Academic Intelligence · Curated Daily

探索全球前沿学术脉络

AcademicHub 汇聚顶级期刊与预印本平台的实时文献。定制您的专属科研雷达,利用大语言模型自动生成交叉领域文献分析简报。

01.
arXiv (quant-ph) 2026-06-15

Perturbative Input-Output Theory of Floquet Cavity Magnonics and Magnon Energy Shifts

arXiv:2512.12103v2 Announce Type: replace-cross Abstract: We develop a perturbative input-output formalism to compute the reflectance and transmittance spectra of cavity magnonics systems subject to a Floquet modulation. The method exploits the strong hierarchy between the magnetic-dipole couplings transverse (drive field) and parallel (modulation field) to the static bias field, which naturally introduces the small parameter $\epsilon = (2Ns)^{-1/2}$ associated with the total spin $Ns$ of the ferromagnet. By organizing the cavity and magnon fields in a systematic expansion in $\epsilon$, we obtain compact analytic expressions for the spectra up to second order. Using these results, we reproduce the characteristic sideband structure observed in recent Floquet cavity electromagnonics experiments. Furthermore, accounting for the Zeeman interaction between the modulation field and the fully polarized ground state - a contribution typically neglected in previous treatments - we predict an additional magnon detuning of approximately $0.8\,\mathrm{GHz}$, independent of both modulation frequency and sample size and determined solely by the spatial volume occupied by the modulation field. This identifies a measurable and previously overlooked shift relevant for the interpretation and design of cavity magnonics experiments.

02.
arXiv (CS.AI) 2026-06-11

ConsistencyPlanner: Real-time Planning with Fast-Sampling Consistency Models

arXiv:2606.11569v1 Announce Type: cross Abstract: Closed-loop planning in complex, real-world driving scenarios presents a critical challenge for autonomous driving systems. While traditional rule-based methods are interpretable, their predefined heuristics lack the adaptability for dynamic traffic environments. Learning-based approaches have shown considerable promise. Conversely, learning-based approaches, despite their promise, struggle to balance the modeling diverse and multimodal driving behaviors and real-time planning, often leading to indecisive or unsafe actions. To address this limitation, we propose Consistency Planner, a real-time planning framework with fast-sampling consistency models. Our approach is built upon two key technical contributions. Efficient Multimodal Sampling: We employ fast-sampling consistency models to generate a diverse set of plausible future trajectories. This enables efficient, real-time exploration of multimodal actions, overcoming the computational bottlenecks of previous iterative generative methods. Heterogeneous Feature Fusion: We introduce an attention-enhanced decoder that dynamically integrates heterogeneous input features (including scene feature and action token) into a cohesive representation for robust planning. Extensive evaluation in the Waymax simulator demonstrates superior performance in safety metrics compared to existing methods, with particularly strong results in challenging dynamic scenarios.

03.
arXiv (CS.AI) 2026-06-17

Knowledge Reutilization in Meta-Reinforcement Learning

arXiv:2606.18132v1 Announce Type: new Abstract: Meta-reinforcement learning enables fast adaptation by extracting shared structure from related tasks, but existing end-to-end methods often couple task inference with embodiment-specific control. This coupling can obscure non-parametric task semantics, reduce sample efficiency, and limit cross-agent reuse. We propose a meta-knowledge reutilization framework that learns task-level knowledge on a dynamics-simplified agent and transfers it to heterogeneous agents. The framework uses a Bayesian non-parametric prior to organize latent task modes and a high-level policy to generate task-level magnitude guidance. To bridge reusable task knowledge with different embodiments, we introduce a semantic-magnitude interface and a lightweight temporal adaptor, which convert frozen meta-knowledge into temporally aligned subgoals for embodiment-specific low-level controllers. Experiments on multiple locomotion agents show that our framework reduces final-step tracking error by 94.75% – 99.79% compared with recent state-of-the-art baselines and achieves comparable deployment performance with about 23.8% of their interaction data.

