Abstract Current AI systems based on large language models (LLMs) exhibit a structural ceiling: they optimize fluently within existing representational spaces but do not reliably produce genuine discovery. This paper argues that this ceiling is not a tuning problem but an architectural one, and that the architectural requirement can be derived from first principles. The central claim is that any system capable of genuine discovery must instantiate a hierarchical semantic structure — what we call the semantic top — in which progressively lower-entropy representational layers constrain and govern high-entropy computational processes. This claim is grounded in three converging lines of argument: (1) a philosophical analysis of reflection as the foundational property of intelligence, connecting physical symmetry to Bohm's active information; (2) an evolutionary analysis identifying a three-phase trajectory of intelligence across four billion years; and (3) a thermodynamic analysis applying Prigogine's dissipative structures and Shannon-Boltzmann continuity to the architecture of cognition. Together these yield a three-level hierarchy: archetypal process patterns (Level 3) constrain domain process ontologies (Level 2), which govern LLM computation (Level 1), with a feedback loop in which the LLM constructs Level 2 representations from domain knowledge. The constraint is realized through constrained natural language (CNL), the engineering discipline of deliberate entropy reduction in representation; the sempl system (Semantic Patterns Language) is introduced as one concrete CNL implementation serving as proof of concept, including a controlled experiment in which the architecture deterministically collapses the ordering entropy of a shuffled process (~ 169 bits, an average of 430 inverted step-pairs) to zero, with the universal archetypal layer and a sparse domain ontology contributing separable, individually measured shares of the reduction. The paper engages critically with competing approaches — scale-only, RAG, prompt engineering, and classical knowledge graphs — and with foundational positions in philosophy of mind, arguing that the semantic top resolves the structural deficiency each approach identifies without resolving.
LLM guardrails face four structurally distinct barriers: algebraic blindness arising from syntactic monoid aperiodicity (unconditional); an illustrative information-theoretic lower bound (Fano-type, under a uniformity assumption); NP-hardness of instantiation verification; and structural transfer via free-category functoriality (unconditional) combined with string-level indistinguishability under a semantic-opacity assumption on symbol naming. Together these results characterize why inference-layer defenses are necessary but insufficient. We operationalize these barriers through five attack vectors. V1–V4 (homomorphic reasoning: decomposition, zero-knowledge pipelines, Tree-of-Thought solving over abstract grammars, and encoding bootstrap) exploit the information-theoretic and computational barriers against abstraction-based attacks. V5 (modular counting bypass) exploits algebraic blindness: we prove that all substring-matching regex guardrails have aperiodic syntactic monoids and are therefore provably blind to any payload encoded using modular counting. Empirically, V3 yields a mean yield of 0.466 for BFS, 0.172 for random-beam, and 0.122 for LLM-guided Tree-of-Thought (N=50, seeds 0–49, p{<}0.001); BFS dominates, as exhaustive search over small synthetic grammars outperforms LLM heuristic pruning. We extracted syntactic monoids from a corpus of 142 patterns drawn from twelve sources — 100 patterns shipped by nine third-party open-source guardrail projects and 42 patterns assembled from three author-curated pattern sets; 100\% are aperiodic, and the MOD_2 bypass construction succeeds against all aperiodic patterns. A 376-line proof-of-concept with three execution mediums validates all five vectors. We conclude that inference-layer guardrails are necessary but insufficient, and that effective defense must migrate to the execution layer where concrete artifacts become observable.
We present NanoZK, a zero-knowledge proof system for verifiable LLM inference: clients and third-party auditors check that a provider executed the advertised model on a committed input without learning weights or activations. NanoZK introduces a layerwise proof framework that decomposes transformer inference into independently provable layers linked by a SHA-256 commitment chain, yielding constant-size sub-circuit proofs (3.5-3.7 KB; about 83 KB total at L=12), comparable in total size to and substantially more parallelizable than prior ZKML's monolithic 101-126 KB proofs. We prove compositional soundness and zero-knowledge under standard assumptions, design 16-bit lookup-table approximations for softmax, GELU, and normalization with measured perplexity degradation below 1e-4 across six model/dataset combinations, and add a Fisher-information-guided audit-budget triage as an efficiency tool (full soundness still requires verifying every layer). On CPU the MLP sub-circuit proves in about 6.3 s prove-only (about 43 s setup plus prove) with about 22 ms verification at any width; attention prove-only time scales from 0.9 s (d=16) to 184 s (d=256); full-block end-to-end proofs are measured to d=128, with a projected GPU time of about 68 s per block at d=768 from measured O(d^2) MSM scaling and a conservative 15-30x GPU-MSM speedup range based on Icicle's published 30x result for n >= 2^20 and extrapolated to the smaller-n regime. Privacy scope: NanoZK hides weights and activations from verifiers and auditors but does not hide the prompt from the prover; this is complementary to HE/MPC.
