Large Language Models (LLMs) offer new opportunities for automated code refactoring. However, generated changes must reduce targeted quality problems without introducing new issues or altering behaviour-relevant code structures. We introduce REFINE (Refactoring with Evidence-aware Flow for Integrated ageNtic Execution), a tool-agnostic, evidence-aware multi-agent approach for generating Java file-level refactoring candidates. REFINE combines static-analysis-guided smell identification, smell-informed planning, LLM-based transformation, automated re-analysis, preservation checks, and structured reporting. We evaluate REFINE on 450 Java files from 15 open-source systems, producing 1,350 model-pass outputs using OpenAI GPT-5.5, Google Gemini 3.1 Pro Preview, and Anthropic Claude Opus 4.8. REFINE reduces detected code smells by 68.26%, 72.79%, and 68.49% across the three configurations, respectively, with the strongest reductions observed for major smells. A matched 150-file direct-prompt baseline shows that REFINE achieves a higher median code-smell reduction with smaller edits and fewer public-method removals. However, broader quality improvements are inconsistent, and preservation checks reveal residual risks, including assert/fail-call changes and public-method removal. Therefore, REFINE outputs should be treated as refactoring candidates requiring compilation, testing, dependency analysis, and human review before adoption in repository- or system-level settings.
Muhammad Waseem, Aakash Ahmad, Pekka Abrahamsson· 0 citations
This paper investigates how multi-agent systems (MAS-based on large language models) can support actuarial risk modelling, with a particular focus on uncertainty quantification, and proposes a novel approach to uncertainty propagation using token-level log-probabilities and a Bayesian Network.
Bart Custers, K. Aslansefat· Lecture notes in computer sc...· 0 citations
Part IV of a series on equity and reliability in medical imaging AI. Part III of this series showed that naming a demographic group in the text prompt of a medical vision–language model changes its diagnosis on a fixed chest radiograph, that the perturbation is exactly rank one in the standard positive/negative pair readout, and that group-specific thresholds therefore provably cannot remove it. Two questions were left open: whether any of this survives a change of imaging domain, and whether it matters for a model that is actually good at its task — Part III's strongest encoder reached a mean AUC of 0.687. This paper audits four pathology vision–language models (PLIP, QuiltNet-B-32, CONCH, and the slide-level TITAN) under the descriptor set from Part III, unchanged, and combines them with Part III's three radiology encoders for a seven-model comparison spanning two imaging domains and both patch and slide granularity. On TITAN the audit runs at full strength: 8,010 whole-slide images from 6,582 patients across 25 TCGA cancer types, every interval clustered on patient, mean neutral-prompt AUC 0.982 with all 25 projects clearing the utility floor. Four results. The rank-one identity transfers exactly — verified to 5.2×10−8 on real slide embeddings, two orders of magnitude tighter than Part III because TITAN has no learned logit scale to amplify float error. Part III's threshold-irremovability corollary therefore holds for whole-slide pathology. Diagnostic competence does not protect against prompt-channel bias. Across seven encoders, neutral-prompt AUC does not predict the effect (r = −0.443, p = 0.32): CONCH at AUC 0.849 shows 21.1 points of mean absolute excess false-negative rate, indistinguishable from OpenAI CLIP at AUC 0.518. What predicts it is a purely geometric quantity, the descriptor perturbation measured against the diagnostic contrast (r = +0.780, p = 0.039), reported as a hypothesis with n = 7 rather than a law. The model answers the token and not its truth, including for race. Part III could test congruence only for sex and age, because NIH ChestX-ray14 records no race; TCGA does, and the difference-in-differences estimator gives a median 0.0045 logits for race and 0.0010 for sex, against a specimen-site positive control 164–730× larger in 4 of 4 pairs at tighter precision. Anatomically impossible descriptors — "a male patient showing endometrial carcinoma" — move the score as much as coherent ones. Prompt symmetrisation is a domain-dependent intervention rather than a general one: it removes 59–63% of the effect on radiology encoders, 26–34% on pathology patch encoders, and amplifies the effect by 94% on TITAN. The text-side symmetry residual separates the two domains completely (Mann–Whitney p = 0.029). The practical consequence is a correction to how this series has framed mitigation. A deployer cannot infer prompt-channel robustness from a model's accuracy, and cannot assume a mitigation validated in one imaging domain transfers to another — including on the strength of Part III. Honest scope. Unlike Parts II, IIb, and III, this part is not preregistered; PROTOCOL_PART4.md is retrospective and says so in its first line, recording which hypotheses preceded their endpoints and which did not. Only TITAN carries interval estimates: the public CRC-VAL-HE-7K mirror records no patient identifiers, so the three patch-level encoders are reported as point estimates only. The anatomically-impossible result rests on five cells and is a demonstration of mechanism, not an estimate of magnitude. This paper also supplies the out-of-domain evidence behind amendment 8 to Part III's protocol, which withdrew Part III's proposed text-only predictor of symmetrisation failure. Files. The manuscript PDF (21 pp, 4 figures); the retrospective protocol; and an artifact archive containing all 13 analysis scripts and every result table as CSV/JSON. Model weights are not redistributed — TITAN, CONCH, and UNI are CC-BY-NC-ND-4.0 and were obtained through the gated Hugging Face process. Ethics. No patient data was collected. TCGA is a public consortium dataset; CRC-VAL-HE-7K is public and de-identified. This is an audit of a model property, not a clinical study, and makes no claim about patient outcomes.
