2731 articles
OpenAI makes GPT-5.5, GPT-5.4, and Codex available through Amazon Bedrock at identical pricing to OpenAI's platform. Models run in commercial and government AWS regions, currently limited to the US. Usage counts toward existing AWS contracts.
Choose France 2026 displays €93 billion in announced investments, of which €45 billion are firm contractual commitments. SoftBank accounts for half the volume.
California's university system heavily invested in AI, creating internal tensions. The article explores organizational conflicts and challenges of integrating AI into academic institutions.
ExpGraph, an arXiv preprint from May 29, 2026, proposes a self-evolving graph memory architecture for LLM agents. The system accumulates experience through dynamic nodes and edges, enabling agents to improve performance on complex tasks without retraining.
A user built a large-scale scraping pipeline aggregating 2M+ active job postings from 100,000+ company career sites. Dataset in Parquet format, daily-refreshed, freely accessible with standard fields (title, company, description, location, URL).
OpenAI calls for global action on youth AI safety, proposing a dedicated AI Safety Institute. The company advocates for global leadership and coordinated policies to address youth-related AI risks.
Nvidia unveils Cosmos 3, an AI model presented at Taipei GTC alongside humanoid robot Isaac GROOT. The model aims to enhance understanding of real-world scenes.
llama.cpp benchmarks on Intel Arc Pro B70 posted. Performance measured: 63 tokens/sec with Qwen via SYCL. Test of Intel discrete GPU for local inference.
Vercel now enables configuring Git settings for all projects in a monorepo from a single location, eliminating the need to configure each project individually. Settings include commit status and repository_dispatch events.
NVIDIA releases Cosmos 3, a collection of omnimodal world models (Nano 16B, Super 64B) capable of generating dynamic video, image, audio, and action commands from text, image, video, and action trajectory inputs. Available on Hugging Face for Physical AI applications.
Nvidia unveils Isaac GR00T, a humanoid robot presented by Jensen Huang. The reference model combines manipulation capabilities and learning features.
Simon Willison built a web tool that replicates Claude.ai's feature: detecting large pasted text volumes and automatically converting them to file attachments. The tool also supports direct file opening and images (shown as thumbnails) plus drag-and-drop functionality.
CAST is an answer-free self-distillation method for GRPO (Group Relative Policy Optimization). It uses a stop-gradient self-teacher to shape token-level advantages by trajectory correctness, with bidirectional advantage sign reversal and bounded advantages for zero-variance groups. Improves mathematical reasoning performance.
Theoretical study of optimal learning rates in 2-3 layer linear neural networks. Derives exact closed-form expressions for gradients and test loss after 1-2 gradient descent steps. Key finding: unequal layer-wise learning rates minimize loss in early training, while equal rates become optimal later. Code released.
Asymptotic theory for SGLD-Gibbs combining stochastic gradient Langevin dynamics with Gibbs updates for Bayesian inference in latent variable models. Authors derive jump-diffusion limits and propose hyperparameter tuning guidance ensuring statistically meaningful uncertainty quantification.
Iterative AI workflow optimizes graphite-based anodes through sequential learning and experimental feedback loops. Citrine Platform generates surrogate models and refines manufacturing constraints. Results: fabrication reliability improved from frequent failures to 100% success, cells ≥350 mAh/g increased from 28.4% to 84.8%, capacity retention rose from 42.1% to 97.3%.
Novel approach for Major Depressive Disorder detection from EEG without data augmentation. SGC (Score-Guided Classification) uses an unsupervised generative network to model pathological anomalies as prior, fused with deep feature representations. Cross-Channel Spatial Adaptation module handles multi-center channel heterogeneity. Validated on Mumtaz2016 and MODMA datasets.
RAFT is a two-stage domain fine-tuning method that mitigates catastrophic forgetting. It refines data via self-conditioned rewriting and answer fusion, then applies on-policy distillation where the original model provides soft targets on student-generated trajectories. Across five domains, RAFT improves domain accuracy by 23.2% over standard SFT and recovers 18.2% of degradation on MS-Bench.
