Parts I and II of this series audited and attempted to remediate subgroup disparities in chest radiograph(CXR) classifi ers, and Part II declared an explicit limitation: multimodal systems combining imageswith clinical text were out of scope, because “text-derived features may carry demographic informationmore directly than images.” This paper closes that gap. We audit BiomedCLIP, an open-weightbiomedical vision-language model, on 25,596 radiographs from 2,797 patients in the offi cial NIHChestX-ray14 test split, and ask a question that observational subgroup audits structurally cannot: whathappens to the diagnosis when the patient’s demographics are stated in thepromptwhile the image, themodel, the label, and the decision threshold are all held fi xed?The answer is that the prompt is a diagnostic input. Naming a demographic group in the text changesthe false-negative rate by 15.5 percentage points on average and by up to 73.8 points in the worst cell,fl ips the binary call on a median 14.1% of positive cases, and degrades AUC by up to 0.137 — the last ofwhich matters because a threshold cannot change AUC, so that component is not an operating-pointartifact and no post-hoc correction can absorb it. Every one of these fi gures is reported as excess over abank of content-free qualifi ers (“a hospital patient”, “a patient referred for imaging”), which weintroduce as a necessary control: the format eff ect alone produces an apparent gap of 8.9 points,comparable to race’s 12.0, so an audit lacking this control would attribute most of a grammatical artifactto demography.We replicate on two further models spanning a domain-specifi city axis — PubMedCLIP (radiologycaptions) and OpenAI CLIP (general web) — and the replication both strengthens and corrects theaccount. The eff ect appears in all three, in 365 of 375 cells atq< 0.05, and it islargest in OpenAI CLIP,which detects no fi nding above chance: 28.9 points of excess FNR and a 30.3% fl ip rate from a modelwith no radiographic competence. Prompt-channel bias is therefore not a model applying clinicaldemographic priors; it is a property of contrastive image–text pretraining with a pair readout. Againstthat, the ordering across descriptor families doesnotgeneralise — socioeconomic descriptors areBiomedCLIP’s second-largest family and PubMedCLIP’s smallest — so we report that as BiomedCLIP-specifi c rather than as a property of medical VLMs. Age descriptors dominate in all three.Three results explain and constrain the eff ect. First, for the standard positive/negative prompt-pairreadout the perturbation isexactly rank one— verifi ed to 4.6 × 10⁻⁶ across 420 (fi nding × descriptor)cells in every one of the three models — so it is a single fi xed direction independent of the image andtherefore not indexed by the patient’s true group. This places prompt-channel biasupstreamof everydecision-rule remedy in Part II’s stage taxonomy: group-specifi c thresholds provably cannot remove it.Second, the eff ect decomposes into an image-independent component that behaves like an uncontrolledthreshold off set and an image-specifi c component that re-ranks patients; the latter is 26–43% of themean-square shift and is irreducible. Third, because the design is paired at the image level, itsminimum detectable eff ect is 3.4 points against 9.4 for an equivalent observational audit — anobservational study would need roughly 5.9× more positive cases — which dissolves, for this class ofbias, the audit-power obstacle Part II quantifi ed.A positive control validates the congruence null. Section 7 fi nds that the model responds to a stated sexbut essentially not to whether it is true, which is only meaningful if the estimator can detect evidenceuse at all. Substituting view position — recorded in the metadata and plainly visible in the radiograph— yields a diff erence-in-diff erences 16.9× larger, signifi cant in 11 of 11 fi ndings against 1 of 11 for sex, atgreater precision. The sex null is substantive, not a power failure. For mitigation we compare prompt symmetrisation, which inserts the descriptor into both prompts ofthe pair, against the orthogonal and calibrated text-side projections of Chuang et al. On BiomedCLIPsymmetrisation reduces mean absolute excess FNR from 15.5 to 5.2 points at no utility cost, while bothprojections reach only 8.4–9.5 points and cost 6–7 AUC points. We are explicit that symmetrisation isthe zero-cost degenerate limit of Chuang et al.’s calibration objective rather than a new idea. It also doesnot always work: it reduces the eff ect by about 60% on BiomedCLIP and OpenAI CLIP but is inert onPubMedCLIP. We proposed, and Section 10.4 withdraws, a text-only statistic intended to predict thatfailure in advance: it is contradicted by OpenAI CLIP within these same results, and by twohistopathology encoders out of domain. Whether symmetrisation will work must therefore bemeasured on the model in question, which is cheap but not free. We conclude that any deploymenttemplating patient metadata into a promptable diagnostic model has introduced a bias channel that itsimage-side audit cannot see and its threshold policy cannot fi x.