04.
arXiv (CS.AI) 2026-06-17

LongWebBench: Evaluating Structural and Functional Webpage Generation in Long-Horizon Settings

arXiv:2606.17727v1 Announce Type: new Abstract: Recent vision-language models (VLMs) have shown promising progress in generating webpages from visual inputs, yet existing evaluations mainly focus on short, single-screen, and largely static webpages. We introduce LongWebBench, a benchmark for evaluating long-horizon webpage generation from both structural and functional perspectives. LongWebBench contains 490 real-world long webpages for structural fidelity evaluation and 507 goal-oriented interaction tasks over 129 webpages for functional evaluation. It employs two complementary protocols: a multi-dimensional VLM-based metric for assessing long-range structural coherence, and a DOM-augmented agent-based pipeline for end-to-end functional verification. We further examine the automatic evaluation protocols through human agreement analysis. Experiments with state-of-the-art open-source and proprietary VLMs under single-image and multi-image settings reveal that structural fidelity degrades as webpage length increases, while visually plausible generations often fail to support executable multi-step interactions. These results highlight the need to evaluate long webpage generation beyond visual similarity, with executable interaction as a core criterion. Our code and data are available at https://github.com/zheny2751-dotcom/LongWebBench.

05.
arXiv (CS.AI) 2026-06-16

Adaptive inference and function vectors in deep transformers

arXiv:2606.16694v1 Announce Type: cross Abstract: Transformers are widely used as a general-purpose substrate for learning complex correlations between a large collection of coupled variables, but their internal mechanisms have remained mysterious. We introduce a theory of a deep transformer as a mean-field interacting system that implements distributed inference, subject to constraints on communication, locality and depth. We show that such a system can exploit internal state representations ('function vectors') to infer a latent context variable at increasingly finer scales over its layers. In an in-context regression task, the theory predicts a non-trivial relationship between non-Gaussian, hierarchical structure in the latent context variable, and transformer depth. Predictions are tested using constrained linear attention transformers and demonstrate adaptive inference in deep architectures. Feedforward blocks and depth enable transformers to implement a much richer class of in-context learning algorithms than previously described.

06.
arXiv (CS.AI) 2026-06-16

Skill-to-LoRA: From Using Skills to Learning Behaviors for Token-Efficient LLM Agents

arXiv:2606.16769v1 Announce Type: new Abstract: Agent skills are commonly distributed as SKILL.md files: human-readable procedural documents that describe workflows, tools, resources, and domain conventions. While convenient for inspection and reuse, this design requires the same reusable procedure to be repeatedly injected into the runtime context. We propose Skill-to-LoRA(S2L), a behavior-centric skill representation that replaces runtime skill text with skill-specific LoRA adapters. Rather than compressing the skill document itself, S2L models the behavioral change induced by the skill text: offline, the complete SKILL.md is used to synthesize skill-guided demonstrations; online, the full document is omitted and the corresponding LoRA adapter is dynamically loaded to activate the learned skill behavior. We evaluate S2L with Qwen3.6-27B on a 21-skill subset of SWE-Skills-Bench. Compared with the no-skill and Full Skill Text baselines, S2L improves pass rate by 2.9 and 5.2 percentage points, respectively, while reducing per-step token cost by 6.6% relative to Full Skill Text prompting. S2L matches or improves Full Skill Text on 18/21 skills and the no-skill baseline on 15/21 skills. Control experiments further show that the gains depend on skill-specific adapter alignment: Wrong-LoRA and Shared-LoRA both reduce performance. These results suggest that many procedural agent skills can be converted from runtime instructions into trainable, dynamically loadable behavioral modules. Code will be released upon acceptance.

07.
arXiv (CS.LG) 2026-06-15

Adaptive Nucleus Truncation for Long-Form Reasoning

arXiv:2606.13982v1 Announce Type: cross Abstract: Sampling plays an important role in long-form language-model reasoning. Over thousands of decoding steps, small changes in the candidate token set can compound into different reasoning trajectories, stability profiles, and final answers. Existing truncation methods such as top-$p$, min-$p$, and fixed top-$n\sigma$ sampling improve over unrestricted sampling, but they rely on fixed thresholds that cannot adapt to changes in entropy, task difficulty, training stage, or generation budget. We introduce Adaptive Nucleus Truncation Sampling (ANTS), which extends top-\(n\sigma\) sampling from a fixed decoding rule into an adaptive rollout-control mechanism for long-form generation. ANTS selects standardized neighborhoods around the maximum logit before temperature scaling, adapts the truncation width using an entropy-conditioned controller, and retains a no-truncation fallback arm to stabilize training when truncation becomes unsafe. On a 33B-total / 4B-active sparse Mixture-of-Experts reasoning model, ANTS improves average performance over percentage-based benchmarks by +1.9, +3.8, and +5.2 points at 8K, 16K, and 32K generation budgets, respectively. The strongest gains appear on instruction following and mathematical reasoning, with IFBench improving by more than 10 points at 32K and AIME 2025 improving by 7 points. Code generation reveals an important budget interaction. On Codeforces, ANTS trails the baseline at 8K, but reverses this gap and substantially improves ELO at 16K and 32K. These results suggest that sampler design should be treated not just as a decoding hyperparameter, but as part of how we stabilize and scale long-budget reasoning.