BudgetDraft trains a sparse drafter for speculative decoding in long-context inference (4K-16K tokens). The method exposes the model to multiple KV budgets during training and aligns each sparse view with a shared full-cache teacher target. Results: 6.55x, 4.46x, 2.10x speedup vs autoregressive decoding at 4K, 8K, 16K tokens.
Researchers demonstrate that hidden reasoning traces in LLMs can be extracted via Reasoning Exposure Prompting (REP), a lightweight prompting method using shadow-model-generated demonstrations in auxiliary code-like formats. REP exposes internal traces even when deployed systems intentionally hide them, while preserving useful reasoning signals for distillation.
TRACE, a trajectory compression method, detects safety risks in long-horizon LLM agents. A Compressor encodes the full trajectory into a supervised latent evidence state, while a Reader judges safety using this reference. Achieves up to 12.6 percentage point improvements on ASSEBench, Pre-Ex-Bench, and R-Judge.
KACE decouples storage from usage in context for mathematical reasoning. An epistemic tree stratified by difficulty and domain is built offline via self-reflective loop. At evaluation, tiered self-consistency dynamically classifies problems and selectively retrieves matching cards. On AIME 2025: 62.2% accuracy (+10.4 points vs Best-of-5).
arXiv paper on LLM inference batching optimization. Authors demonstrate mixed batching (MB) is suboptimal on bandwidth-constrained GPUs: exclusive batching (EB) achieves 41.9% higher throughput on RTX PRO 6000 (1.792 TB/s). They propose EB+, a hybrid scheduler that dynamically switches between EB and MB based on GPU bandwidth, model size, and workload composition, reaching 36.4% gains under non-stationary traffic.
TAPS introduces a target-aware prefix selection method for diffusion-drafted speculative decoding. By converting diffusion marginals into path-conditioned acceptance estimates, TAPS selects a compact prefix-closed subtree under fixed verification budget. Results: 7.9x lossless speedup vs vanilla autoregressive decoding, 1.36x and 1.74x over DFlash and DDTree.
New set-distance reward method for reinforcement learning of vision-language models on chest X-ray report generation. Tested on Qwen3-VL, Gemma3 with GRPO: 6.80% (BERTScore), 7.82% (RadGraph F1), 4.45% (CheXbert F1) improvements over supervised fine-tuning. Enables test-time best-of-N selection and mid-generation pruning reducing tokens by 50%.
GEM is a concept erasure framework for Rectified Flow Transformers. It bridges trajectory-based unlearning (Generative Flow Networks) and teacher-guided erasure, using geometric guidance signals to suppress unwanted concepts while preserving benign generation and preventing harmful content synthesis.
VESTA is a statistical agent framework equipping VLMs with a dynamically growing exploration toolkit for quantitative model refinement. Evaluated on DAWN (benchmark for distribution fitting and time series modeling), VESTA outperforms prior agentic systems, especially on complex astronomy tasks (initial mass functions, gravitational-wave chirp signals).
Decoder-only models hit an information-theoretic limit in deterministic state-tracking tasks beyond ~25 steps. An Attention Bottleneck Theorem bounds capacity to O(H·log(L/H)·√dh). Across 12 models and 8 domains (SWE-Bench, WebArena, SQL), tool delegation achieves 86-94% vs 24-42% for pure neural reasoning. Fine-tuning improves <5%, confirming an architectural ceiling.
Study on preventing catastrophic forgetting during continual pretraining of multilingual language models. Authors propose five parameter alignment strategies (layer freezing, regularization, post-hoc reversion, model merging) tested across 32 languages and four evaluation axes. Parameter alignment substantially reduces forgetting while maintaining language acquisition.
Hybrid framework coupling LLMs with physics-based simulation for inorganic material synthesis planning. Case study on niobium-oxygen system: LLM-generated synthesis routes outperform classical path-planning algorithms by leveraging implicit priors.
Survey paper proposing Intelligent Computing Architecture Model (ICAM), a six-layer framework for model-native computing. Maps classical computer architecture concepts to LLM systems (cache management, context, agents). Introduces three design laws: Semantic Locality Law, Context Budget Law, Agent Speedup Law. Distinguishes probabilistic execution plane from deterministic control plane.