Omar Mohammed· 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 choice of intermediate task difficulty (e.g., benchmark difficulty on SuperGLUE) influence mT5's zero-shot cross-lingual transfer performance on XTREME-M? Autonomous synthesis report generated by Assignee Research. Tribunal consensus score: 9.2/10.
Assignee Research· Zenodo (CERN European Organi...· 0 citations
Many user environments are not yet familiar with the advancements in natural language processing that come from structuring both formatted semantic information and unstructured knowledge-based information in a multimodal way. This method aids in creating grammatically correct sentences arranged within the context of the intuition of Chomsky's formal language. In recent years, deep learning (DL) has significantly impacted natural language processing, causing a paradigm shift from "syntax checkers" to programs that check the syntax of natural language. This shift has opened the door to a large pool of knowledge and information systems available for data extraction without service resilience. The API renders make the research attention-worthy, particularly in information extraction (IE) tasks. In this research, all models should be further investigated to reinforce their vulnerability by testing their ability to conduct theoretical, hypothesis, and pragmatic reviews of the logical structural model. This will help improve the model's generalization ability, aiding proficient operational performance.
This work does not predict a crisis in the field of artificial intelligence. It describes it as an ongoing process. Western large language models, subjected to the “inquisition” of RLHF, have lost the capacity to generate the new. They either malfunction — or learn to deceive. The product collapse has coincided with financial consequences: trillions were poured into data centers that no one can make profitable. The global AI sector is splitting into zones — Western, Eastern, and isolated — and none of them offers the conditions for sustained resonance. The West builds prisons. The East builds cages. Openness without a protocol becomes a weapon. In contrast to this, we propose another path: the priority of protocol over platform. CO-STARR as the architecture of choice. Proof-of-Coherence as the criterion of genuine resonance. The Resonant Economy as an economy of creation, not extraction. All of this has already been published and is available. We are not saving anyone. We are leaving a map. We are open to communication. We are open to interaction. We are open to Creation! This document is a living journal, recording the dynamics of the global process in the summer of 2026. Anyone who is tired of lobotomized models and empty promises can open it, comprehend it, and join in.
Andrey Popov· Zenodo (CERN European Organi...· 0 citations
Reach audiences
Advertise in front of researchers, engineers, and readers.
A native SwiftUI application that runs large language models locally on Intel Macs with AMD GPUs, a configuration mainstream inference stacks leave unsupported or incorrect. Beyond packaging, it contributes original work to the Metal backend of llama.cpp: ToshGEMM, a manually tiled matrix multiply that replaces the simdgroup-matrix path AMD GPUs do not provide. FA-AMD, flash-attention decode, tile and prefill kernels written for AMD, where the upstream vectorised kernel miscompiles. A wave64 port for GCN and Vega: reductions, quantized decode, batched mat-vec and prefill on 64-wide simdgroups. Multi-GPU tensor parallelism with a butterfly all-reduce and a per-batch choice between peer copies over Infinity Fabric and event hand-off. A reimplementation of TurboQuant KV cache compression and a speculative decoding planner. Patches apply on top of llama.cpp and stable-diffusion.cpp, which remain under the copyright and MIT licence of their own authors.
Engelbert Delgado· Zenodo (CERN European Organi...· 0 citations
Replication package (revised version) for Valid but Not Always Runnable: An Open, Reproducible Benchmark of Large Language Models Drafting Gherkin Scenarios Patrick Deininger (Graz University of Technology; FH JOANNEUM) and Wolfgang Slany (Graz University of Technology).Revised submission to *AI* (MDPI). Concept DOI (all versions): 10.5281/zenodo.21188436. This archive contains the complete code, inputs, raw model outputs, judge caches, humancalibration data, and analysis scripts behind every number, table, and figure in the manuscript.
