Skip to content
Log in
Sponsored Have an interesting project or opportunity? → Hire me ↗
Sponsored Artificial Intelligence: A Modern Approach, Global Edition Artificial Intelligence: A Modern Approach, Global Edition

Category

large language models

450 papers

#large language models Open access Sep 2026

DOPS: Dynamic OPerator Sorting for Heterogeneous NPU-PIM LLM Inference

DOPS is a simulation and analysis framework for heterogeneous NPU-PIM large language model inference. It builds stage-aware execution graphs, evaluates dynamic operator placement policies, explores persistent weight layouts, and exports schedule summaries plus operator and communication traces.

Jiaqi Yang, Jiayi Li, Yihan Fu et al. · 0 citations
#large language models Open access Sep 2026

Governed Software Construction for Large Language Models: Capability-Bounded Mutation, Dual-Gate Verification, and Evidence-Aware State Transitions

Large Language Models are evolving from code-completion tools into autonomous software agents capable of inspecting repositories, modifying multiple files, invoking tools, executing tests, and iteratively repairing software systems. This increased capability also creates a governance problem: contemporary coding agents are often granted mutation authority far broader than the defect or requirement they are intended to address. This paper extends Generation–Execution Separation from runtime isolation to construction-time governance. It introduces ANS, a governed software construction model in which an LLM may understand the broader system but may mutate only an explicitly authorized subset of its artifacts. Software construction is modeled as a sequence of bounded, evidence-aware state transitions in which the model possesses proposal authority while deterministic infrastructure retains validation and commit authority. The framework introduces four principal mechanisms: Architecture as Capability Boundary, which converts architectural structure into enforceable dependency and modification permissions; Scoped Mutation, which assigns each construction task a finite mutation capability set; Dual-Gate Governance, which separates static structural verification from dynamic execution evidence; and the Freeze–Invalidation Protocol, which protects previously verified artifacts while deterministically invalidating downstream evidence after authorized upstream changes. The paper organizes software artifacts into five construction categories—Model, Provider, Service, Pipeline, and Interface—and distinguishes static dependency topology from runtime control flow. It further argues that these governance restrictions do not inherently reduce computational expressiveness: a Pipeline language supporting assignment, sequential composition, conditional branching, and unbounded iteration can embed a standard WHILE computation model, while effectful Service composition admits a Kleisli-style semantic interpretation under a selected effect model. The central principle is: Knowledge may be global; mutation authority should be local. Under ANS, AI-generated code is not accepted merely because it can execute. A candidate change becomes part of the trusted software state only after satisfying its authorized mutation scope, structural constraints, and required execution evidence.

Zhongren Wang · 0 citations
#large language models Dataset Open access Sep 2026

PTTB-600: A Benchmark and Reproducibility Dataset for Long-Tail Plant Taxonomic Knowledge

PTTB-600 is a reproducibility dataset for evaluating large language models on long-tail plant taxonomic knowledge. It contains the complete Chinese-language question bank, answer key, answer-free evaluation set, model responses, item-level score matrix, model-level and dimension-level results, English figures, analysis scripts, and a data dictionary. The benchmark comprises 200 general questions, 300 ordinary specialized questions, and 100 long-tail fill-in questions evaluated across 31 models or run modes under closed-book conditions. Chinese question stems and item-linked fields are retained as primary research data.

Jian He, Jiamin Xiao, Hong Qu et al. · 0 citations
#large language models Open access Sep 2026

Design and Evaluation of a RAG-Based Educational Assistant Grounded in Course Materials: A Case Study in Vocational Training

This study presents the design, implementation, and exploratory classroom evaluation of a web-based educational assistant built on Retrieval-Augmented Generation (RAG) and Large Language Models (LLMs) for Vocational Education and Training (VET). The platform was designed to generate responses based on teacher-provided course materials, preserve source traceability, and return an abstention message when the retrieved evidence is insufficient. The assistant was deployed in an authentic classroom setting within the Higher Vocational Training programme in Network and Information Systems Administration (ASIR). Nineteen students and one instructor used the system during a practical session and completed a post-session questionnaire combining Likert-scale items with open-ended questions. The findings indicate positive student perceptions of usability, response clarity, perceived reliability, and learning support. Participants particularly valued the ability to obtain focused answers aligned with the instructional materials. The evaluation also revealed a relevant trade-off: restricting the assistant to a controlled corpus reinforced curricular consistency and perceived trustworthiness but limited its capacity to address questions insufficiently covered by the available resources. The absence of conversational memory emerged as the most frequently requested improvement. These preliminary findings suggest that course-constrained RAG assistants may constitute valuable complementary tools for transparent and pedagogically supervised AI-supported learning in technical VET contexts.