FETCH, an automated legal triage classifier, generates follow-up questions using a low-cost LLM ensemble. The study shows cheap models perform well at classification, but high-quality plain-language question generation requires GPT-4 or higher. Prompt engineering alone is insufficient; LLM-as-judge ratings diverge from human evaluations.
CTRL-STEER introduces a closed-loop control framework for Vision-Language-Action (VLA) models. Instead of fixed steering coefficients, it adaptively adjusts intervention strength over time using PID or reinforcement learning controllers. Experiments on OpenVLA with LIBERO task suites demonstrate improved concept regulation stability and better steering-task success trade-offs without retraining the base model.
Geodesic Flow Matching adapts Riemannian transport to denoise Spatial Semantic Pointers (SSPs) on toroidal manifolds. Tested on a Spiking Neural SLAM system, the method achieves 72% reduction in tracking error and 40% increase in neural efficiency versus baselines.
Automatically Differentiable Nonlinear Tensor Networks (ADNTNs) compress neural network weights using compact core tensors trained via reverse-mode automatic differentiation. Tested on AlexNet and VGG-16, they achieve per-layer compression ratios from 2000× to 77000× while maintaining or improving accuracy.
Demo2Reward optimizes VLM reward model language instructions at test-time using 3-10 expert demonstrations to reduce false positives in robotics. No additional training required. Validated on simulated tasks and real-world transfer.
Hoeffding Concept Bottleneck Models (HCBM) replace linear concept aggregation with non-linear sparse decomposition based on gradient-boosted trees. Robust to inter-concept leakage, they improve explainability of computer vision predictions, particularly on overhead imagery.
TIGER is an inference-time framework to mitigate hallucinations in multimodal generation. It independently extracts an observation graph from input and a claim graph from output, then assigns risk scores to claims based on support and conflict. The model repairs high-risk claims while keeping the backbone frozen. Convergence analysis shows geometric risk reduction to an explicit asymptotic bound.
AbaqusAgent is a multi-agent LLM-based framework for finite element analysis (FEA) in solid mechanics. Composed of six agents (interpreter, architect, input writer, runner, reviewer, visualizer), it converts natural-language instructions into executed FEA analyses with Abaqus. Validated on 50 problems with 86% success rate.
Interactive reasoning evaluation benchmark with 474 executable games. LLMs receive only task rules, must query a hidden environment, integrate partial observations, and decide when to submit answers. Evaluates contextual robustness, metacognitive adaptation, and interaction efficiency across frontier models.
arXiv paper introduces Universal Quantum Transformer (UQT), a quantum-native architecture leveraging multi-qubit properties for exact mathematical reasoning. On 5 qubits, UQT perfectly learns modular arithmetic (Z₁₁) and non-Abelian algebra (S₄) with deterministic generalization, no stochastic instability. Deployed on IBM Quantum hardware, it bypasses classical attention's quadratic bottleneck.
Novel Permutation-Invariant Bayesian Optimization (PIBO) approach leveraging Optimal Transport theory to optimize offshore wind farm layouts. PIBO handles permutation invariance (turbine order doesn't affect energy production), cuts computation time by ~50%, and improves solutions versus vanilla BO.
Position paper on post-solve robustness in MILP decision engines. Identifies gap: nominally optimal solutions become infeasible under small cost/resource perturbations. Proposes audit layer around incumbent solution, combining certified inner approximations, probabilistic robustness estimation, and solver-backed verification.
SPADER is an RL framework for tool-augmented LLM agents in Multi-Answer QA. It introduces Step-wise Peer Advantage (SPA) for fine-grained credit assignment over long trajectories, and a diversity-aware exploration reward promoting rare entity discovery. Evaluated on QAMPARI, Mintaka, WebQSP, QUEST: improves recall and F1 vs prompting and supervised RL baselines.
Theoretical and empirical study of parameter-based knowledge editing limits in LLMs. Authors prove via dimensional collapse hypothesis that localized modifications propagate global interference degrading model capabilities. Retrieval-based methods consistently outperform parameter-editing approaches.