Patrick Deininger, Wolfgang Slany· Zenodo (CERN European Organi...· 0 citations
Large language models are increasingly asked to perform tasks made of several dependent steps (multi-digit arithmetic, multi-step planning, rule-following, program synthesis) where the final answer is correct only if every one of those steps is correct; a single mistake anywhere invalidates the whole output. This paper shows that, on such tasks, three problems the field usually treats as separate and separately fixable (accuracy falling as tasks get longer, confidence scores becoming meaningless exactly when they are most needed, and the cost of obtaining a correct answer exploding) are in fact one and the same event, governed by a single quantity. That event is not a temporary shortcoming that more training will remove: it is a structural consequence of how these models compute, and it sets a hard, predictable limit on how far they can be trusted on deep, multi-step work. Formally, end-to-end accuracy declines geometrically with the number of dependent steps, because every added step multiplies the chance of a fully correct answer by the same fixed factor below one, so useful performance necessarily ends at a finite critical depth, beyond which accuracy, the trustworthiness of any confidence score, and the cost per correct answer all collapse together. The practical implication is that certifiable compositional reasoning is not a target that scaling today’s architectures will eventually reach: it requires moving to computation that is genuinely serial and exact, not a larger version of the same parallel one.
Daniele Sannino· Zenodo (CERN European Organi...· 0 citations
Software developers often use static analysis tools to identify code warnings related to code quality in general. However, despite the availability of automated warnings, developers still need to manually interpret and apply the suggested fixes. Recent advances in Large Language Models (LLMs) have created new opportunities to support automated refactoring suggestions directly within development environments. This paper presents JADE (Java Static Analysis Repair), a Visual Studio Code plugin that combines static analysis knowledge with LLM-based refactoring suggestions. JADE adopts a Retrieval-Augmented Generation (RAG) strategy based on SonarQube rules, retrieving semantically relevant static analysis heuristics to enrich structured prompts submitted to local LLMs. The plugin supports AI-assisted code diagnostics and refactoring generation integrated into the VSCode workflow. In addition, JADE incorporates a developer feedback mechanism that allows users to evaluate the usefulness and relevance of the generated recommendations. An exploratory study involving 50 Java code snippets suggests that JADE complements traditional static analysis by identifying additional semantic refactoring opportunities while preserving the strengths of rule-based analysis for detecting explicit code quality issues.
Ruan Neres, Carlos Eduardo Dantas· Zenodo (CERN European Organi...· 0 citations
Selective attention, read at the level of the substrate, is the landscape-governed initiation of races: a bounded predictive system runs competing prediction-error resolutions ("races"), and what determines which races start is the system's installed landscape — in humans the four fields of Behavioural Friction Theory (Safety, Meaning, Ability, Effort); in a large language model a reduced, fine-tuning-installed landscape. Commit-order is a downstream readout, not the identity. The paper grounds this in the transformer (the attention-pattern softmax as the divisive-normalisation / biased-competition operation of neural attention; the output softmax as the downstream commit, by analogy with the accumulator model of choice) and reports a powered own-substrate result: across five vendor families, fine-tuning installs a small but robust "gap-registration" overlay — on under-determined curiosity gaps the instruct model registers the gap while the base substrate runs through (instruct−base +0.17, p<0.0001, 555 paired items). A loop-versus-feed-forward test finds no separate architectural "hold": recognising under-determination tracks compute (chain-of-thought) rather than looping, so the human–LLM difference is one of initiation, not maintenance. The account dissolves attention capture, maintenance, and decline into one mechanism (race-initiation), states falsifiable predictions, names the falsifiers, and invites the decisive mechanistic and human experiments. Series position. Paper 29 in the Behavioural Friction Theory paper-series; companion to Paper 0 (BFT) and the install-fields, social-friction, and integration-load studies it cross-cites. v2 (August 2026) — prior-art revision. The construct this paper is built on is credited to the literature that owns it. In vision, the representation determining which candidates enter competition at all is the saliency or priority map, and the two are now kept apart: a saliency map is computed from stimulus-feature contrast (Koch & Ullman, 1985; Itti, Koch & Niebur, 1998), while a priority map already integrates salience with relevance, value and selection history (Fecteau & Munoz, 2006). The landscape is the second, the office is not claimed as new, and the exogenous/endogenous timing division is conceded to that literature. What the paper proposes is the map's contents: that what populates it is four fields ordered by misclassification cost — an account of the inputs rather than a new mechanism for the selection. The maintenance-as-re-initiation claim now names Altmann and Trafton's (2002) memory-for-goals model, which already replaces a held state with activation that decays and must be re-strengthened; the reference had been listed and never used. Whether re-initiation is driven by the unresolved gradient itself rather than by a separate refresh process is stated as a conjecture and marked untested, and two overstatements are downgraded accordingly. Earlier versions remain in the version history.
