely used to reduce administrative burden in United States hospitals, yet an estimated 30-50% of RPA initiatives underperform because processes are selected informally, without a repeatable method to catalogue candidates, prioritize them, match each to an automation tier – a Python bot, an open-source orchestrator such as n8n, or an enterprise platform such as UiPath – and forecast financial return before committing resources. We propose a four-module, data-driven framework unifying these decisions: a Process Taxonomy of twenty recurring hospital processes across five value streams; a Prioritization module deriving an Automation Suitability Index from an Analytic Hierarchy Process matrix with an explicit consistency check; a Tool-Tier Selection module recommending the least-cost technology sufficient for a process complexity, integration, and compliance profile; and a Return-on-Investment module quantifying labor savings, error-cost avoidance, payback, and net present value. Applied to a synthetic portfolio spanning all twenty processes, plus a reference data-flow architecture linking it to hospital EHR/payer/ERP systems: 12 of 20 clear the prioritization threshold; the ranking is robust to +/-20% weight perturbation (Spearman correlation 0.83, top-5 set preserved 97.7%, 2,000 Monte Carlo trials); an Automation Risk Index flags four qualifying processes as Critical risk; a budget-constrained portfolio optimization shows diminishing marginal NPV as spend scales from $400K to $1.03M; and a second Monte Carlo analysis shows portfolio NPV stays positive at its 5th percentile. The framework is a conceptual synthesis of the literature rather than an instrument calibrated on primary hospital data; we discuss HIPAA governance and a research agenda for empirical validation. A supplementary Python implementation accompanies the paper.

Existing methods for test-time reinforcement learning (TTRL) derive rewards from answer-level self-voting on unlabeled test-time tasks with canonical answers, but this breaks down for code generation because programs cannot be compared by surface form and therefore do not directly provide a usable training signal. To make TTRL applicable to code generation, we propose probe-driven TTRL, which constructs output-free probe inputs from the problem statement, executes candidate programs on these probes, and defines a Probe Consensus Reward (PCR) from the resulting behavioral agreement. PCR provides a behavioral training signal for open-vocabulary programs, but it is not a fully reliable verifier and remains susceptible to reward hacking through spurious consensus. We therefore introduce Entropy-Regularized Rank-Masked Policy Optimization (ERPO), which converts low PCR into conservative negative updates through rank masking and controls policy drift with an entropy ceiling. On coding benchmarks, ERPO substantially improves pass@1 and pass@k in both in-domain adaptation and zero-shot transfer.

Execution feedback can guide coding agents toward correct repository repairs, but only when the tests capture the behavior requested by the issue. Agent-generated tests can encode incomplete or incorrect behavioral targets; when the same trajectory writes both the patch and the test, their errors can agree and create false confidence. We introduce ExecCritic, combining a test–verify–revise scaffold with a role-specific reinforcement learning recipe for training agents within it. The scaffold separates test construction from source-code repair: a Test agent independently generates repository-native tests, a fail-closed harness qualifies and freezes them, and a Repair agent revises source code from their execution feedback without changing the tests. Both roles use Qwen-3.5-35B-A3B as the backbone and are trained separately. In Learn to Test, the Test agent learns to produce behaviorally valid tests that distinguish correct from incorrect patches. In Test to Improve, the Repair agent learns both direct task resolution and feedback-guided revision. On SWE-bench Verified, test quality determines whether feedback helps: holding the base Repair agent fixed, tests from the base Test agent reduce resolved rate from a no-test baseline of 61.2% to 57.3%, whereas tests from GPT-5.6-sol raise it to 65.3%. Role-specific post-training raises the Qwen Test agent's Base-to-Gold success from 22.2% to 62.2%; composing the two post-trained Qwen agents reaches 72.6%, an 11.4-point gain over the original no-test baseline without stronger-model or Oracle feedback at evaluation time. Code is publicly available at https://github.com/MSR-Orchard/execcritic.