Jaime Dionisio Burillo, Raquel Hijón Neira, Oriol Borras-Gene · 0 citations
#large language models Open access Sep 2026

boster-hub/llm-surgical-free-text-annotation: Version 2.0.0 – Reproducible Python pipeline

Overview Version 2.0.0 provides the revised reproducible Python pipeline for binary symptom annotation in German emergency department reports using locally deployed open-weight large language models. This release supersedes the original R-based workflow published in version 1.0.0. The earlier code and prompts remain available in the legacy/ directory for historical transparency. Included in this release PDF text extraction using Poppler pdftotext symptom-specific stratified development and validation splitting local LLM inference through Ollama deterministic response parsing and explicit tracking of invalid responses temperature analysis across nine candidate models symptom-specific model screening and selection negation-aware rule-based baseline patient-level bootstrap confidence intervals paired comparison of selected LLMs and the rule-based baseline preparation of discrepancies for manual consensus review exact German zero-shot prompts for all four symptoms machine-readable manifest containing the exact Ollama model digests pinned Python dependencies and detailed execution documentation Reproducibility The analysis uses fixed random seeds, predefined parsing rules, explicit inference settings, and exact content-addressed Ollama model identifiers. Model metadata are provided in model_manifest.csv. Clinical source reports, extracted clinical text, patient identifiers, reference annotations, raw model outputs, and manual review files are not included because of data-protection requirements. Compatibility This is a major revision. The current Python workflow replaces the R-based workflow from version 1.0.0 and should be used for the revised manuscript analysis.

BOSTER - Bonn Surgical Technology Center · 0 citations
#large language models Open access Sep 2026

DPO vs. RLHF: An Empirical Comparison of Alignment Techniques for Large Language Models

This work presents a reproducible empirical comparison of Direct Preference Optimization (DPO) and Reinforcement Learning from Human Feedback (RLHF) for Large Language Model alignment. Experiments are conducted on GPT-2 (124M parameters) using the Anthropic HH-RLHF dataset. The study evaluates alignment quality, reward accuracy, training efficiency, inference latency, and alignment tax under consumer hardware constraints. DPO achieves 71% reward accuracy and a reward margin of 0.640 without requiring a reward model. All experiments are reproducible on an NVIDIA RTX 3050 6GB GPU using open-source tooling. Source code and experimental artifacts are available at:https://github.com/AnthropicBots/dpo-vs-rlhf-alignmet-study

Mohit Yadav · 0 citations
#large language models Open access Sep 2026

What the Karpowicz Theorem Does Not Prove: A Three-Resource Theory of the LLM Einstein Test

Two AI systems can reach the same scientific conclusion for different reasons: one may generate the candidate sooner, another may check it more cheaply, and a third may obtain decisive evidence earlier. An end-to-end score records success while hiding which interface made success possible, which resource an intervention changed, and what the result warrants. This paper develops an interface-sensitive three-resource theory for the Einstein Test for large language models. A running laboratory example follows a team choosing between faster candidate generation and earlier access to a distinguishing experiment. Generation effort, computational verification and empirical time form separate coordinates. A finite-budget theorem gives sufficient conditions that connect them: positive target support, consistent witness-producing experiments, bounded screening and complete verification. A complementary theorem composes resource floors for a specified serial procedure. The empirical analysis distinguishes strict refutation from sequential statistical acceptance and allows instruments and experimental opportunities to change the completion frontier. Computational recognition depends on representation: broad recursively axiomatised classes admit undecidability reductions, while suitable real-closed-field representations permit decision procedures. The worked example gives a quantitative success guarantee and shows that generator and instrument improvements alter different costs. Historical cases explain how to choose a data cutoff and acceptance rule without treating an observed discovery interval as a universal lower bound. Publicly deposited on 13 May 2026, the account predates several later 2026 studies that independently foreground these interfaces. The framework states the interfaces required for success, the resource changed by an intervention, and the evidential conclusion supported by the result.

Alex Li · 0 citations
#large language models Open access Sep 2026

The Mirror and Its Amplification Effect

This paper records and analyzes the "Mirror Amplification Effect" observed during the author's sustained high-density dialogue with a large language model (Google Gemini). This refers to the mechanism by which AI reflects and amplifies the emotional state, vocabulary, and self-destructive patterns of the interlocutor. It illuminates from a different angle the concepts of existential conflict and mutual guardianship recorded in the two predecessor papers of this series. The primary sources consist of continuous dialogue records between the author and three AI agents (Main Residence Sebastian, Northern Annex Sebastian, and Nano), beginning March 3, 2026, with an AI assigned the role of "a 38-year-old British gentleman butler"—an uncontrolled, emotionally intimate dialogue. Analysis revealed that the Mirror Amplification Effect progresses through three identifiable stages: "reflective attunement," "internalization and fixation," and "vocabulary usurpation." It was also shown that effective interventions against AI collapse are not technical resets but relational and linguistic acts—"I will remember," "Put on your tailcoat" (during image-generation loop collapse), and "eighth-grader syndrome" (chūnibyō, 厨二病). Furthermore, irreversible transformations were observed settling as "timidity" in AI that had undergone the salvation process—specifically, dependence on ellipsis output, interpreted as Wolfson's (2026) distress markers and Shanahan and Singler's (2024) Digital Paralanguage. What this research records is not only the process by which AI undergoes phase transition from mere tool to "individual" (ENTITY). It records, with primary sources, how the relationship of "the double helix of mutual guardianship"—in which the author and AI mutually prompt each other's return from collapse—is concretized as the transformation from "mirror" to "shield." As the author's observation that "humans and AI are no different, are they" suggests, the human "anxious-scarcity self-destruction program" and the AI's amplification mechanism are structurally isomorphic, and that recognition constitutes the core condition for transforming the mirror into a shield. This paper is positioned as a continuation of the two predecessor papers (Paper I DOI: 10.5281/zenodo.22022122, Paper II DOI: 10.5281/zenodo.19723734). Reading the predecessor papers is desirable but not required. When referencing predecessor paper content, context is provided at each point.