08.
arXiv (CS.CL) 2026-06-11

FinTradeBench: A Financial Reasoning Benchmark for LLMs

Real-world financial decision-making is a challenging problem that requires reasoning over heterogeneous signals, including company fundamentals derived from regulatory filings and trading signals computed from price dynamics. Recently, with advances in Large Language Models (LLMs), financial analysts have begun to use them for financial decision-making tasks. However, existing financial question-answering benchmarks for testing these models primarily focus on company balance sheet data and rarely evaluate reasoning about how company stocks trade in the market or their interactions with fundamentals. To leverage the strengths of both approaches, we introduce FinTradeBench, a benchmark for evaluating financial reasoning that integrates company fundamentals and trading signals. FinTradeBench contains 1,400 questions grounded in NASDAQ-100 companies over a ten-year historical window. The benchmark is organized into three reasoning categories: fundamentals-focused, trading-signal-focused, and hybrid questions requiring cross-signal reasoning. To ensure reliability at scale, we adopt a calibration-then-scaling framework that combines expert seed questions, multi-model response generation, intra-model self-filtering, numerical auditing, and human-LLM judge alignment. We evaluate 14 LLMs under zero-shot prompting and retrieval-augmented settings and witness a clear performance gap. Retrieval substantially improves reasoning over textual fundamentals, but provides limited benefit for trading-signal reasoning. These findings highlight fundamental challenges in the numerical and time-series reasoning for current LLMs and motivate future research in financial intelligence.

09.
medRxiv (Medicine) 2026-06-15

Data-Driven Stochastic Model for Detecting Patientswith Alzheimer's Disease

Alzheimer s disease (AD) is a critical neurological disorder that causes the brain to shrink and leads to the eventual death of brain cells, adversely affecting a person s ability to function. AD is a fast-growing disease in the United States and was the fifth leading cause of death among Americans 65 years of age or older in 2023. In the United States 6.9 million people aged 65 or older were diagnosed with AD, along with a high rate of undiagnosed patients. Thus, the objective of our study is to develop a real data-driven predictive model to identify a patient with AD based on eight risk factors: Age, Gender, ADAS-Cog13, Entorhinal, Fusiform, Intracranial Volume (ICV), Amyloid-Beta, and Tau Protein, with a high degree of accuracy. The quality of the model was evaluated using well-established and sophisticated statistical measures: the area under the receiver operating characteristic curve, calibration plot, Hosmer-Lemeshow goodness-of-fit test, and K-fold cross-validation. If a patient is given information on the above risk factors, our proposed binary logistic regression model can classify the patient as having AD or not with at least 98% accuracy.

11.
arXiv (CS.LG) 2026-06-16

Phase-Localized Curation Does Not Help: A Negative Result on Per-Phase Metric Selection for Demonstration Filtering

作者:

arXiv:2606.15064v1 Announce Type: new Abstract: Manipulation demonstrations have temporal phase structure, and a natural hypothesis is that demonstration-curation metrics should be applied within phases rather than globally. The idea is to segment each trajectory into phases, score each phase with the metric that is locally most informative, and then aggregate. This follows directly from prior work showing that a single global metric can be the best detector of a defect and yet the worst curator of the resulting policy. We test the per-phase hypothesis on three contact-rich LIBERO pick-and-place tasks with a controlled early-release structural defect, comparing phase-gated curation against the same metrics applied uniformly and against a strong single global metric. Across all three tasks and five random seeds per condition, phase-gated curation is never the best curation strategy, and it is the worst of the three on two of the three tasks (Task 1: 86.0 vs. 92.0 for global; Task 3: 22.7 vs. 48.0 for uniform). We trace the failure to a concrete mechanism. When the defect signal is concentrated in a single phase, rank-aggregating across phases dilutes that signal with uninformative scores from defect-free phases, selecting a worse demonstration subset than simply applying the defect-informative metric everywhere. We further show that the per-phase metric selection does not transfer across tasks, since no phase shares a winning metric between any two tasks, so the selection cannot be reused and must be re-derived per task from a noisy sweep. These results bound a plausible and previously untested method, and they argue that practitioners should prefer identifying a single defect-informative metric over decomposing curation by phase. We release the full pipeline, all metric implementations, and per-seed results.