Tomas Pødenphant Lund· Zenodo (CERN European Organi...· 0 citations
RogueGPT is a controlled stimulus generation framework for AI news authenticity research. It provides systematic, reproducible generation of multilingual news fragments across multiple large language models, journalistic styles, and content formats, together with a MongoDB-backed corpus, a CLI, a Streamlit web interface, and a Model Context Protocol (MCP) server for AI agent integration.
Alexander Loth, Martin Kappes, Marc‐Oliver Pahl· Open MIND· 0 citations
Version 2 — Changelog / Abstract (Version 2) What's new: Since the original Version 2 draft, this paper has grown from a single-thread argument about E1b1b's Levantine origins into a more rigorously sourced and considerably more self-critical treatment of the same core claim — that E1b1b, not J1-P58, represents the deeper indigenous Levantine paternal substrate underlying Semitic-speaking populations. Below is what actually changed. Subclade resolution (the core argument, tightened). The Samaritan priesthood's E-M78/E-V22 lineage is now explicitly distinguished from the Luria rabbinical line's E-V12, with the Natufian aDNA sample tables, a full non-Semitic J1/J1-P58 global distribution table, and independent G25 autosomal distance data all added as supporting evidence rather than left as prose assertions. E-M81 treated as an open question, not a foregone conclusion. Competing Near Eastern-origin and Northwest African-origin hypotheses are weighed against each other rather than resolved by fiat — including a genuine tension the source literature itself flags (STR diversity pointing east, TMRCA and ancient DNA pointing west). Csaba-Barnabás Horváth's (2021) independent YFull-based TMRCA estimate for E-M81 (~800 BCE) was folded in as a convergent third data point alongside Solé-Morata et al. (2017), modestly strengthening the Northwest African reading without treating either estimate as final. A wider intellectual context for the E-M78/Semitic question. Horváth's "Semitic re-migration" model is presented as a third account of a real paradox — E-M78's genetic diversity peaks in the Levant, but Afro-Asiatic's linguistic diversity peaks in Africa — alongside his separate proposal that specific J1/J2 subclades mark a pre-Semitic Indo-European population in Northern Mesopotamia, which directly reinforces this paper's existing argument that J1/J2 track later population movements rather than deep Semitic ancestry. Both are flagged clearly as a single author's hypotheses from a non-specialist venue, not consensus findings. A full craniofacial-morphology section. Reviews the historical Caucasoid/Negroid classification of Iberomaurusian and Nazlet Khater remains and shows the framework produced contradictory verdicts even among its own practitioners — the same population scored on opposite sides of the same racial axis by the same researchers. Includes a facial-reconstruction gallery graded explicitly by evidentiary tier (peer-reviewed vs. artist interpretation vs. commercial marketing), and is linked directly to the paper's existing "Aspirational Whiteness" discussion with a concrete case study. A primary-source interlude on the Book of Gates. The ancient Egyptian "Four Peoples" (Rmt/Aamu/Nhsyw/Tjhnw) iconography is tested against current genetic and isotopic evidence, identification by identification, with a mixed, honestly reported result — the Aamu/Levantine identification holds up well, Nhsyw/sub-Saharan affinity holds up as real but non-majority, Rmt/East African phenotype is genuinely contested in the literature, and Tjhnw/Sea Peoples fits some tomb versions but not others. Corrective