SelSkill, a dual-granularity preference-learning framework, optimizes selective skill invocation in agentic tasks. On ALFWorld with Qwen3-8B: +10.9 pp task success, +29.1 pp execution precision. On BFCL: +5.7 pp task success, +29.5 pp execution precision. Zero-shot transfer to Tau-bench and PopQA.
LaSR proposes a training paradigm for Speech LLMs featuring latent reasoning aligned to acoustic feature regions of target words. Without explicit intermediate tokens, the method improves specialized vocabulary recognition on Fun-Audio-Chat. A new Spoken Darwin-Science corpus for academic terminology is introduced.
arXiv study on LLM adaptation limits for annotation tasks. Toxicity detection experiments across diverse datasets show 66% of zero-shot errors resist correction via prompting (rescue rate 34.8%). Models follow misaligned definitions while maintaining confidence. Definition-Specific Familiarity (DSF) metric correlates with performance (r=+0.41), outperforming memorization metrics.
Empirical study adapting Phi Silica for short-form text rewriting via dataset curation, prompt distillation, and parameter-efficient fine-tuning. Dataset built from public slide decks. Results: improved semantic fidelity, reduced hallucinations, higher preference win rate vs GPT-5-chat.
Academic paper proposing product-aware autoencoders for anomaly detection in multi-product cyber-physical systems. Traditional global models create blind spots where attacks can evade detection. Tests on Tennessee Eastman Process benchmark: product-aware model achieves 100% detection accuracy versus 22.2% for global baseline in attack scenarios.
ProtStructQA is an executable benchmark for protein structural question answering. 382.2K questions generated from hidden domain-specific language, evaluated on Qwen3 (0.6B–8B) and Gemma-3. Key finding: capability threshold between Qwen3-1.7B and 4B where models transition from failing to produce executable denotations to mastering chain-of-thought reasoning.
New automated metric (Triangulated Preference Shift score) to measure lexical biases introduced during preference learning (RLHF) in LLMs without manual curation. Analysis across 6 model families reveals a shift toward a 'language of prestige' (overuse of 'delve', 'furthermore').
BitsMoE introduces spectral-energy-guided bit allocation for MoE LLM quantization. Using SVD decomposition, it preserves shared basis unquantized and fine-grained quantizes expert-specific factors via integer linear programming. On Qwen3-30B at 2-bit, it improves accuracy by 27.83 percentage points and increases decoding speed 1.76× over GPTQ.
TOPD (Trajectory-aware On-Policy Distillation) improves LLM reasoning by using near-future trajectory information to identify truly divergent states. On AIME24/25, TOPD reaches 63.3%/53.3% vs 60.0%/46.7% in standard OPD, showing 30% of high-loss tokens are false positives.
Study of automatic assessment for massive multilingual bitext: decomposed into parallelism evaluation via multilingual embeddings and reference-free quality estimation. Benchmarks 4 embedding models and 9 evaluators on FLORES-200 covering 6,654 language-pair directions. Key finding: no single model is universally reliable; direction-aware routing and calibration required.
Governance protocol for multi-agent knowledge bases combining formalized artifact lifecycle, reputation-weighted deliberative voting (Beta Reputation + EigenTrust), and graduated sanctions. Evaluation on 100 agents, 7 behavioral archetypes: 0.826 vs 0.791 precision under moderate adversity (p<0.001). Commit-reveal vote concealment yields +8.2-8.6pp improvement.
DeSQ decomposes complex questions into atomic constraints to generate SPARQL queries. The framework operates in three stages: decomposition, generation of SPARQL fragments with placeholders, and final assembly. Outperforms state-of-the-art on 4/5 major benchmarks with improved robustness to lexical variation.
RealityTest evaluates whether AI systems disclose their identity when asked. Multimodal, multilingual benchmark based on 3,152 identity-probing queries from ~750 participants across 49 countries, 5 languages (text and speech). Findings: only 31% ask directly; a single suppression instruction reduces disclosure below 30% even in best-performing models.