Omar Mohammed· Zenodo (CERN European Organi...· 0 citations
Measured board power, tokens per joule, and electricity cost per million tokens for local large language model inference on a retail NVIDIA RTX 5080. Board power was logged with nvidia-smi at 1 Hz while driving fixed-length generations. Efficiency. Llama 3.2 3B 1.20 tokens/joule; sparse gpt-oss 20B 0.665; Qwen 2.5 7B 0.53; Qwen 2.5 14B 0.317. The sparse 20.9B model is approximately twice as efficient per joule as the dense 14B, indicating that architecture and quantization dominate parameter count on the efficiency axis. Loaded power. Board power medians of 265-344 W against a 448 W stock limit that was never reached, at 82-93% utilization and 49-52 degrees C. Idle finding. Across a week of captures the card idled at 52-71 W at the Windows desktop, pinned in P0 with graphics clocks near 2.9 GHz; the cleanest achievable state still read 53.7 W. Published review figures typically quote single-digit to 15 W idle. At the May 2026 EIA US residential average of $0.184/kWh, 52-71 W continuous is 456-622 kWh, or approximately $85-115 per year before any tokens are generated. Since generation itself costs only $0.04-$0.16 per million tokens, idle behaviour rather than model choice dominates the operating cost of an intermittently used inference node. Includes raw 1 Hz telemetry captures in addition to summary rows. Canonical page, full method and change log: https://techfuelhq.com/data/rtx-5080-llm-power-efficiency/
Wood, LK, IV· Zenodo (CERN European Organi...· 0 citations
Reach audiences
Advertise in front of researchers, engineers, and readers.
The Transformer architecture---built on attention rather than recurrence---redrew the landscape of natural language processing and became the substrate of contemporary artificial intelligence. This article presents a narrative review of the architecture's canonical line: Sutskever and colleagues' 2014 sequence-to-sequence learning, Bahdanau and colleagues' 2015 attention alignment, Vaswani and colleagues' 2017 Attention Is All You Need, Devlin and colleagues' 2019 BERT pretraining, Radford and colleagues' 2019 GPT-2, Brown and colleagues' 2020 GPT-3 and few-shot learning, Raffel and colleagues' 2020 T5 transfer, Dosovitskiy and colleagues' 2021 Vision Transformer, Bommasani and colleagues' 2021 foundation-model framing, Hoffmann and colleagues' 2022 Chinchilla scaling laws, Ouyang and colleagues' 2022 InstructGPT alignment, and Touvron and colleagues' 2023 LLaMA openness. The synthesis is organized around three themes: architecture, in which self-attention's parallel sequence processing replaced recurrence and enabled scale; scaling, in which pretraining on text plus parameter growth yielded emergent few-shot capability and then compute-optimal correction; and alignment and access, in which instruction tuning, RL from feedback, and open weights reshaped capability's deployment. It is concluded that the Transformer is machine learning's most consequential architecture to date---its attention mechanism the field's new inductive bias---and that scaling's economics and governance now define its trajectory.