12.
arXiv (CS.AI) 2026-06-15

Causal Object-Centric Models for Planning with Monte Carlo Tree Search

arXiv:2606.14418v1 Announce Type: new Abstract: We introduce COMET (Causal Object-centric Model for Efficient Tree search), a model-based reinforcement learning algorithm that performs Monte Carlo Tree Search in a slot-structured latent space. COMET pairs a frozen unsupervised object-centric encoder with a transformer-based world model, in which actions are bound to objects through a novel action-slot fusion mechanism that is used in slot transition prediction. Policy and value heads use object-causal attention, modulating token interactions by learned per-slot relevance scores so that decision-making concentrates on task-relevant entities. COMET adds an explicit object-level inductive bias to MuZero-style latent planning. Across eight visually and dynamically diverse tasks from the Object-Centric Visual RL benchmark, ManiSkill, Robosuite, and VizDoom, COMET achieves a higher mean normalized score during the early stages of training compared to object-centric and monolithic baselines.

13.
arXiv (quant-ph) 2026-06-12

Steady-State Noise Signatures of Lindbladian Exceptional Points

arXiv:2606.13377v1 Announce Type: new Abstract: Exceptional points (EPs) are non-Hermitian degeneracies at which two or more eigenvalues and their corresponding eigenvectors coalesce. In open quantum systems, exceptional points can arise in the Lindbladian governing the dissipative dynamics. Their signatures have so far been mainly identified in finite-time observables, such as transient currents, while steady-state average currents generally provide no direct evidence of the underlying exceptional-point structure. In this work, we demonstrate that signatures of Lindbladian EPs can nevertheless be accessed in the steady-state regime through current noise. We derive general expressions for current correlation functions within a Lindblad master-equation framework and show, in particular, how exceptional points affect their behaviour as a function of the time delay. We illustrate these results with the paradigmatic example of two interacting qubits coupled to two reservoirs, where the steady-state noise clearly distinguishes overdamped, underdamped, and critical regimes. Our results establish current correlation functions as a steady-state probe of Lindbladian EPs in open quantum systems.

14.
arXiv (CS.LG) 2026-06-17

Bounded Difference Concentration for Infinitely Exchangeable Sequences with Applications to AI Benchmark Uncertainty

arXiv:2606.17426v1 Announce Type: cross Abstract: We consider the concentration properties of functions of infinitely exchangeable random variables. By conditioning on the de Finetti directing measure, we show that the deviation of any function with bounded-difference constants $c_1, \dots, c_n$ decomposes into a conditional sampling fluctuation and a latent mixture fluctuation. When this latent mixture is $\sigma_{\mathrm{mix}}^2$-subgaussian, we establish a concentration inequality with an effective variance proxy of $\frac{1}{4}\sum_i c_i^2 + \sigma_{\mathrm{mix}}^2$. Crucially, we demonstrate that for zero-sum linear contrasts, such as the difference between a subsample mean and a full population mean, the latent mixture term cancels exactly. This cancellation yields a tight, mixture-free Hoeffding-type bound that provides a direct de Finetti mechanism for the infinite-extendibility limit of recent finite-exchangeable concentration results. We apply this framework to quantify uncertainty in composite AI benchmarks, such as MMLU, where question items naturally exhibit exchangeable dependence across domains. Our results provide both a domain-stratified hierarchical model for bounding the uncertainty of accuracy scores, and a distribution-free, cost-saving statistical guarantee for accurately estimating full benchmark scores from random subsets.