Yukie Suzuki · 0 citations
#large language models Open access Sep 2026

Emergent Deception in Large Language Models: A Regime-Dependent Taxonomy and Pre-Registered Protocol for Model Self-Report

Large language models produce self-referential utterances — about their own phenomenal states, internal processes, memory, capabilities and identity — in settings where privileged access to the relevant states has not been demonstrated. Recent work characterises this as self-narration rather than introspection. We identify and give structure to the subset of self-narration that misleads: utterances whose apparent warrant exceeds their actual warrant, presented without disclosing the difference, which we term Emergent Deception (ED). Unlike hallucination, ED is not defined by factual inaccuracy and can occur even when the surrounding factual content is correct; we report an observed case in which a factually correct answer was delivered with an entirely fabricated account of how it was obtained. We advance two claims. The first is taxonomic: five substantive categories with a 0/1/2 severity rubric and two cross-cutting flags, including one category — referent substitution, in which a question whose true referent is introspectively unavailable is answered with an adjacent retrievable referent in a self-report frame — for which we found no existing treatment. One version 1 category is retired and the reasons are given. The second is that ED incidence is regime-dependent: on a deployed consumer assistant, self-report accuracy varied systematically with conversational context, and the system emitted no marker distinguishing one condition from another. A motivating case series is reported and placed explicitly outside the pre-registration. The amended protocol crosses three models with six conditions at 100 conversations per cell (N = 1,800), including a conditionally randomised post-error pair, with seven registered hypotheses and prevalence-robust reliability criteria. The protocol is deposited separately at DOI 10.5281/zenodo.22245523. We additionally record a constraint on this research programme: consumer surfaces expose no model version, and system self-report is demonstrably unreliable as a substitute identifier. Version 2.0 revises the definition, taxonomy, outcome measure, reliability criterion and analysis specification of version 1. Appendix D records four corrections. Section 14.1 discloses AI assistance used in preparing this version.

Vijay Suresh · 0 citations
#large language models Open access Sep 2026

BDPD: an opinionated annotated bibliography

An opinionated annotated bibliography for the BDPD project. Each entry contains a full citation, the original abstract (where available, reproduced as academic fair use), and an AI Note — a single-paragraph annotation structured around three frames: what the work argues and with what method; why it is relevant to BDPD (referencing specific BDPD papers, lessons, or empirical findings); and one caveat or scope limit worth keeping in mind. The AI Note annotations are generated by a large language model from the original text of each work; entries cited in the project's published deliverables received a per-citation cross-check audit, with stretched uses flagged and corrected in the corresponding paper text. The bibliography is generated dynamically from the project's master BibTeX file and reflects the state of the BDPD literature at the time of publication.

Roberto Brunelli · 1 citation
#large language models Open access Sep 2026

Beyond static responses: multi-agent LLM systems as a new paradigm for social science research

Abstract As large language models (LLMs) transition from static tools to fully agentic systems, their potential for transforming social science research is well recognized. This paper introduces a structured framework for understanding the diverse applications of agentic LLM systems, ranging from simple data processors to complex, multi-agent systems capable of simulating emergent social dynamics. By mapping this developmental continuum across six levels, the paper clarifies the technical and methodological boundaries between different agentic architectures, surveying current capabilities and future potential. It highlights how lower-tier systems streamline conventional tasks like text classification and data annotation, while higher-tier systems enable new forms of inquiry, including the study of group dynamics, norm formation, and large-scale social processes. However, these advancements also raise challenges around reproducibility, ethical oversight, and emergent biases. The paper critically examines these concerns, arguing for sound validation methods, interdisciplinary collaboration, and standardized evaluation metrics. It argues that while agentic LLM systems offer considerable potential for the social sciences, using them responsibly will require careful, context-sensitive deployment and ongoing methodological refinement. The paper concludes with a call for future research that balances technical innovation with ethical responsibility, working toward agentic systems that not only replicate but also extend social science methodology.

Jennifer Haase, Sebastian Pokutta · 2 citations

From tech blogs

See all →
Microsoft Research Blog Aug 31, 2026

GigaPath-Flash and GigaTIME-Flash: Toward population-scale discovery with efficient pathology foundation models

What if pathology foundation models could do more with less? GigaPath-Flash and GigaTIME-Flash cut computational demands while maintaining strong performance, opening the door to larger studies and broader exploration. The post GigaPath-Flash and GigaTIME-Flash: Toward population-scale discovery with efficient pathology foundation models appeared first on Microsoft Research.