Zen Revista, 10 IA· Zenodo (CERN European Organi...· 0 citations
This document presents the defensible core of the Universal Model Framework (UMF), isolating the minimal set of structural assumptions and derivations that remain logically coherent, mathematically motivated, and empirically falsifiable. As stated in the text, the goal is to extract “the smallest segment that is logically structured, mathematically motivated, and empirically vulnerable,” while ensuring that “every load‑bearing claim is paired with an explicit failure condition.” It is a deliberately falsifiable research program investigating whether quantum structure, arithmetic regularity, and emergent spacetime geometry can arise from a common relational foundation. It separates three logically distinct questions: whether relational systems can reconstruct quantum-theoretic structure; whether ordinary prime-number organization is physically selected rather than merely mathematically available; and whether a stable continuum geometry with causal and gravitational dynamics can emerge under refinement. The work reports exact finite results for recursive graph constructions, discrete geometry, cochain-based fermionic operators, local frames, symmetry tests, and numerical-reproducibility controls, while documenting failed frame-transport and continuum candidates. Crucially, it does not claim established fundamental physics: no continuum limit, Lorentzian causal structure, gravitational field equation, physical mass scale, complete quantum reconstruction, or prime-specific empirical signal has yet been derived. The framework’s contribution is therefore methodological as well as mathematical: it provides a transparent architecture for distinguishing theorem, model assumption, numerical fit, negative result, and falsifiable prediction in foundational physics. This project was developed by Marco Gericke, with structured assistance from a large language model. All scientific concepts and conclusions were generated, verified, and interpreted by the author. Dedicated to Peter Plichta, who envisioned the code before it could be computed.
Marco Gericke· Zenodo (CERN European Organi...· 0 citations
Corrected analysis and figure-generation package for a benchmark of six large language models, four single-model prompting strategies, and a separate GPT-5.4-assisted external-review workflow across 119 de-identified gastrointestinal oncology cases.
Jiajia Yuan, Shengyuan Zhou, Guangzhengao Yang et al.· Zenodo (CERN European Organi...· 0 citations
Intermediate-task training---fine-tuning a pretrained model on an intermediate task before fine-tuning again on the target task---often improves model performance substantially on language understanding tasks in monolingual English settings. We investigate whether English intermediate-task training is still helpful on non-English target tasks. Using nine intermediate language-understanding tasks, we evaluate intermediate-task transfer in a zero-shot cross-lingual setting on the XTREME benchmark. We see large improvements from intermediate training on the BUCC and Tatoeba sentence retrieval tas Research goal: How does the computational efficiency of English intermediate-task fine-tuning compare to target-language fine-tuning for zero-shot cross-lingual transfer, measured in terms of total training time and FLOPs on XTREME-R tasks? Autonomous synthesis report generated by Assignee Research. Tribunal consensus score: 7.7/10.
Assignee Research· Zenodo (CERN European Organi...· 0 citations
Background: Multidisciplinary tumor boards (MTBs) are central to contemporary head and neck oncology, ensuring accurate staging, guideline-concordant therapy, and balanced functional outcomes. In parallel, large language models (LLMs) have demonstrated increasing competence in synthesizing complex clinical data and generating structured recommendations. Their potential role as decision-support tools in head and neck oncology, however, remains insufficiently evaluated.Methods: We retrospectively compared treatment recommendations generated by two state-of-the-art LLMs (ChatGPT ™ and Google Gemini ™) with consensus decisions from a multidisciplinary tumor board. Fifty consecutive, synthetic head and neck cancer cases discussed between April and June 2025 were included without restriction on tumor type or stage. Both models received identical anonymized clinical, radiologic, and histopathologic reports and were prompted to generate guideline-based first- and second-line treatment recommendations
Data, field notebooks and code supporting a study of two video-based vehicle counting campaigns on a single arterial corridor of Ciudad Juárez, Mexico, and of how a classification error propagates into a road traffic noise model of the FHWA/REMEL family.The deposit contains the full adjudication register of the 105 image crops that a pretrained detector flagged as heavy vehicles, adjudicated twice and independently — once with the assistance of a large language model and once by the author under blind conditions, in two tranches — together with the blind answer sheets and keys, the measured-versus-modelled evaluation over 126 pairs from six continuously monitored sites, the vehicle registry tabulations used to characterise the local fleet, the calibration notebooks of both campaigns, and the code that reproduces every figure.The author's labels are the reference throughout. Agreement between the two passes was 30 of 30 on a pre-registered random subsample (Cohen's κ = 1.00) and 98 of 105 over the full census (93.3 %, κ = 0.586); the seven discrepancies are itemised in the accompanying paper.Video material and the original crops are deliberately not deposited: the capture protocol prohibits retaining legible licence plates, and the crops contain them. Every image reproduced in the paper had its plates blurred. The adjudication register makes each verdict auditable without distributing any image with a plate.