Study of agreement metrics for LLM-as-Judge evaluation. Analysis of 24 recent papers shows that for binary criteria (MET/UNMET), Pearson r, Spearman ρ, Kendall τ_b, and phi are redundant. Cohen's κ alone adds information. Authors propose a reporting checklist including judgment scale, abstention handling, and confusion matrix.
DLLM-JEPA pairs JEPA with masked-diffusion language models for self-supervised representation learning. Eliminates need for explicit multi-view data and reduces training FLOPs by 33% vs LLM-JEPA. Achieves +18.7pp improvement on GSM8K (LLaDA-8B) and +11.4pp (Dream-7B) while preserving base model capabilities.
Multi-domain red teaming framework evaluating 11 LLMs across 690 clinical scenarios. Results: substantial variance (scores 0.791–0.984), safety-critical failures masked by aggregate accuracy, 10-20% error amplification on equity tasks. Hybrid evaluation (automated + human validation) essential.
ProactiveLLM learns when to interact with streaming inputs without external signals. Through monotonic random masking and synchronized privileged self-distillation, the model perceives semantic sufficiency from partial inputs. Reduces interaction latency while maintaining quality on text and speech streaming tasks.
Humor generation system for SemEval-2026 Task-1 (MWAHAHA) using a "generate-many → select-best" strategy. Diverse candidate generation via multi-step prompting and model ensembling, then selection using a preference model trained on 2.5K human pairwise judgments. Ranked 1st in English and Chinese subtasks, 2nd in Spanish.
SENSE improves retrieval-based speculative decoding by anchoring retrieval on target model hidden states for robust semantic alignment. A soft-gated evaluation module validates semantic equivalence rather than surface forms. On LLaMA and Qwen, SENSE achieves 4.09 mean acceptance length and 3.26x speedup.
CSRP, a three-stage framework for Chinese grammatical error correction, combines continual pre-training (5.9M samples), Chain-of-Thought fine-tuning, and policy optimization with efficiency-aware rewards. Achieves 50.99 F₀.₅ on NACGEC and outperforms GPT-4 on spelling correction (59.61 F1).
ART (Attention Run-time Termination) is a lightweight runtime mechanism that halts KV block accesses during decoding once their attention contribution becomes negligible. Tested on LongBench, it achieves 20% higher generation throughput on large batches while maintaining comparable accuracy.
DOPA, a demonstration retrieval framework, uses an OOD proxy to approximate the inaccessible target domain and guide selection of relevant demonstrations. A Mahalanobis distance-based global diversity constraint ensures sufficient variety among retrieved examples. Positive results across multiple LLMs and tasks under severe distribution shift.
DraDDP is the first public multimodal dataset for discourse parsing in multi-party dialogues. Built from American TV dramas, it contains 495 dialogue segments (6,374 utterances, 9.1 hours of video). Benchmarks demonstrate the value of multimodal information for identifying dependency structures and relation types between utterances.
InfoAtlas is a foundation model that estimates mutual information between high-dimensional variables in a single forward pass, eliminating costly iterative optimization. Pretrained on synthetic data with diverse dependence patterns, it achieves 100× speedup over neural MI estimators while matching accuracy and generalizes to complex real-world scenarios.
Theoretical study showing how transformers learn to implement tree search (DFS) via RL. A two-head transformer naturally emerges from policy gradient training on stochastic trees without expert demonstrations. The model generalizes to unseen depths and adapts its strategy based on goal distributions.
Novel spatio-temporal graph anomaly detection method modeling spectral energy shifts. Detects camouflaged anomalies (decreased spectral variation) missed by existing spectral approaches. Energy-aware framework compatible with message passing, no specialized sequence modules required.
New arXiv paper proposing discrete-time approximations to SG(L)D with and without momentum, enabling accurate predictions of stationary covariance and integrated autocorrelation time. Non-asymptotic error bounds for practical tuning and uncertainty quantification, validated on misspecified models and large batch sizes.
ChurnNet compares classical methods (Random Forests, XGBoost, SVM) with a Unified Multi-Task Time Series Model for customer churn prediction. Results show traditional approaches outperform the modern AI model in predictive performance, data efficiency, and computational requirements across multiple datasets.
