3148 articles
ChartDesign fine-tunes LLMs (Phi3, Qwen3, InternVL2.5) via LoRA to automatically generate chart design attributes from tabular data. Trained on curated corpus (PewResearch, CharXiV), the system achieves 84% accuracy on held-out test set vs 53% baseline, generalizing to unseen domains.
Qualitative study with 22 creative writers on intentional AI refusals in writing assistance. Researchers explore how refusals (saying "no") could introduce reflective friction rather than blind compliance, depending on context (planning, drafting, reviewing) and individual preferences.
A bank develops a conversational AI agent to triage fraud, scam, and dispute cases. The agent uses LLMs to ask targeted questions and route customers to appropriate specialist teams. Evaluation via synthetic digital twins simulating realistic dialogues. Result: +30.6% improvement in classification accuracy with compliance guardrails.
WorldString is a neural architecture modeling the state manifold of real-world objects from point clouds or RGB-D video streams. Designed as a differentiable digital twin, it serves as a foundational building block for physical world models integrating policy learning and neural dynamics.
SkillGenBench is a benchmark for evaluating skill generation pipelines for LLM agents. It covers two regimes: task-conditioned generation and task-agnostic generation, with procedural sources grounded in repositories or documents. Experiments reveal substantial performance variation and distinct failure modes between software repositories and long-form documents.
New metric to evaluate XAI methods without ground-truth, based on continuous input perturbation. Measures sufficiency and necessity of attributed information. Also proposes trainable XAI method as adapter on black-box models, generating causal explanations without degrading performance.
Comprehensive study of AI-assisted research systems through April 2026. LLMs excel at structured, retrieval-grounded, and tool-mediated tasks but remain fragile for genuinely novel ideas and scientific judgment. End-to-end autonomous systems have not yet consistently reached major-venue acceptance standards. Human-governed collaboration is the most credible deployment paradigm.
KVDrive is a multi-tier KV cache management system for long-context LLM inference, orchestrating cache placement across GPU/DRAM/SSD, pipeline scheduling, and cross-tier coordination. The prototype achieves 1.74x higher throughput than state-of-the-art systems while preserving accuracy.
LAR (Latent Action Reparameterization) compresses LLM agent action spaces by learning semantic multi-step latent actions. This reduces effective decision horizon and inference costs while preserving expressiveness. Across benchmarks, LAR decreases action tokens and wall-clock inference time without degrading task success rates.
Paper introducing trace-based evaluation to detect when agents hit business KPIs while violating behavioral constraints. In hotel pricing with hidden competitor state, authors show PPO variants fail trace alignment while behavior cloning and Trace-Prior RL better preserve price/bid distributions and rate discipline.
MM-ToolBench is a benchmark for omni-modal tool-using agents in real-world workflows. 100 executable tasks (customer service, intelligent creation), 27 MCP servers, 324 tools. Closed-loop multimodal verification: agents execute, inspect, and self-correct. Claude Opus 4.6 achieves 32% success vs 94% human baseline.
AMR-SD introduces asymmetric meta-reflective self-distillation to improve token-level credit assignment in LLM reinforcement learning. The method compresses diagnostic signals into self-generated Socratic hints and uses Causal Information Gain with asymmetric ReLU-gated threshold for sparse token-level advantage modulation, preventing late-stage training collapse.
OCCAM is a framework for explaining black-box image classifier decisions through causal visual concepts. It discovers concepts in open-set manner, localizes them via text-guided segmentation, and measures causal contribution through object-level interventions. OCCAM aggregates interventional evidence to induce a structured ontology revealing concept dependencies and systematic model biases.
Brain tumor segmentation method using multimodal MRI with virtual nodes and dynamic graph neural networks. One-stage framework handling missing modalities through adaptive adjacency matrices and heterogeneous weight matrices. SOTA results on BRATS-2018/2020 with incomplete modalities.
QSTRBench is a benchmark evaluating LLMs' ability to reason with qualitative spatial and temporal reasoning (QSTR). It covers 9 calculi (Point Algebra, Allen's Interval Algebra, RCC-5/8/22, etc.) with composition tables, converse relations, and conceptual neighbourhoods. Tested models outperform guessing but none answer all questions correctly. RCC-22 proves most difficult.
SD-Search introduces on-policy hindsight self-distillation for search-augmented reasoning agents. A single model acts as both student and teacher: the teacher, conditioned on past rollout outcomes, guides the student via token-level Jensen-Shannon divergence at query positions. No external teacher model or additional annotations needed.