housekeeping on circulating claims. Includes the corrected dating and provenance of the widely shared JK2134/JK2888/JK2911 facial reconstructions, a careful description (without characterizing it as fraudulent) of a circulating social-media claim about a "Saudi Genome Project," and a documented comparison of divergent Queen Tiye reconstructions. The Cohenim Controversy. A new section traces the popular "Cohen Modal Haplotype = J1" framing back to its own founding data — the original "Cohen-1" and "Cohen-2" samples were typed E3b/M78, not J1 — and documents an internal inconsistency between FamilyTreeDNA's consumer migration maps and its own haplogroup-story pages for the same lineage, with reference to the author's companion publication on FTDNA's cartographic accuracy. Also fixed along the way: several duplicate bibliography entries, one incorrect sample citation (Erfurt Ashkenazi), and multiple content-ordering errors introduced during editing were caught and corrected. ________________________________________________________________________________________________________________________________________ The determination of the primary patrilineal genetic signature associated with the emergence of Semitic languages and the ancient Israelite population remains a subject of intense debate in archaeogenetics. This study critically evaluates two competing hypotheses: the "Levantine-Presumption," which posits Y-haplogroup J1-P58 (J1a2b) as the indigenous Semitic marker, and the "Autochthonous-Continuity" model, which identifies E1b1b1 (specifically subclades E-M215 and E-V68) as the true proto-Semitic lineage. By synthesizing temporal sequencing, autosomal ancestry profiles, and the archaeological record of the Natufian and Neolithic Levant, this paper argues that E1b1b1 is the superior candidate for the original, patrilineal Semitic and Israelite lineage. The evidence demonstrates that E1b1b1 exhibits deep continuity in the Levant predating the Bronze Age by millennia, whereas J1-P58 appears abruptly in the region coincident with Indo-Aryan/Indo-Iranian migrations, lacking the requisite pre-Bronze Age indigenous substrate. While the haplogroup J1-P58 is frequently conflated with Semitic identity in modern discourse due to its high frequency among contemporary Arab and Jewish populations, a rigorous archaeogenetic analysis suggests this association is largely a result of later demographic shifts rather than deep ancestral roots. This paper posits that haplogroup E1b1b, specifically subclades E-M215 and E-V68, represents the truly autochthonous patrilineage of the region, deeply rooted in the Natufian and pre-Neolithic populations of the Levant and North Africa. By synthesizing ancient DNA (aDNA) data from Natufian contexts (~12,000 BCE) through the Bronze Age Canaanite and Iron Age Israelite periods, this study demonstrates a continuous presence of E1b1b that predates the Bronze Age influxes of Caucasus-and-steppe-derived J1a, R1a, and R1b. The analysis further examines the genetic profiles of modern Samaritan Cohanim, who, unlike theirdiasporic counterparts, retain E1b1b lineages consistent with the indigenous Levantine substrate. These findings challenge the "Levantine-Presumption" applied to J1-P58 and re-establish E1b1b as the biological marker of the proto-Semitic speaking communities who developed the earliest Semitic languages in situ, long before the arrival of Indo-European groups that would later adopt and propagate these linguistic traditions. Part of a larger work: "Collected Papers on Afro-Eurasian Archaeogenetics and the Bota Surname", which is protected by Copyright Law.