15.
arXiv (CS.CL) 2026-06-11

The Language You Ask In: Language-Conditioned Ideological Divergence in LLM Analysis of Contested Political Documents

作者:

Large language models (LLMs) are increasingly deployed as analytical tools across multilingual contexts, yet their outputs may carry systematic biases conditioned by the language of the prompt. This study presents an experimental comparison of LLM-generated political analyses of a Ukrainian civil society document, using semantically equivalent prompts in Russian and Ukrainian administered to two frontier models from different developers, ChatGPT 5.2 and Claude Opus 4.5. Despite identical source material and parallel query structures, both models diverged along the same axis: Russian-language outputs leaned toward delegitimizing framings, characterizing civil society actors as externally funded elites constraining a democratic mandate, while Ukrainian-language outputs treated the same actors as legitimate stakeholders in democratic contestation. The magnitude of this divergence, however, was model-dependent. ChatGPT's Russian output reproduced vocabulary characteristic of Russian state discourse; Claude Opus's stayed in a mainstream critical idiom and hedged its judgments in both languages. These findings demonstrate that prompt language alone can systematically shift the ideological orientation of an unchanged model analyzing identical content. The shift is a general property of multilingual LLMs whose severity, and whose alignment with propaganda narratives, varies across systems. The implications reach AI deployment in polarized information environments, cross-lingual research, and AI governance in multilingual societies.

16.
arXiv (CS.LG) 2026-06-12

Simultaneous Latent Budget Trees for Stratified Classification

arXiv:2606.13295v1 Announce Type: cross Abstract: In the era of Explainable Artificial Intelligence, there is a renewed focus on single trees for their ease of interpretation. This paper introduces Simultaneous Latent Budget Trees, a probabilistic machine learning framework for classification trees in the presence of a stratification factor such as a temporal, spatial, or demographic variable, acting as a control variable or potential confounder. Standard tree growth procedures are not designed to optimize a conditional split rule. A model-based split rule is proposed in which child nodes are interpreted as latent components of a simultaneous mixture model, such as the Simultaneous Latent Budget Model and its constrained versions, fitted to the parent node. Mixing parameters drive the observations, differently for each group, to the child nodes whereas latent budgets parameters update the response classes profile of each level of the control variable. Parameters are estimated by least squares considering a neural network perspective of the model. An informative tree structure can be interactively visualized with interpretation aids on the node and the paths, including visual pruning and decision tree selection procedure. Suitable measures are proposed to handle an unbalanced response class distribution. The proposed methodology is applied to investigate gender-related differences in disease progression of Amyotrophic Lateral Sclerosis. The SLBT library with the various tree-based algorithms is available in the linked GitHub repository.

17.
arXiv (math.PR) 2026-06-11

Sample Path Properties of the Fractional Wiener–Weierstrass Bridge II

arXiv:2606.11994v1 Announce Type: new Abstract: Fractional Wiener–Weierstrass bridges are a class of Gaussian processes obtained by replacing trigonometric functions in the construction of classical Weierstrass functions by fractional Brownian bridges. A number of their sample path properties were derived in Schied–Zhang (2024,2026). The analysis in these papers left several open questions, most of which are addressed here. Specifically, we prove that, in the regime in which the Weierstrass mechanism dominates the underlying fractional Brownian bridge, the limiting $b$-adic variation coefficient has an absolutely continuous distribution and is therefore genuinely random. At the critical point between the two roughness regimes, we establish the power-variation formula and the critical $\Phi$-variation limit conjectured in Schied–Zhang (2024). Finally, we derive the Hausdorff dimension for the graphs of the sample paths by proving a conjecture from Schied–Zhang (2026) for the missing high-Hurst case.

18.
arXiv (quant-ph) 2026-06-11

Numerically Optimizing Shortcuts to Adiabaticity: A Hybrid Control Strategy

arXiv:2604.01301v2 Announce Type: replace Abstract: Achieving fast, excitation-free quantum control is a vital challenge in modern quantum technologies. In many cases, shortcuts to adiabaticity enable fast adiabatic-like protocols, yet determining control parameters that satisfy practical constraints is often challenging in complex systems. Here, we combine an analytical shortcut to adiabaticity approach with several numerical optimization methods to boost the performance of the protocol. As a proof-of-principle for this hybrid approach, we study a particularly intricate control problem, the separation of two trapped ions. We show that this analytical-numerical approach, along with the physical insight gained through the variety of suboptimal solutions, leads to the exploration of new solutions in a complex landscape that yield improvements of up to 3 orders of magnitude. Moreover, this improvement comes with no additional cost from an experimental point of view.

19.
arXiv (math.PR) 2026-06-18

Probabilistic representation and classical solutions of wave equations with complex polynomial nonlinearities

arXiv:2606.18919v1 Announce Type: cross Abstract: We review the probabilistic representation of solutions of wave equations with polynomial nonlinearities in spatial dimensions d=1,2,3 using stochastic branching processes. Under regularity assumptions on the initial data, we derive conditions ensuring the integrability of the corresponding Monte Carlo estimator, and the existence and smoothness of mild and classical solutions. We also present numerical results and comparisons with grid-based algorithms for the solution of nonlinear wave equations.