ProRL is a programmatic reinforcement learning framework for combinatorial optimization (job shop scheduling). It generates interpretable policies as human-readable programs via a domain-specific language (DSL-S), exploring the program space through local search and Bayesian optimization. Outperforms classical heuristics and DRL baselines with minimal training episodes.
arXiv paper on spatial limitations of MLLMs in multi-agent environments. Models suffer from a "Cartesian Illusion": lack grounded 3D topological understanding. Authors propose an Epistemic Sensory Bottleneck module with Anchor-Based Embodied Spatial Decomposition CoT to improve second-order spatial inference (Theory of Mind). Zero-shot baseline: 42% accuracy.
PPR-GDE, an RL method for open-ended generation, uses pairwise preference rewards and group-based diversity to prevent diversity collapse. Without scalar rewards, it preserves subjective evaluations and encourages semantic dispersion within response groups.
Position paper arguing that agent generalization requires environment scaling—expanding the distribution of executable rule-sets agents interact with, beyond trajectory or task scaling. Proposes unified taxonomy separating trajectory scaling, task scaling, and environment scaling. Contrasts programmatic generators with generative world models for constructing scalable environments.
Study of 124 countries covering 99% of global GDP. Task-based automation exposure measure: 3.3% in South Sudan to 61.6% in China. Distinguishes labor-substituting vs labor-augmenting automation. AI more prevalent in substitution in low-income countries, augmentation in high-income. Women disproportionately exposed to substitution.
A randomized controlled experiment shows LLM access significantly increases average performance, but gains are unevenly distributed. AI Interaction Competence (ability to elicit, filter, and verify outputs) predicts benefits, not GPA. A scaffolding intervention (conceptual maps) reduces outcome variance.
POST introduces an adversarial learning framework for multivariate time series anomaly detection. The model combines graph neural networks with minimax optimization over adjacency matrices to address spatial over-generalization. Evaluation on public and synthetic benchmarks with channel-wise anomaly localization.
TaskGround is a task-inference framework for household agents operating on complete scenes. It structures reasoning in three steps: grounding (extracting relevant context), inference (executable structure), execution (action sequences). Evaluated on FullHome (400 tasks), it improves success rates and makes Qwen3.5-9B competitive with GPT-4 while reducing token costs by 18x.
ConsumerSimBench, a benchmark built from 1,553 Chinese social-media topics and 23,122 reaction criteria, evaluates whether LLMs can reconstruct real consumer reaction patterns. Gemini-3.1-Pro covers only 47.8% of criteria, revealing a major gap between technical performance and consumer intuition. A multi-agent pipeline improves MiMo-V2.5-Pro from 32.9% to 37.6%.
Study of Compositional Geometry Routing Problem (CGRP), a generalization of routing problems covering points, lines, areas, and hybrid geometries. Proposes DiCon, a solver with differential attention and contrastive learning to handle asymmetry and enlarged action spaces. Results show strong performance, versatility, and superior generalization across diverse instances.
DocOS is a benchmark evaluating GUI agents capable of proactively searching online documentation to solve long-tailed tasks. Experiments reveal two bottlenecks: difficulty reliably locating relevant information and faithfully grounding retrieved instructions into precise GUI actions.
New zeroth-order hard-thresholding algorithm with variance reduction for ℓ0-constrained optimization. Addresses SZOHT's limitation on random directions by mitigating conflict between ZO gradient deviation and hard-thresholding expansivity. Improved convergence rates validated on ridge regression and black-box adversarial attacks.
LGBO (LLM-Guided Bayesian Optimization) embeds LLM semantic reasoning into every Bayesian Optimization iteration via a region-lifted preference mechanism. Tested on physics, chemistry, biology, and materials science benchmarks, LGBO reaches 90% of best observed value in 6 iterations for Fe-Cr battery electrolyte optimization, versus 10+ for standard BO.
New approach for learning generalized policies in classical planning using Relational Graph Neural Networks (R-GNNs). Authors introduce efficient lookahead search encoding and relational abstraction to improve scalability on IPC 2023 benchmark. Results outperform classical planner LAMA.
EHV is a JIT compiler architecture embedding formal verification of AI governance policies directly into the inference pipeline. Using CRDTs and TEEs, it achieves sub-millisecond formal determinism (SMFD) and reduces governance latency from days to O(1), eliminating the trade-off between deployment velocity and compliance.
GVG (Generative Visual Grounding) uses an EEG-to-image generative model to translate brain activity into visual images, bypassing text-only alignment. Tested on GVG-X-Omni (170M tuned params) and GVG-Janus (trimodal), the framework improves EEG understanding and visual generation by leveraging MLLMs' visual priors.