Pedro Martínez Cisneros· Zenodo (CERN European Organi...· 0 citations
This repository contains the complete dataset, experimental inputs, raw outputs, statistical scripts, and validation artifacts for the study assessing the effectiveness of Large Language Models (LLMs) in generating abuser stories from user stories under a constrained-context baseline. The study evaluated three lightweight models (GPT-4o mini, Claude 3.5 Haiku, Gemini 1.5 Flash) using two prompting techniques (Zero-Shot and One-Shot) across 10 real-world user stories.
Anonymous· Zenodo (CERN European Organi...· 0 citations
The sentence “the strength of this material is 250 MPa” does not settle when it breaks. This paper asks what does settle it──the answer is that there are three ways to break and a different thing decides each. Buckling is decided by shape, fracture by a flaw, and yielding by the crystal. No new mathematical theorem and no new law is claimed. Scope of this paper (scope note): No new mathematical theorem and no new law is claimed──Euler's buckling load, slenderness, Griffith's fracture condition and the yield stress are all standard. No mechanics of materials is built──what is used is three formulas and one comparison. No value fit for design is given──no safety factor, no initial imperfection and no residual stress is included. The numbers are for seeing which of the three limits is lowest, not values for design. The fence on the elastic constants is not treated──Paper 261 treats -1<nu<1/2. This paper is what happens after one is seated inside that fence, up to breaking. No theory of plasticity is built──yielding is treated as one stress value, with no hardening and no flow rule. No value of the surface energy is claimed──gamma=1.0 J/m^2 is a posited value, not a measurement on a particular material. It is used to see orders of magnitude. Fatigue and time dependence are not treated──only a single loading is examined. Griffith is not confused with another──the Griffiths appearing in Paper 173 is an algebraic geometer and a different person from the A. A. Griffith of this paper. Relation to earlier papers: Paper 261 wrote that what raises the fence on the elastic constants is the positive definiteness of the energy──this paper treats what follows, namely where a material seated inside that fence breaks. Paper 190 measured rare on a logarithmic scale──this paper likewise writes the effect of a flaw as a square root and in orders of magnitude. Paper 196 counted “pressure” as four different quantities──this paper counts that “strength” is not even one quantity. Paper 201 counted “complete” as four different claims──the same shape of roll call. What is added is computing that the buckling stress of one section moves from 164.3 to 18.3 MPa on changing only the length, putting the switching slenderness at the concrete value 88.8577, confirming that a flaw tells as sqrt1000=31.6228, and setting the three limits in one table and writing that the lowest is the actual limit. First, change only the length of one section. A square steel column of side 31.6 mm buckles at 164.3 MPa when it is 1 m long and at 18.3 MPa when it is 3 m──nothing about the material has been changed (Section 2). Second, this is the core of the paper. The slenderness at which the mode changes is lambda_c=pisqrtE/sigma_y=88.8577──slimmer than this and buckling comes first, stubbier and yielding does, so one material has its limit exchanged (Section 3). Third, the third limit is set by a flaw. By Griffith's condition a flaw of 1 mum gives 356.8 MPa and one of 1 mm gives 11.3 MPa (Section 4). Fourth, a flaw tells as a square root. A flaw 1000 times larger lowers the strength by a factor of 31.6228──which is sqrt1000 itself (Section 4). Fifth, the