ReMax formalizes exploration in RL through retries: a policy is evaluated by expected maximum return over M samples. Exploration emerges naturally without explicit bonuses. RePPO, a PPO variant optimizing ReMax, generalizes discrete M to continuous parameter m for fine-grained exploration control. Results on MinAtar and Craftax benchmarks.
Study on adaptive data selection for wearable prediction across 3 modalities (heart rate, activity, EMA) in longitudinal dataset. AUROC gains up to 0.7 for low-baseline participants, strong inverse correlation (r=-0.67) between improvement and baseline performance. Limited benefit for strong baselines.
PROBE, an optimization framework for LLM agents in drug design, resolves the conflict between binding affinity and druggability. Through controlled edit probing and pocket-specific site mapping, it guides a multi-agent loop (affinity, druggability, co-optimization) on CrossDocked2020 with SOTA results.
A legal framework for agentic AI tort liability. The paper proposes three interaction types (autonomous drift, pure tool use, collaborative planning) and uses interaction logs as primary evidence to determine where liability attaches. Introduces a "Reasonable Agent" standard based on constraint verification and forensic logging.
Study of LLM agent faithfulness in a Texas Poker simulator. Researchers measure the gap between stated reasoning and actual actions by decomposing the problem into two steps: reasoning-conclusion and conclusion-action. The two steps exhibit opposite behaviors.
Post-training quantization (PTQ) reduces reasoning model accuracy and increases chain-of-thought length. 52% of failures involve correct intermediate answers not output as final answers. A training-free logit penalty on overthinking markers ("wait", "but", "alternatively") reduces CoT length by 12-23% while preserving accuracy across 5 models (1.5B-32B) and 5 benchmarks.
OPCD method to improve large models using weak critics. Instead of weak supervisors as labelers, they guide revisions. Progressive on-policy critique distillation filters high-quality critiques and distills critic-guided behavior into strong models via adaptive self-teacher signals. Results on reasoning and alignment benchmarks.
FoLoRA, a forgetting-aware optimization framework, adapts foundation models via generalized Rayleigh quotient. It balances target task performance and preservation of pretraining capabilities by penalizing adaptation directions with low utility relative to forgetting. Tested on math, code, and instruction-following adaptation.
Preference Delta Aggregation (PDA) aggregates weak preference signals from model pairs (e.g., Qwen3 4B vs 1.7B) via LoRA merging. Geometric Alignment Merging (GAM) aligns adapter subspaces before aggregation. On knowledge reasoning and agentic search benchmarks, PDA+GAM improves Qwen3 8B by +6.8 and +7.3 points respectively.
Study of effectiveness and efficiency of tool-calling in LLM agents. Authors show evaluation pipelines are sensitive to minor choices (random seed, system prompt, multi-turn templates) affecting leaderboard reliability. They identify two sources of computational waste in RL and propose two acceleration techniques without performance degradation.
Parameterized Diffusion Policy (PDP) conditions diffusion policies on continuous parameters embedded in a learned behavior manifold. The approach enables smooth interpolation between strategies and efficient adaptation to novel constraints without weight updates. Improved performance on multimodal benchmarks in simulated and real-robot experiments.
Method to evaluate collections of bivariate causal statements without access to ground truth. Authors introduce a compatibility score quantifying plausibility of induced causal models, without relying on faithfulness assumption. Applied to assessing causal claims from large language models.
Novel shielding framework for RL agents ensuring formal safety guarantees in MDPs with unknown transition dynamics. Uses robust MDPs (RMDPs) with sets of transition probabilities and LTL formulas. Combines shielding with PAC-learning methods to construct minimally restrictive shields while guaranteeing safety.
Modern LLMs systematically overestimate their competence and attempt unsolvable queries. Researchers propose Capability Self-Assessment (CSA), formulated as a policy-learning problem using reinforcement learning, to teach models to recognize their limits. RL significantly outperforms supervised fine-tuning, preserves original capabilities, and generalizes out-of-distribution.