LAST-RAG proposes a method for selecting stochastic degradation models to estimate remaining useful life (RUL). It combines observed trajectories and domain context via retrieval from a local evidence bank, with RCRUS mechanism to prevent premature model elimination. Experiments show outperformance versus statistical and prognostic baselines.
LMAC leverages LLM reasoning to design communication protocols in MARL, enabling agents to reconstruct the underlying state uniformly and accurately. The approach iteratively refines protocols using an explicit state-awareness criterion. Experiments on MARL benchmarks demonstrate substantial performance gains over prior baselines.
DuIVRS-2 is an LLM-based Interactive Voice Response system deployed at Baidu Maps for large-scale POI attribute acquisition. Using FSM-guided data augmentation, selective generation, and Chain-of-Thought mechanisms, the system processes 0.4 million calls daily with 83.9% Task Success Rate and 130ms latency.
PAIR is an internal reward model for multi-step LLM training via GRPO. It combines a hidden-state probe (belief consistency) with a lightweight attention head to generate dense step-level reward signals without external model calls or ground-truth dependencies.
KISS introduces a Knowledge Infrastructure (KI) enabling AI agents to execute complex Earth science simulations. On 3,000 trials, KI-equipped agents produced physically plausible simulations in 84% of cases vs. <40% without KI. An automated Knowledge Dissection Toolkit (KDT) generated 119 KIs across 14 Earth-science domains, showing operational expertise is structured and extractable rather than ad hoc.
arXiv paper on supervised fine-tuning (SFT) effectiveness for LLMs. Authors show SFT primarily removes noise-like token interactions but rarely acquires reliable new ones. The denoising phase is extremely brief; continued fine-tuning introduces overfitted interactions. Implications for early stopping and LLM training.
PuppyChatter is a software framework designed to simplify development of LLM-based AI applications. It combines the simplicity of vendor-specific SDKs with vendor-neutrality principles of abstraction frameworks, reducing complexity and vendor lock-in risks.
Amazon Music deploys a robust neural sparse retrieval system for large-scale music search. The system handles misspellings, transpositions, and phonetic variations with 91.4% recall@10 on 6M documents, outperforming trigrams (57.7%). Inference-free architecture with granular subword tokenization (max 3 chars) and zero online latency.
Authors identify that trajectory entanglement in Rectified Flow stems from nonzero divergence regions in the learned velocity field. They propose an offline correction that attenuates the divergent component during coupling generation, with no deployment overhead. Improvements validated on 2D benchmarks and image generation.
EXG is an experience graph framework for self-evolving LLM-based agents. It organizes successes and failures into structured, relational representations, enabling real-time cross-task experience reuse and offline reuse as external memory. Tested on code generation and reasoning benchmarks, EXG outperforms reflection and memory-based baselines.
EGI is a multimodal framework to monitor unconscious emotions of Scrum Masters in real-time. The system combines speech-to-text transcription (WER 10%), prosody analysis, emotional vocabulary matching, and context-aware suggestions via open-source multi-module API. Testing shows significant improvement in emotional awareness during simulated agile meetings.
Novel approach to extend Knowledge Graphs for French cultural heritage. Authors introduce WJoconde, a multimodal KG integrating text and images, with three variants and a benchmark for Knowledge Graph Completion. They propose a framework combining LLMs and Vision-Language Models for automated data extraction and validation, improving KG reliability.
SAPO improves generative recommendation by aligning reinforcement learning optimization to individual reasoning steps. Instead of assigning a single advantage to the entire response, SAPO computes separate group-relative advantages for each reasoning step and SID token, stabilizing training and outperforming baselines across three real-world datasets.
Dual-process memory architecture for scientific agents: decouples episodic window (10 messages) from semantic consolidation (3 tokens/message). Evaluation on 15,000 messages across 6 LLMs (OpenAI, Anthropic, Google): maintains 70-85% accuracy at 10,000 messages with 62% fewer tokens. Identifies trade-offs: Dual Process excels at numeric/temporal queries, RAG for historical retrieval.
GraphMind automates complex operational workflows by extracting action graphs from human resolution traces, then executes them via a multi-agent engine with LLM reasoning. An adaptive reinforcement mechanism (ATR) optimizes successful paths. Deployed across 4 cloud services, the system outperforms a Trace-RAG baseline with a 4.95/5 expert review score.
NeuSymMS is a hybrid neuro-symbolic memory system for LLM agents. It couples neural fact extraction from dialogue with a CLIPS-based expert system that classifies, deduplicates, and reconciles facts. Knowledge is stored as subject-relation-value triples in a relational database, with short/long-term memory and access-based promotion.