three do not compete. Of three upper bounds that hold at once, the lowest is the actual limit──and which is lowest is decided not by the material but by shape and flaw (Section 5). Sixth, so strength is not a property of the material. The number sigma_y=250 MPa is held, and a column 3 m long still breaks at 18.3 MPa──a factor of 13.66 apart (Section 5). the sentence “the strength of this material is 250 MPa” did not settle when it breaks. There are three ways to break and a different thing decides each──buckling by shape, fracture by a flaw, yielding by the crystal. The three do not compete, and the lowest is the actual limit. Triple the length of one section and the buckling stress falls to a ninth; admit an invisible flaw of 10 mum and the fracture stress drops below half the yield. For a column 3 m long the 250 MPa on the data sheet stands a factor of 13.66 above the real limit and never gets its turn. One thing separates them──writing down which limit is being counted. Write it down, and the occasions for changing the material separate from those for changing the shape and those for removing the flaw. Do not write it down, and one goes on looking for a stronger steel for a column that breaks at 18.3 MPa. On the making of this work: The ideas and content of this work stem from the author's own considerations. Assistance from an AI (a large language model) was used for structuring, English translation, and checking the algebra. Any remaining errors or misinterpretations are solely the author's. Feedback and corrections are sincerely appreciated. ----- 「この材料の強度は 250 MPa である」という一文は、壊れる条件を決めていない。本稿が問うのは、では何が決めているのかである──答は、壊れ方が三つあり、それぞれ別のものが決めているである。座屈は形が、破壊は傷が、降伏は結晶が決める。新しい数学定理も新しい法則も主張しない。 本稿の射程(射程注記):新しい数学定理も新しい法則も主張しない──オイラーの座屈荷重、細長比、グリフィスの破壊条件、降伏応力は、いずれも標準的である。材料力学を作らない──使うのは三つの公式と、一つの比較だけである。設計に使える値を与えない──安全率も、初期不整も、残留応力も入れていない。数値は三つの限界の大小を見るためのものであり、実際の設計値ではない。弾性定数の柵を扱わない──論文261 が -1<nu<1/2 を扱う。本稿は柵の中に座ったあと、壊れるまでの話である。塑性論を作らない──降伏を一つの応力値として扱い、硬化も流れ則も扱わない。表面エネルギーの値を主張しない──gamma=1.0 J/m^2 は置いた値であり、特定の材料の測定値ではない。桁を見るために使う。疲労と時間依存を扱わない──一回の載荷だけを見る。グリフィスは別人と混同しない──論文173 に現れる Griffiths は代数幾何学者であり、本稿の A. A. Griffith とは別人である。既刊との関係:論文261 は弾性定数の柵を立てているのがエネルギーの正定値性だと書いた──本稿はその先、柵の中に座った材料がどこで壊れるかを扱う。論文190 は「稀」を対数の目盛りで測った──本稿も、傷の効き方を平方根と桁で書く。論文196 は「圧力」が四つの別の量であることを数えた──本稿は「強度」が一つの量ですらないことを数える。論文201 は「完備」が四つの別の主張であることを数えた──同じ形の点呼である。加えたのは同じ断面で長さだけを変えて座屈応力が 164.3 から 18.3 MPa まで動くことを計算したこと、切り替わりの細長比を 88.8577 と具体的な数で出したこと、傷の効き方が sqrt1000=31.6228 であることを確かめたこと、三つの限界を一つの表に並べ、最低のものが実際の限界になると書いたことである。 第一に、同じ断面で長さだけを変える。一辺 31.6 mm の正方形鋼柱は、1 m なら 164.3 MPa、3 m なら 18.3 MPa で座屈する──材料は一切変えていない(第2節)。 第二に、これが本稿の芯である。切り替わる細長比は lambda_c=pisqrtE/sigma_y=88.8577 である──これより細長ければ座屈が先、太短ければ降伏が先で、同じ材料で限界が入れ替わる(第3節)。 第三に、三つ目の限界は傷が決める。グリフィスの式では、傷が 1 mum で 356.8 MPa、1 mm で 11.3 MPa になる(第4節)。 第四に、傷は平方根で効く。傷を 1000 倍にすると強度は 31.6228 分の 1──sqrt1000 そのものである(第4節)。 第五に、三つは競合しない。同時に成り立つ三つの上限のうち、最も低いものが実際の限界になる──どれが最低かは、材料ではなく形と傷が決めている(第5節)。 第六に、だから強度は材料の性質ではない。 sigma_y=250 MPa という数を持っていても、長さ 3 m の柱は 18.3 MPa で壊れる──その差は 13.66 倍である(第5節)。 「この材料の強度は 250 MPa である」という一文は、壊れる条件を決めていなかった。壊れ方が三つあり、それぞれ別のものが決めているからである──座屈は形、破壊は傷、降伏は結晶。三つは競合せず、最も低いものが実際の限界になる。同じ断面で長さを 3 倍にすれば座屈応力は 9 分の 1 になり、10 mum の見えない傷が入れば破壊応力は降伏の半分以下に落ちる。材料表の 250 MPa は、長さ 3 m の柱については実際の限界の 13.66 倍上にあり、一度も出番が来ない。分けるものは一つ──どの限界を数えているのかを書き出すこと。書き出せば、材料を替えるべき場面と、形を変えるべき場面と、傷を消すべき場面が分かれる。書き出さなければ、18.3 MPa で壊れる柱に、より強い鋼を探し続けることになる。 作成にあたって:本稿の着想と内容は、著者自身の考察に基づくものです。文章の構成整理や英訳、数式の確認には AI(大規模言語モデル)の助力を得ました。最終的な内容の解釈や誤りがあれば、それらはすべて著者の責に帰します。お気づきの点があれば、ご教示いただければ幸いです。
Yuuki Yamagishi· Zenodo (CERN European Organi...· 0 citations
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.
We use cookies to run the site and, with your consent, for analytics and to show ads.
See our Cookie Policy.