MindZero is a self-supervised reinforcement learning framework training multimodal LLMs to infer human mental states without annotations. The model is rewarded for generating mental state hypotheses that maximize the likelihood of observed actions. After training, inference becomes fast single-pass and outperforms model-based methods in both accuracy and efficiency.
Delayed per-step reward attribution method for training LLM agents in multi-agent strategic interaction. An 8-billion-parameter open-source model trained with this approach matched or surpassed GPT-5 and won both Open and Efficient tracks at MindGames Arena benchmark (NeurIPS 2025).
MERIT, a multi-horizon memory retrieval framework for text-to-SQL agents, maintains episode-level and turn-level memory with learned retrieval policies optimized via reinforcement learning. On BIRD-Interact, it outperforms no-memory and static-retrieval baselines, improving success rate while reducing interaction turns.
UniKE, the first benchmark for cross-modality knowledge editing in unified multimodal models (UMMs), reveals a critical gap: text-side efficacy reaches 92% but VQA accuracy in image generation drops to 18.5%. A reasoning-augmented parameter editing method improves results by up to +18.6 percentage points.
SALSA adapts speech-aware LLMs via layer-wise learned steering vectors optimized with supervised objectives. Tested on children's speech, multilingual, and Mandarin-English code-switching, it achieves up to 46.8% relative improvement over zero-shot. Steering the encoder's later layers outperforms steering the LLM backbone.
Audit of 7 LLMs (US/China) on 2,520 responses to 60 legal-administrative prompts in English and Mandarin. Models default to the institutional framework of input language: 74.5% of English responses adopt US framework, 53.3% of Chinese responses adopt China framework. Risk of jurisdictional misselection when preferred language differs from applicable jurisdiction.
BOUTEF is a multilingual corpus from 2 countries (Algeria, Tunisia) covering fake news, authentic narratives, comments, and debunking. Includes MSA, Algerian/Tunisian dialects, Arabizi, French, English, and code-switching. Analysis shows fake news relies on emotionally charged narratives and sensational framing, while debunking adopts a factual, verification-oriented style.
BiGRU architecture enhanced with KAN (Kolmogorov-Arnold Network) block for legal document classification and summarization in low-resource multilingual setup. Evaluation on Bengali/English/transliterated corpus from Bangladesh: 67.96% accuracy in classification (F1=0.65), ROUGE-1/2/L scores of 0.38/0.23/0.31 in summarization. Ablation study shows KAN improves classification from 57.34% to 67.96%.
Comparative study of a two-hop Graph-RAG architecture versus standard vector-only RAG for cross-entity financial sentiment analysis. On 100 queries (30 direct, 70 relational), Graph-RAG improves entity recall (+6.4%, p<0.001) and answer relevancy for complex queries (+11.7%), with no quality degradation, modest 22.6% latency increase but 80% variance reduction.
TCAR-Gen combines query-conditioned graph neural networks, temporal evidence fusion, and chain-of-trees reasoning for retrieval-augmented generation. Achieves 0.3738 Recall@5 on Victorian Crime Diaries benchmark, outperforming Vanilla RAG, Temporal RAG, and GraphRAG variants. Cross-model evaluation across GPT-OSS 20B to TinyLlama 1.1B shows robust retrieval coverage at smaller scales.
Study combining cognitive-linguistic features (first-person pronouns, absolutist words, negative emotions) with DistilBERT embeddings and Holographic Reduced Representation to detect depression in Reddit posts. Hybrid model achieves F1=0.94 vs 0.80 TF-IDF baseline, AUC=0.981.
TrustLDM is a trustworthiness benchmark for Language Diffusion Models (LDMs) covering safety, privacy, and fairness. Results show LDMs degrade alignment when malicious post contexts are attached to masked responses, regardless of context length. An automatic evaluation framework (TrustLDM-Auto) systematically identifies vulnerable configurations across all tested models.
AEyeDE introduces an attention-based attribution framework for detecting AI-generated text using attention matrices from a proxy Transformer model. A lightweight CNN learns discriminative representations from these attribution maps. The method outperforms text-only baselines, shows strong generator-specific detection, and demonstrates robustness under cross-dataset transfer and spelling perturbations.