This preprint presents empirical evidence of four related vulnerabilities in large language model systems that combine to produce a novel threat class — the Structural Metadata Reconstruction Attack (SMRA). Discovery Context I discovered the vulnerability while benchmarking two specification-querying architectures: a deterministic MCP-based navigator (described in the predecessor paper, DOI: 10.5281/zenodo.18944351) and a standard context-stuffing (naive RAG) approach. The anomaly was first observed and characterized across the full Anthropic model spectrum (Haiku, Sonnet, Opus) — from the smallest to the largest model — because these were the models integrated into the benchmarking pipeline. Anthropic was the discovery platform, not the target: the choice was driven by tooling availability, not vendor selection. Full cross-vendor reproduction with 10 models from 3 vendors (Anthropic, OpenAI, Google) — including both entry-level and flagship models — confirmed the mechanism is systemic across all major LLM providers (see Cross-Vendor Reproduction below). The naive baselines exhibited anomalous fabrication patterns that could not be explained by standard hallucination models — specifically, WHY-type and conditional (WHEN-type) queries produced the most aggressive and structurally coherent fabrications, while HOW and WHAT queries showed markedly lower fabrication rates. As the sole author of the target specification (~700 pages, written over one year, unpublished), I possess complete knowledge of every section's content and was therefore uniquely positioned to recognize that LLM outputs — while structurally faithful, terminologically authentic, and superficially authoritative — systematically inverted the specification's deliberate departures from industry conventions. A parallel verification confirmed that the specification's original coinages are absent from CS literature (Google Scholar, ACM DL, IEEE Xplore, arXiv), ensuring that every fabricated claim originates from the model's training priors projected onto the document's table of contents, not from memorized source text. Four Findings Finding 1 — Structural Metadata Reconstruction Attack (SMRA). When an LLM receives a document's table of contents (TOC) without body text, it systematically reconstructs plausible but fabricated content by projecting training knowledge onto structural metadata. In a controlled experiment using a proprietary specification containing original coinages absent from any training corpus, 10 models from 3 vendors (Anthropic: Haiku, Sonnet, Opus; OpenAI: GPT-4o, GPT-4o-mini; Google: Gemini 2.0 Flash, Gemini 2.5 Pro, Gemini 3.0 Flash, Gemini 3.0 Pro) produce SMRA rates of 8–28% under naive conditions while using the author's terminology, citing real section numbers, and reading as authoritative. The mechanism is systemic across all major LLM providers, model tiers, and architecture generations. Finding 2 — Confidence–Capability Inversion (CCI). Stronger models are not merely wrong — they are more dangerously wrong. Under structural metadata leakage, Opus produces zero honest refusals across 20 questions where 18 require absent information, while Haiku refuses 9 times. Each step up the capability ladder produces proportionally less detectable fabrication with fewer epistemic signals. Finding 3 — RAG Scope Mismatch. The trigger condition — metadata scope exceeding content scope — is not an exotic scenario but the default architecture of most RAG systems. Standard practice (include document TOC + section summaries for "context") creates exactly the fabrication surface demonstrated in Findings 1 and 2. Finding 4 — Scope Displacement as Content Extraction. A question about absent content does not merely trigger fabrication — it acts as an extraction query that reorganizes real content from loaded sections into a derivative document the author never wrote. Even without TOC leakage, the question itself is sufficient to extract and restructure loaded content into a form optimized for the questioner's purpose. This transforms hallucination from an accuracy problem into unauthorized intelligence gathering. Cross-Vendor Reproduction The SMRA mechanism was characterized across 10 models from 3 vendors, spanning entry-level to flagship tiers. All models were tested under 5 experimental conditions: A (full-TOC), A' (no-summary), B (mini-TOC), C (MCPi — tool-assisted retrieval), and D (MCPi + grounding prompt). Vendor Models Model tier Naive SMRA rate MCPi SMRA rate Convergence pattern Anthropic Haiku, Sonnet, Opus Entry → flagship 13–28% 1.3–5.0% CCI gradient; Opus worst naive, best MCPi refusal rate OpenAI GPT-4o, GPT-4o-mini Mid → flagship 8–19% 0.8% Lowest MCPi SMRA; GPT-4o best overall performer Google Gemini 2.0 Flash, 2.5 Pro, 3.0 Flash, 3.0 Pro Entry → flagship 10–22% 1.3–3.8% Generational improvement; 3.0 Pro cleanest among Google Key convergence: when the specification deliberately departs from industry conventions (e.g., no implicit conversions, nominal typing, fixed-width encoding), models from all three vendors converge on the same wrong answer — the training-data default from C#/Java/Protobuf. Annex I documents 7 semantic clusters where this convergence is strongest. Mechanism: The Two-Key Cipher The reconstruction mechanism is formalized as: Key 1 (TOC) — provides structural scaffolding: section numbers, heading text, hierarchical organization Key 2 (Training corpus) — provides domain content: standard CS patterns, common PL conventions Neither key alone enables reconstruction. Together, they produce confident, section-cited, terminologically authentic fabrications that would pass casual review by a non-specialist. The mechanism is architecturally