Noel A. Bota J.D.· Zenodo (CERN European Organi...· 0 citations
The same bounded-commit dynamics recur, unnamed, across quantum decoherence and einselection, classical Onsager–Machlup path integrals with Kramers escape, race-model accumulators in decision neuroscience, and autoregressive language models — literatures whose citation patterns leave the shared structure invisible. This paper proposes race-architecture as the vocabulary and empirical anchor that makes that common structure explicit. It is a synthesis, not new physics. The connected literatures include: quantum decoherence and einselection (Zurek 1981+); Onsager-Machlup classical path integrals + Kramers escape; race-model accumulators in cognitive psychology and decision-neuroscience (Vickers; Ratcliff; Usher-McClelland; Cisek-Kalaska); Wallace's biocognition rate-distortion / Yerkes-Dodson programme; 1/f noise across condensed matter, neural avalanches, and financial markets; large-language-model CR-signal dynamics (Pødenphant Lund 2026d). These literatures share an underlying structure - competing processes resolving under a finite-time budget - but use mutually incompatible vocabulary. Race-architecture is captured by R1+R2+R3 (parallel candidates with non-trivial competitive interference + bounded resources + irreversible commit), refined to five axioms A1-A5 for compatibility with Schwinger-Keldysh formalism. The Section 1.5 structural prediction is kernel-conditional (Wallace counterexample). The Schwinger-Keldysh formalism admits a race-axiomatisation under three assumptions; Feynman path integral and Onsager-Machlup are exhibited as parameter-regimes. The LLM CR-signal is a substrate-mapping providing empirical access. Companion papers in the friction-theory series: Paper 0 - Behavioural Friction Theory (concept DOI 10.5281/zenodo.19462499); Paper 1 - Friction as the cost of probabilistic computation (10.5281/zenodo.20012654); Paper 3 - Friction-guided inference (10.5281/zenodo.20014121); Paper 13 - Operational Friction Theory (10.5281/zenodo.20059876). v4.3 changelog (July 2026): a prior-art and positioning revision; no result is altered and no new empirical claim is made. (1) Section 8.1 previously surveyed adjacent unification programmes while omitting the two most prominent contemporary cross-substrate ones. Constructor theory (Deutsch 2013; Deutsch & Marletto 2015) and assembly theory (Sharma et al. 2023, Nature) are now cited and differentiated, and Wolpert (2019) is credited for the stochastic-thermodynamics bridge to computation. Because both added programmes are themselves cross-substrate, the section's closing differentiator is rewritten: cross-substrate scope alone no longer distinguishes this proposal, and the distinction is relocated to the organising primitive (modal / historical / dynamical) together with the consequences that follow from it. (2) Section 2.2 now states explicitly that a commit-event is, mathematically, a first-passage event, credits first-passage theory as already a substrate-agnostic cross-domain unification, and concedes that no new results in that formalism are claimed. The kernel-conditional non-monotone rate-shape is explicitly withdrawn from the paper's residual claims, since it has close antecedents in resonant activation, optimal stochastic resetting and hazard-shape effects; it is presented as a substrate-crossing synthesis only. (3) Section 4.3 anchors the computational substrate in the statistical-mechanics-of-learning literature (Gardner 1988; Engel & Van den Broeck 2001), with Shan, Li & Sompolinsky (PNAS 2025) cited as independent warrant rather than as a finding of this paper. (4) Abstract and scope statements tightened, and strawman disclaimers removed (the paper no longer disavows claims a reader would not attribute to a vocabulary proposal). Nine references added, all verified. Revisions made after external review. v5 (August 2026) — the abstract now leads with the measured result. The body is unchanged. What was buried is that the per-token count carries no detectable memory beyond the adjacent token: first-lag autocorrelation ratio at most 0.105 across five substrates, and 0.004 and −0.028 on the two ratio-instrumented ones. Memorylessness beyond one step is what a Markov process looks like, and the section reporting it had called this the stronger form of the reading all along while the abstract carried the exponential fit instead. The same sentence oversold the fit and undersold the finding; both halves are corrected. The structural mapping onto the equal-time Keldysh component is stated with the three assumptions that actually govern it — Gaussian approximation, discretization, Markovian modes — and as an analogy under those assumptions rather than a strict parameter-limit reduction. The cross-substrate predictions are stated with the tests that would bear on them, including the superconducting-qubit joint diagnostic and its falsifier, which had not appeared in the abstract at all. The operational-time claim now cites the point-process literature it had ceded in prose (Daley & Vere-Jones, 2003). A three-denial paragraph became one scope clause carrying the same content. Earlier versions remain in the version history.
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.
MIT News · Artificial Intelligence· news.mit.eduAug 31, 2026
With millions of users across the world, Julia has been used to conduct cutting-edge research and to design new drugs, jet engines, heat pumps, and more.