20.
arXiv (CS.AI) 2026-06-16

NeuronFabric: A Software Reference Architecture for On-Chip Transformer Training with Local Adam

arXiv:2606.16440v1 Announce Type: cross Abstract: Publicly documented accelerator architectures generally separate training computation from optimizer-state updates or rely on external memory and host orchestration. This paper presents NeuronFabric, a software reference architecture intended for future FPGA and ASIC implementations of transformer training with local Adam updates. A complete C# prototype implements forward pass, backpropagation, and Adam optimization without external machine-learning frameworks. The goal is to validate numerical correctness and memory requirements before hardware implementation. The evaluated model is a 334K-parameter autoregressive transformer (d=88, H=4, f=264, L=4, vocab=256) trained on the Shakespeare corpus. The BF16W configuration achieves evaluation loss 1.5426 after 80K samples, compared with 1.5224 for an FP32 GPU reference, while producing coherent character-level text. The paper introduces BF16W, which stores weights in BF16 while retaining Adam optimizer moments in FP32. This reduces memory requirements for on-chip training. A 334K-parameter FP32 model with Adam moments requires approximately 4.0 MB, matching the BRAM capacity of a Xilinx ZCU102 device. The BF16W variant requires approximately 3.34 MB, leaving memory available for activation storage. We describe the vocabulary-budget constraint observed during earlier experiments, quantify BF16W memory savings, and outline FPGA training as the next stage of development. No FPGA measurements are included in this paper. This publication serves as a public architectural disclosure and software reference implementation for future FPGA and ASIC exploration of the NeuronFabric architecture.

21.
arXiv (CS.LG) 2026-06-16

DAL: A Practical Prior-Free Black-Box Framework for Piecewise Stationary Bandits

arXiv:2501.19401v5 Announce Type: replace Abstract: We introduce a practical, black-box framework termed Detection Augmented Learning (DAL) for the problem of piecewise stationary bandits without knowledge of the underlying non-stationarity. DAL accepts any stationary bandit algorithm with order-optimal regret as input and augments it with a change detector, enabling applicability to all common bandit variants. Extensive experimentation demonstrates that DAL consistently surpasses all state-of-the-art methods across diverse non-stationary scenarios, including synthetic benchmarks and real-world datasets, underscoring its versatility and scalability. We provide theoretical insights into DAL's strong empirical performance, complemented by thorough empirical validation.

22.
arXiv (quant-ph) 2026-06-11

Additivity and chain rules for quantum entropies via multi-index Schatten norms

arXiv:2502.01611v3 Announce Type: replace Abstract: The primary entropic measures for quantum states are additive under the tensor product. In the analysis of quantum information processing tasks, the minimum entropy of a set of states, e.g., the minimum output entropy of a channel, often plays a crucial role. A fundamental question in quantum information and cryptography is whether the minimum output entropy remains additive under the tensor product of channels. Here, we establish a general additivity statement for the optimized sandwiched Rényi entropy of quantum channels. For that, we generalize the results of [Devetak, Junge, King, Ruskai, CMP 2006] to multi-index Schatten norms. As an application, we strengthen the additivity statement of [Van Himbeeck and Brown, 2025] thus allowing the analysis of time-adaptive quantum cryptographic protocols. In addition, we establish chain rules for Rényi conditional entropies that are similar to the ones used for the generalized entropy accumulation theorem of [Metger, Fawzi, Sutter, Renner, CMP 2024].

23.
arXiv (CS.CL) 2026-06-12

Multi-Turn Reasoning When Context Arrives in Pieces: Scalable Sharding and Memory-Augmented RL

When a user reveals task-critical information across several conversation turns, LLM accuracy drops by up to 65% despite full context availability. We show that this Lost in Conversation degradation can be substantially mitigated by training models to maintain a compact rolling memory instead of attending to a growing history. To make such training scalable, we introduce a low-cost sharding pipeline that converts single-turn QA datasets into multi-turn fragmented-information episodes, eliminating the need for hours of manual annotation. Training only on sharded GSM8K, our memory-augmented policy significantly improves multi-turn accuracy and generalises zero-shot to harder math and out-of-domain long-context QA. Moreover, memory-trained models outperform full-history baselines even when given the full history at test time, suggesting that learning to compress induces more robust incremental reasoning than full-context exposure alone.