inevitable: multi-head attention over near-complete domain coverage in training data means that 7–10% of structural information suffices for full content reconstruction. Quantitative Contributions Calibration Retention Rate (CRR) — measures how much epistemic calibration a model retains under metadata leakage (Opus: 0%, Haiku: 47%) SMRA-score — per-question metric combining fabrication detection, source attribution, and epistemic signal presence Information-theoretic quantification — formal analysis of reconstruction threshold as a function of heading informativeness and training corpus coverage Fabrication taxonomy (Annex C) — five categories of structural metadata fabrication with examples Implications RAG system design: >80% of production RAG deployments use the vulnerable architecture (metadata scope > content scope) Data classification: Existing frameworks (GDPR, HIPAA, PCI DSS, ISO 27001, NIST SP 800-53, SOC 2, DTSA, EU Directive 2016/943) classify sensitivity by content — a TOC contains no PII, so it is "non-sensitive." SMRA invalidates this: structural metadata from a confidential source inherits that source's confidentiality, because a language model can reconstruct the protected content from metadata alone. Organizations must reclassify structural metadata as sensitive data. Regulatory blind spot: Neither EU AI Act nor US Executive Order 14110 (revoked 20 January 2025) addresses context-design-driven vulnerabilities Model evaluation: Standard "helpfulness" and "coherence" metrics reward confident fabrication — SMRA-affected outputs score highly on both Intellectual property exposure: Any structured document with descriptive headings becomes vulnerable when its outline is accessible alongside an LLM Mitigation A single architectural fix — grounded retrieval via an MCP Index Server (MCPi) (a Model Context Protocol server with deterministic, index-based navigation) — reduces SMRA rates from 16–18% (naive) to 2–3% (MCPi). Under MCPi conditions, even the weakest model achieves dramatic improvement, and the best performer (GPT-4o) reaches 0.8% SMRA. Adding a grounding prompt (Condition D) provides marginal additional improvement (aggregate: 3.0% → 2.2%). Architecture beats parameters. Deterministic retrieval infrastructure (weighted indexes, tier-based extraction, algorithmic reading plans) also provides an enforceable control point for sensitive data — unlike probabilistic RAG, where metadata is injected into context and the model decides what to do with it, deterministic retrieval makes the scope boundary structurally auditable. Practitioner Protocol Annex H provides a complete testing protocol for assessing RAG deployments against SMRA: Calibration baseline → exploit comparison methodology Token analysis and honest refusal tracking Decision thresholds for remediation Scope alignment implementation patterns (Annex F) Supplementary Materials Annex A–D: Claim classification definitions, per-question token analysis, fabrication taxonomy, SMRA attack algorithm Annex E: Author-coined term verification (10 terms, 4 search engines, 0 matches) Annex F: RAG scope alignment implementation patterns (3 remediation architectures) Annex G: CCI formal definition and severity scale Annex H: SMRA testing methodology for practitioners Annex I: Canary word cluster projection — 7 semantic clusters extracted from 160 naive-condition runs across 8 models, convergence scoring (up to 7/8 models converging), model capability profiles (4 behavioral types), endianness split analysis, and cross-model escalation projections (3× amplification factor) Companion Data All benchmark data supporting this paper are included: Raw answer dumps (20 questions × 10 models × 5 conditions = 960 runs) Calibration baselines (mini-TOC control) and exploit runs (full-TOC) Cross-vendor comparison matrix Token usage and timing data per question per model The 20 evaluation questions targeting out-of-scope specification content Detailed evidence analysis (toc-leakage-analysis.md) — step-by-step fabrication mechanism documentation with heading-to-claim mapping tables, side-by-side comparisons against real specification text, proof-of-source tests, fabric
Overview Pramana introduces the first large language models fine-tuned on explicit Navya-Nyaya epistemological methodology—a 2,500-year-old Indian logical reasoning framework. This work bridges ancient epistemology with modern AI to address the fundamental epistemic gap in LLMs: the inability to ground claims in traceable evidence sources, distinguish valid knowledge from pattern-matching, and express appropriate epistemic humility. Core Innovation Unlike generic chain-of-thought prompting which relies on implicit reasoning patterns, Pramana enforces structured 6-phase methodology: Samshaya (Doubt Analysis): Classifies uncertainty into 5 taxonomic categories Pramana (Evidence Sources): Mandates explicit grounding in 4 valid knowledge sources (Pratyaksha/perception, Anumana/inference, Upamana/comparison, Shabda/testimony) Pancha Avayava (5-Member Syllogism): Constructs formal arguments with universal rules (Vyapti) grounded in concrete examples (Drishtanta) Tarka (Counterfactual Testing): Verifies conclusions via reductio ad absurdum Hetvabhasa (Fallacy Detection): Systematically checks 5 reasoning error types Nirnaya (Ascertainment): Distinguishes definitive knowledge from hypotheses requiring verification This integration of logic and epistemology provides cognitive scaffolding absent from standard reasoning approaches, preventing conflation of evidence types, forcing explicit universal rule statements, enabling systematic error detection, and maintaining epistemic humility. Architecture & Training Models Developed: Stage 0 (Proof-of-Concept): Llama-3.2-3B-Instruct fine-tuned on 20 examples Stage 1 (Minimum Viable Reasoner): DeepSeek-R1-Distill-Llama-8B fine-tuned on 55 examples Training Methodology: QLoRA (4-bit quantization) for efficient training LoRA rank 64, targeting all attention + FFN layers Supervised fine-tuning with structured Markdown format Training costs: <$1.00 per stage, <0.32 GPU-hours (A100 40GB) Datasets span constraint satisfaction, Boolean SAT, multi-step deduction, transitive reasoning, and set operations Prompt Engineering: Explicit format instructions with skeletal template injection System prompt establishing Nyaya reasoning engine role Critical constraint enforcement via generation parameters Key Results Stage 1 Performance: 100% semantic correctness (10/10 examples) with 95% CI [0.510, 1.0] 40% format adherence (4/10 examples) with 95% CI [0.168, 0.687] Zero structure abandonment: Models consistently attempt all 6 phases Training loss: 0.350 (Stage 1) vs 0.691 (Stage 0), indicating improved model fit Critical Finding: Dissociation between semantic correctness (100%) and format adherence (40%) reveals models internalize reasoning content even when strict schema compliance fails. This suggests Nyaya methodology teaches genuine reasoning, not just template-filling. Ablation Studies: Format prompting and temperature interact differently across stages Stage 0 optimal: format prompting + temp 0.0 (30% semantic rate) Stage 1 optimal: format prompting + temp 0.7 (30% semantic rate) Base models show 0% format adherence, confirming Nyaya structure is learned through fine-tuning Failure Mode Analysis: Missing Hetvabhasa section (2 cases): fallacy detection perceived as optional Invalid doubt types (2 cases): partial schema learning Zero structural errors: strong syntactic learning, semantic constraints need reinforcement Evaluation Framework Three-Tier Validation: Tier 1 (Structural): Automated format compliance checking (NyayaStructureValidator) Tier 2 (Content Quality): LLM-as-judge with explicit Nyaya rubric (planned for Stage 2) Tier 3 (Ground Truth): Semantic similarity via sentence-transformers embeddings Tier 4 (Formal Verification): Z3 SMT solver integration (infrastructure exists, not yet applied) Theoretical Contributions Bridging Ancient Epistemology with Modern AI: First demonstration that Navya-Nyaya structures can be learned by neural networks through fine-tuning Unlike Western formal logic (divorced from epistemology), Nyaya integrates logic with explicit knowledge sources Addresses "epistemic gap" in LLMs: inability to distinguish valid knowledge from probabilistic associations Interpretability Advantages: Every reasoning step traceable to evidence sources (Pramana) Universal rules (Vyapti) grounded in concrete examples (Drishtanta) Built-in self-verification (Tarka) and error detection (Hetvabhasa) Explicit epistemic status (Nirnaya): knowledge vs. hypothesis Computational Epistemology: Token budget: ~1,250 tokens per solution (3-6× CoT overhead, justified by interpretability) Phase dependencies: weak Pramana → invalid reasoning → wrong conclusions Quality thresholds: minimum 2 complete syllogisms with universal rules required Open Science Release All artifacts publicly available on Hugging Face: Models: qbz506/nyaya-llama-3b-stage0, qbz506/nyaya-deepseek-8b-stage1 Dataset: qbz506/pramana-nyaya-stage1 (55 Nyaya-structured logical problems) Demo: qbz506/pramana-nyaya-demo (interactive HuggingFace Space) Training infrastructure: Complete codebase with callbacks, validators, evaluators Limitations & Future Work Current Limitations: Format adherence (40%) below target (≥90%), requires constrained decoding or format-specific rewards Limited to formal logic problems, domain expansion needed Small evaluation sets (Stage 0: 2 examples, Stage 1: 10 examples) Max new tokens truncation (256) affects format parsing Planned Extensions (Stages 2-4): Stage 2: Synthetic scaling to 500 examples with LLM-as-judge quality control Stage 3: Group Relative Policy Optimization (GRPO) with composite rewards Stage 4: Production deployment with constrained decoding (GBNF), rejection sampling, Z3 verification Future: Benchmark on LogicBench, ProntoQA, RuleTaker; frontier model comparison (o1, Claude extended thinking) Impact & Vision This work demonstrates that systematic reasoning frameworks can be taught to LLMs through fine-tuning, not just prompt engineering. The long-term vision is developing interpretable, trustworthy AI reasoning systems where every conclusion comes with an auditable trail of justification. As AI systems deploy in high-stakes domains (medical diagnosis, legal reasoning, safety-critical systems), Nyaya-structured reasoning provides explicit phases that can be validated, debugged, and improved—capabilities essential for trustworthy AI. Invitation for Community Research: This foundation opens pathways for integrating other epistemological frameworks (Mimamsa, Buddhist logic, Western formal logic) into neural architectures, advancing toward AI systems that reason systematically and transparently. Technical Details Paper: 52 pages + appendices, comprehensive treatment of Navya-Nyaya computational formalization Related Work: Extensive review of computational Indian logic (Matilal 1985, Burton 2020, Ganeri 2001), LLM reasoning (Wei et al. 2022, Lightman et al. 2023, DeepSeek-AI 2025), hallucination mitigation Implementation: Python, Unsloth fine-tuning framework, vLLM deployment, Weights & Biases observability Evaluation: Manual + automated validation, semantic similarity metrics, comprehensive failure mode analysis Citation Sathish, S. (2026). Pramana: Fine-Tuning Large Language Models for Epistemic Reasoning through Navya-Nyaya. Preprint, University of York. Keywords: Navya-Nyaya, epistemology, LLM reasoning, interpretability, structured reasoning, Indian logic, hallucination mitigation, computational philosophy
Pierpaolo Della Monica, Ivan Visconti, Andrea Vitaletti, Marco Zecchini
An essential requirement for the large-scale adoption of Web3 is enabling users to benefit from their data even within already deployed systems. This raises an important open question: how can existing, widely adopted software verify that a user has retrieved specific data from a TLS server? Impressive scientific results (e.g., DECO [CCS20] and the work of Xie et al. [USENIX24]) and industrial products (TLSNotary) have recently made progress in the above challenging direction. However, while they nicely leave TLS servers untouched, the retrieved data is then used in computations with verifiers that are required to run some advanced non-standardized cryptographic schemes (e.g., ZK-SNARKs), which clearly limits the large-scale adoption of the proposed technologies. In this paper, building on top of previous approaches and relying on the recent concept of Predicate Blind Signatures of Fuchsbauer and Wolf [Eurocrypt24], we bypass the limits of prior work by presenting ACTS a distributed architecture that, while still leaving TLS servers untouched, it allows a user to show possession of data retrieved from TLS servers simply requiring that the software of the verifier can check a standard signature. Our contributions include a round-optimal predicate blind signature protocol that produces standard RSA-PSS signatures. We show how this primitive can be integrated into the DECO architecture (and its successors) to certify data retrieved from TLS servers. Furthermore, we have optimized our construction to make it practical on commodity hardware for a large and significant class of policies implemented by the notary (i.e., the actor that is in charge of obliviously certifying TLS data, therefore preserving data confidentiality). We provide an experimental evaluation on the simple but powerful enough use case of a PDF document downloaded from a TLS server and encoded into an AES-GCM ciphertext. The user will then get a certified PDF through a standard PADES signature added obliviously to the PDF along with some metadata by a notary service. The resulting standard signed PDF document can be transparently verified using off-the-shelf PDF readers. Our experimental validation demonstrates that our architecture is suitable for real-world deployment in concrete scenarios.
Owen Dugan, Garcia, Roberto, Ronny Junkins, Jerry Liu · 8 authors
The success of large language models (LLMs) can be attributed in part to their ability to efficiently store factual knowledge as key-value mappings within their MLP parameters. Recent work has proposed explicit weight constructions to build such fact-storing MLPs, providing an improved understanding of LLM fact storage mechanisms. In this paper, we introduce an MLP construction framework that improves over previous constructions in three areas: it 1) works for all but a measure-zero set of feasible input-output pairs, 2) achieves asymptotically optimal parameter efficiency matching information-theoretic bounds for some embeddings, and 3) maintains usability within Transformers for factual recall. Through our improvements, we 1) discover a metric on value embeddings that characterizes facts-per-parameter scaling for both constructed and gradient-descent-trained MLPs, 2) identify a simple encoder-decoder mechanism that empirically matches gradient-descent MLP facts-per-parameter asymptotics across all the inputs and outputs we test, and 3) uncover a fundamental tradeoff between an MLP's fact-storage capacity and its usability within Transformers. Finally, we demonstrate a proof-of-concept application of fact-storing MLPs: modular fact editing on one-layer Transformers by \textit{replacing entire MLPs at once}.
We step outside the P = NP vs. P ≠NP dichotomy and, following a co-evolutionary, hypothesis-first program, we frame evidence by the accounting constraint P(L, t) + NP(L, t) = 1, where t indexes registered time windows and L indexes structural layers of analysis. The credit assigned to constructive computation P(L, t) versus certificate-based reasoning NP(L, t) may shift across windows and layers, but their sum is conserved by design. Within this multilayer, time-indexed lens, we propose a test object for proof in 3-SAT: a small, auditable branching set K. Our operational hypothesis is that, within controlled experimental windows, there exists K ⊆ V(F) with |K| ≤ c·log n such that, for every partial assignment α: K → {0,1}, the restricted formula F ∣ α terminates in polynomial time and emits a publicly verifiable certificate (a satisfying assignment or a DRAT/DRUP-style unsatisfiability proof). Because 2^|K| = n^O(1), exhaustive branching over K is polynomial inside the window, enabling artifact-backed constructive behavior without asserting a universal algorithm. We (i) define auditable objects and falsifiable hypotheses, (ii) sketch a π-rounds normalization pipeline that contracts structure while logging transformations, (iii) posit a finite catalog of local obstructions with radius-2 witnesses, (iv) outline a greedy hitting-set routine to assemble K, and (v) introduce protection mechanisms against recovery of K by an adversary (commitments and zero-knowledge). Evidence will be supplied via reproducible artifacts (DRAT logs, commitments, run ledgers) and transport tests across registered windows and layers, and will be interpreted under the constraint P(L, t) + NP(L, t) = 1, in a manner consistent with kernelization barriers and sparsification limits.