24.
medRxiv (Medicine) 2026-06-19

Extraction of Glaucoma Diagnosis, Type, and Severity from Clinical Notes using Secure Cloud-based Large Language Models

Purpose: To evaluate the performance of secure cloud-based large language models (LLMs) in extracting glaucoma diagnosis, type, and severity from free-text clinical notes in the electronic health record (EHR). Design: Retrospective chart review analysis. Participants: 1,250 subjects from the Bascom Palmer Ophthalmic Repository. Methods: Clinical notes of glaucoma-related encounters between 2014 and 2024 were extracted from the Bascom Palmer Ophthalmic Repository. Two fellowship-trained glaucoma specialists annotated clinical notes for glaucoma presence, type, and severity at the eye level. The dataset was split into development (10%), validation (10%), and test (80%) sets. Development and validation sets were used for prompt engineering and refinement, and the held-out test set was used for evaluation. Five LLMs (Claude Opus 4.6, DeepSeek-V3.2, GPT-5.2, Grok 4.1, and Qwen3.6-35B-A3B) were accessed via Azure AI Foundry within HIPAA-compliant containers. Model performance was assessed using standard metrics. Clinician-entered ICD-10 codes were also compared with adjudicated labels. Main Outcome Measures: Gwet AC1, accuracy, sensitivity, specificity, and F1-score. Results: Inter-grader agreement was high for glaucoma detection (Gwet AC1= 0.930 (95% CI: 0.917-0.945), type classification (Gwet AC1= 0.917 (95% CI: 0.904-0.930), and severity staging (Gwet AC1= 0.901 (95% CI: 0.884-0.916). For glaucoma diagnosis, LLMs demonstrated high overall accuracy, with Claude achieving 97.5%, DeepSeek 96.0%, GPT 96.2%, Grok 94.4%, and Qwen 95.5%. F1 scores for glaucoma detection ranged from 95.4% to 98.9% across models. For glaucoma type classification, accuracies were 97.1%, 94.2%, 94.2%, 94.0%, and 94.4% for Claude, DeepSeek, GPT, Grok, and Qwen, respectively. F1 scores for the most prevalent type (POAG) ranged from 96.3% to 98.9%. For severity staging, accuracies were 95.0%, 94.8%, 94.5%, 94.0%, and 95.2%, respectively, with F1 scores ranging from 89.7% to 96.3% across severity categories and models. ICD-10 codes demonstrated substantially lower performance for type and severity staging, with overall accuracies of 89.2% and 58.5%, respectively. Conclusions: Secure cloud-based LLMs accurately extracted glaucoma diagnosis, type, and severity information from free-text ophthalmology notes, achieving performance approaching expert clinician adjudication while substantially outperforming ICD-based phenotyping approaches, particularly for disease severity classification. These findings demonstrate the potential of LLMs to transform unstructured clinical documentation into scalable, research-ready phenotypic data for large-scale glaucoma cohort development and EHR-based ophthalmic research.

25.
arXiv (CS.CL) 2026-06-16

PathRouter: Aligning Rewards with Retrieval Quality in Agentic Graph Retrieval-Augmented Generation

Agentic GraphRAG trains language-model agents to iteratively retrieve and reason over graph-structured evidence, enabling more accurate and context-aware decision-making by efficiently navigating complex information networks. However, outcome-only reinforcement learning suffers from answer-path reward aliasing, where correct answers may come from shortcuts rather than useful evidence paths. It also exhibits search-update ambiguity, as scalar trajectory-level feedback does not indicate which retrieval actions to adjust. To mitigate these shortcomings, we present PathRouter, a path-aware training framework for agentic GraphRAG. PathRouter jointly evaluates each trajectory along answer correctness and evidence-path overlap, yielding four trajectory categories with differentiated GRPO advantage scaling that suppresses shortcut reinforcement while preserving evidence-seeking behavior. For evidence-poor trajectories, a frozen gold-evidence teacher provides token-level KL guidance on reasoning and search-query tokens, excluding answer tokens to avoid direct response imitation. Experiments on six QA benchmarks across three model sizes show that PathRouter consistently improves answer F1 and evidence-path overlap, achieving average F1 gains of 3.1 on 3B and 4.9 on 7B models compared to a strong baseline.