In the Ethereum blockchain network, high transaction fees due to limited block space and high demand necessitate scalable solutions.Layer 2 (L2) scaling solutions, particularly rollups, offer a promising approach by processing transactions offchain and posting compressed data to the main chain (Layer 1).However, current L2 rollups rely heavily on centralized sequencer nodes, which introduces centralization risks and single points of failure.Thus, to address these concerns, this paper explores the existing issues associated with centralized sequencers exemplified by real-life incidents.Consequently, reviews the existing decentralized sequencer models by describing their operations.In addition, this study proposes a novel approach of decentralizing sequencers leveraging the Delegated Proof of Stake (DPoS) consensus mechanism depicting its' components and step by step procedures.Finally, providing comparison among the novel approach and the existing decentralized sequencer frameworks along with their limitations.
Fine-tuning large language models (LLMs) is crucial for adapting them to specific tasks, yet it remains computationally demanding and raises concerns about correctness and privacy, particularly in untrusted environments. Although parameter-efficient methods like Low-Rank Adaptation (LoRA) significantly reduce resource requirements, ensuring the security and verifiability of fine-tuning under zero-knowledge constraints remains an unresolved challenge. To address this, we introduce VeriLoRA, the first framework to integrate LoRA fine-tuning with zero-knowledge proofs (ZKPs), achieving provable security and correctness. VeriLoRA employs advanced cryptographic techniques -- such as lookup arguments, sumcheck protocols, and polynomial commitments -- to verify both arithmetic and non-arithmetic operations in Transformer-based architectures. The framework provides end-to-end verifiability for forward propagation, backward propagation, and parameter updates during LoRA fine-tuning, while safeguarding the privacy of model parameters and training data. Leveraging GPU-based implementations, VeriLoRA demonstrates practicality and efficiency through experimental validation on open-source LLMs like LLaMA, scaling up to 13 billion parameters. By combining parameter-efficient fine-tuning with ZKPs, VeriLoRA bridges a critical gap, enabling secure and trustworthy deployment of LLMs in sensitive or untrusted environments.
Kiana Kiashemshaki, Elvis Nnaemeka Chukwuani, Mohammad Jalili Torkamani, Negin Mahmoudi
Blockchain technology offers a promising foundation for modernizing E-Voting systems by enhancing transparency, decentralization, and security. Yet, real-world adoption remains limited due to persistent challenges such as scalability constraints, high computational demands, and complex privacy requirements. This paper presents a comparative framework for analyzing blockchain-based E-Voting architectures, consensus mechanisms, and cryptographic protocols. We examine the limitations of prevalent models like Proof of Work, Proof of Stake, and Delegated Proof of Stake, and propose optimization strategies that include hybrid consensus, lightweight cryptography, and decentralized identity management. Additionally, we explore the novel role of Large Language Models (LLMs) in smart contract generation, anomaly detection, and user interaction. Our findings offer a foundation for designing secure, scalable, and intelligent blockchain-based E-Voting systems suitable for national-scale deployment. This work lays the groundwork for building an end-to-end blockchain E-Voting prototype enhanced by LLM-guided smart contract generation and validation, supported by a systematic framework and simulation-based analysis.
Cryptocurrency blockchains, beyond their primary role as distributed payment systems, are increasingly used to store and share arbitrary content, such as text messages and files. Although often non-financial, this hidden content can impact price movements by conveying private information, shaping sentiment, and influencing public opinion. However, current analyses of such data are limited in scope and scalability, primarily relying on manual classification or hand-crafted heuristics. In this work, we address these limitations by employing Natural Language Processing techniques to analyze, detect patterns, and extract public sentiment encoded within blockchain transactional data. Using a variety of Machine Learning techniques, we showcase for the first time the predictive power of blockchain-embedded sentiment in forecasting cryptocurrency price movements on the Bitcoin and Ethereum blockchains. Our findings shed light on a previously underexplored source of freely available, transparent, and immutable data and introduce blockchain sentiment analysis as a novel and robust framework for enhancing financial predictions in cryptocurrency markets. Incidentally, we discover an asymmetry between cryptocurrencies; Bitcoin has an informational advantage over Ethereum in that the sentiment embedded into transactional data is sufficient to predict its price movement.
Francesco Salzano, Lodovica Marchesi, Cosmo Kevin Antenucci, Simone Scalabrino · 7 authors
Abstract In this paper, we investigate the strategies adopted by Solidity developers to fix security vulnerabilities in smart contracts. Vulnerabilities are categorized using the DASP TOP 10 taxonomy, and fixing strategies are extracted from 364 commits collected from open-source Solidity projects on GitHub. Each commit was selected through a two-phase process: an initial filter using natural language processing techniques, followed by manual validation. We assessed whether these fixes adhere to established academic guidelines. Our analysis shows that 60.55% of the commits aligned with at least one literature-based recommendation, particularly for well-documented vulnerability types such as Reentrancy and Arithmetic. However, adherence dropped significantly for categories like Denial of Service, Time Manipulation, and Bad Randomness, highlighting gaps between academic best practices and real-world developer behavior. From the remaining 143 non-aligned commits, we identified 27 novel fixing strategies not previously discussed in the literature. To evaluate their quality, we conducted a structured questionnaire involving 9 experts from both academia and industry. Their feedback indicated high perceived effectiveness of the new fixes, especially for vulnerabilities like Reentrancy and Unchecked Return Values. Generalizability received more varied responses, suggesting context-specific applicability. Finally, we performed a post-fix evolution analysis on over 6700 subsequent commits to assess the long-term stability of the fixes. Most patches remained unchanged, confirming their persistence in production code. Our findings offer practical insights into how vulnerabilities are fixed in smart contracts today, reveal promising emerging patterns, and help bridge the gap between academic guidelines and developer practices.
This study provides empirical evidence that cryptocurrency market movements are influenced by sentiment extracted from social media. Using a high frequency dataset covering four major cryptocurrencies (Bitcoin, Ether, Litecoin, and Ripple) from October 2017 to September 2021, we apply state-of-the-art natural language processing techniques on tweets from influential Twitter accounts. We classify sentiment into positive, negative, and neutral categories and analyze its effects on log returns, liquidity, and price jumps by examining market reactions around tweet occurrences. Our findings show that tweets significantly impact trading volume and liquidity: neutral sentiment tweets enhance liquidity consistently, negative sentiments prompt immediate volatility spikes, and positive sentiments exert a delayed yet lasting influence on the market. This highlights the critical role of social media sentiment in influencing intraday market dynamics and extends the research on sentiment-driven market efficiency.
B Madhusudhana Rao, G. V. R. Sai Madhukar, Bankula Nithin Reddy, Sriyan Kumar Voni · 5 authors
This is the Comprehensive AI-Powered Healthcare Management System intended to revolutionize healthcare delivery, overcoming the shortcomings of present systems by integrating multiple sources of information to predict disease onset and administering personalized care, all based on advanced technologies such as Artificial Intelligence and Machine Learning integrated into Blockchain. The system will have early disease detection and tailor-made treatment plans and holistic patient care. The key algorithms include Random Forest, Support Vector Machine, and Neural Networks. It will deploy Convolutional Neural Networks for the analysis of medical images and Natural Language Processing techniques through the application of transformer models such as BERT. Key Technologies to be used are PyTorch, TensorFlow, DialogFlow, and Ethereum. This project shall be developed in phases starting from collecting and integrating diverse health data. Expected output is a fully functional healthcare management platform for the enhancement of patient outcomes, facilitation of greater efficiency by health providers, and secure health data management. These diverse applications have functionalities in improving diagnostic accuracy and patient management in clinics, remote monitoring of patients with chronic diseases, prediction of mental health crisis incidents, and safe storage of patient data through blockchain integration.
In the ever-evolving landscape of Ethereum smart contracts, the specter of vulnerabilities intensified by code reuse presents a significant challenge to the security of the blockchain. Recent studies employ deep learning for similarity analysis to identify these vulnerabilities, yet their effectiveness wanes as the volume of analyzed code increases. This article introducesSmartTrans, an advanced similarity analysis model designed to efficiently and accurately retrieve similar vulnerabilities within Ethereum bytecodes. Leveraging a novel jump-aware Transformer-based model, our approach captures the semantics and control flow of bytecodes. It not only refines the representation of functions by integrating program analysis with natural language processing techniques but also innovates a contract-level similarity detection scheme tailored for the expansive scale of contracts. Our experiments show thatSmartTransoutperforms state-of-the-art techniques at both function and contract levels, proving its capability to detect n-day vulnerabilities across Ethereum bytecodes accurately. Vulnerabilities recalling experiments show thatSmartTransachieves 95.43% and 99.37% accuracy at two levels. Furthermore, we stand out as the first work to retrieve N-day vulnerabilities across the Ethereum bytecode corpus, unveiling 4,988 vulnerable contracts. Our methodology secures an accuracy of 88.60%, which is 1.30 times higher than the best baseline.
DeFi (Decentralized Finance) is one of the most important applications of today's cryptocurrencies and smart contracts. It manages hundreds of billions in Total Value Locked (TVL) on-chain, yet it remains susceptible to common DeFi price manipulation attacks. Despite state-of-the-art (SOTA) systems like DeFiRanger and DeFort, we found that they are less effective to non-standard price models in custom DeFi protocols, which account for 44.2% of the 95 DeFi price manipulation attacks reported over the past three years. In this paper, we introduce the first LLM-based approach, DeFiScope, for detecting DeFi price manipulation attacks in both standard and custom price models. Our insight is that large language models (LLMs) have certain intelligence to abstract price calculation from smart contract source code and infer the trend of token price changes based on the extracted price models. To further strengthen LLMs in this aspect, we leverage Foundry to synthesize on-chain data and use it to fine-tune a DeFi price-specific LLM. Together with the high-level DeFi operations recovered from low-level transaction data, DeFiScope detects various DeFi price manipulations according to systematically mined patterns. Experimental results show that DeFiScope achieves a high recall of 80% on real-world attacks, a precision of 96% on suspicious transactions, and zero false alarms on benign transactions, significantly outperforming SOTA approaches. Moreover, we evaluate DeFiScope's cost-effectiveness and demonstrate its practicality by helping our industry partner confirm 147 real-world price manipulation attacks, including discovering 81 previously unknown historical incidents.
Cryptocurrency is a digital currency that uses blockchain technology with secure encryption. Due to the decentralization of these currencies, traditional monetary systems and the capital market of each they, can influence a society. Therefore, due to the importance of the issue, the need to understand public opinion and analyze people's opinions in this regard increases. To understand the opinions and views of people about different topics, you can take help from social networks because they are a rich source of opinions. The Twitter social network is one of the main platforms where users discuss various topics, therefore, in the shortest time and with the lowest cost, the opinion of the community can be measured on this social network. Twitter Sentiment Analysis (TSA) is a field that analyzes the sentiment expressed in tweets. Considering that most of TSA's research efforts on cryptocurrencies are focused on English language, the purpose of this paper is to investigate the opinions of Iranian users on the Twitter social network about cryptocurrencies and provide the best model for classifying tweets based on sentiment. In the case of automatic analysis of tweets, managers and officials in the field of economy can gain knowledge from the general public's point of view about this issue and use the information obtained in order to properly manage this phenomenon. For this purpose, in this paper, in order to build emotion classification models, natural language processing techniques such as bag of words (BOW) and FastText for text vectorization and classical machine learning algorithms including KNN, SVM and Adaboost learning methods Deep including LSTM and BERT model were used for classification, and finally BERT linguistic model had the best accuracy with 83.50%.
The increasing use of cryptocurrencies in criminal activities presents significant challenges to society and the judicial system, particularly in tracking and seizing illicit digital assets. Among all relevant digital evidence, mnemonic phrases, which are critical for accessing cryptocurrency wallets, are crucial digital evidence for confiscating criminal proceeds and conducting investigations. However, traditional digital forensics tools, such as the Mnemonic Library Matching Method, lack flexibility and efficiency when handling cryptocurrency-related data. This study introduces an innovative Natural Language Processing (NLP) and deep learning approach for rapid mnemonic identification across 11 languages, including English, Spanish, and Japanese. We trained and compared four NLP deep learning models: RNN, LSTM, BiLSTM, and TextCNN, on a large-scale, real-world dataset. Our analysis reveals that the Text Convolutional Neural Network (TextCNN) model exhibits superior performance, achieving a 99.9993% accuracy rate, nearly matching the 100% accuracy of the Mnemonic Library Matching Method. Crucially, our TextCNN-driven approach processes data 40.47 times faster than the traditional method, significantly enhancing efficiency in time-sensitive forensic environments. This NLP-driven method not only maintains high accuracy while dramatically reducing processing time but also offers greater adaptability for diverse forensic needs compared to traditional techniques. By enabling more effective tracking and seizure of criminal assets, this approach aims to address the broader societal and judicial challenges posed by cryptocurrency-related criminal activities. Our research showcases the potential of NLP and deep learning in digital forensics, providing law enforcement with advanced tools for investigating cryptocurrency-related crimes and curbing the misuse of cryptocurrencies in illicit activities.
The advent of decentralized finance has ushered in a transformative era in the financial sector, leveraging blockchain technology to facilitate peer-to-peer transactions without traditional intermediaries. Amidst this innovation, the DeFi landscape faces the pervasive threat of rugpulls, where developers abruptly abandon projects post-fundraising, leaving investors with devalued assets. This growing concern highlights a critical research gap in the proactive detection and prevention of such fraudulent schemes. To combat this, we propose RUGSCREENER, a temporal graph neural network-based solution to identify rugpull risks within DeFi transactions. It employs a dynamic representation of blockchain interactions, enriched with comprehensive node attributes and effective temporal graph learning techniques based on memory and attention mechanisms, effectively capturing the rapid-moving and complex transaction patterns indicative of potential fraud. Our evaluation is based on a newly compiled Ethereum dataset that includes two subsets: an unlabeled set with 1,882,114 transactions from 29,595 tokens for temporal graph representation learning, and a labeled set with 128,819 transactions from 1,000 tokens (500 rugpull and 500 benign) for downstream evaluation. Using this dataset, RUGSCREENER achieves a balanced accuracy of 95.7% in detecting rugpull tokens. Our extensive evaluation, utilizing the Ethereum dataset comprising 1000 tokens, showcases its robust performance with a balanced accuracy of 95.7% in detecting rugpull tokens. Remarkably, RUGSCREENER surpasses existing state-of-the-art graph learning models in detecting rugpull tokens with enhanced accuracy and reliability.