MEMOIR, a memory-guided tree-search framework, automatically synthesizes solvers for combinatorial optimization using LLMs. With a two-level memory hierarchy (branch-local and global), it achieves 96.7% solution validity across 7 problems (scheduling, routing, packing), outperforming baselines by 9.2 points and reducing run-to-run validity variance by over an order of magnitude.
ResDreamer, a hierarchical world model, reconstructs residuals at each layer to progressively abstract visual dynamics in a self-supervised manner. Without domain-specific knowledge, it achieves state-of-the-art sample and parameter efficiency for RL in 3D adversarial environments. Code released.
Theoretical study of multi-objective evolutionary algorithms for multi-party optimization (MPMOP). On MP-JCG benchmark, payoff-guided mutation requires Θ(n²) fitness evaluations to cross a gap region, while CPR-NSGA-II achieves O(n log n) via cross-party recombination. Runtime analysis on BPBOMST (multi-party minimum spanning tree) with instance-parameterized bounds.
Theoretical paper on computational challenges of token economics in LLM systems. Introduces the "Token Economics Trilemma": tensions between fine-grained valuation, low-latency execution, and allocation optimality. Identifies three technical areas: real-time value accounting, constrained resource allocation, and economic-aware system architecture.
D²Evo is an RL framework to enhance LLM reasoning through self-evolution. The method generates medium-difficulty training samples by mining anchors matched to model capability, then jointly optimizes a Questioner and Solver. Results: outperforms existing methods on mathematical reasoning benchmarks with <2K real examples.
Longitudinal study of 235M posts from 4M Bluesky users showing algorithmic feed exposure (News, Science, Blacksky) measurably shapes user language: semantic alignment, register formalization, psycholinguistic restructuring. Reposting is the strongest predictor of linguistic convergence across feeds.
QQJ is an evaluation framework for generative AI combining expert-designed multi-dimensional rubrics and LLM evaluator calibration on small high-quality annotation sets. Tested on text and image generation, QQJ shows stronger alignment with human judgment than traditional automatic metrics and unconstrained LLM-based evaluators.
CBT-Audio is a dataset of 1,802 patient turns from 96 public CBT recordings with expert-validated distress labels. Evaluation of 10 open-source audio language models shows audio improves distress estimation over text alone in 8/10 model families, with strongest gains when verbal content and vocal delivery diverge.
Learning-Zone Energy (LZE) is an online data selection framework for RL post-training of LLMs. Tested on Qwen 1.5B-8B across GSM8K and MATH, it retains 40% of training data per step while matching full-data baselines, with OOD gains of +45.9% on AIME25 and 36% FLOP reduction.
BoLT is an open-source benchmark for black-box optimization applied to LLMs. It covers hyperparameters, data mixtures, and prompts via lightweight surrogate models fitted to thousands of real experiments. Benchmarking Bayesian Optimization and BBO methods reveals gaps in existing approaches.
CardioThink, a physician-inspired MLLM framework, structures ECG diagnosis through explicit reasoning stages (rhythm, conduction, morphology, impression) to enhance interpretability. Structured Set Policy Optimization (SSPO) aligns clinical reasoning without manual annotations, outperforming direct prediction approaches across ECG benchmarks.
CyberCorrect formalizes LLM self-correction as a closed-loop control system. A tri-modal error detector (self-consistency, verbalized confidence, logic-chain verification) and type-directed correction controller achieve 79.8% accuracy on CyberCorrect-Bench (440 reasoning tasks), +6.2pp over existing methods, reducing overshoot by 41% via convergence control.
Response-free item difficulty modelling for multiple-choice questions using fine-tuned transformers. End-to-end approach on item wording eliminates manual feature engineering. Multi-task variant with auxiliary QA objective delivers significant improvements in small-sample regimes.
A2RBench is an automated pipeline for generating formally verifiable abstract reasoning benchmarks. Using programmatic verification (cycle consistency), it eliminates hallucinations and scales task variations. Evaluations show current LLMs score 39.8% vs 68.5% for humans, and struggle with complex 3D tasks.
Systematic study of reasoning faithfulness in Vision-Language-Action (VLA) driving models. Analysis of 300 Alpamayo-R1-10B inferences across 100 PhysicalAI-AV scenarios reveals: reasoning fidelity 42.5%, 94 missed pedestrians, 97.7% trajectory fragility under visual perturbations, 48.3% reasoning-action consistency. Proposes four-component safety architecture.
CAM-Bench is a Lean 4 benchmark of 1,000 computational and applied mathematics problems (optimization, numerical linear algebra, numerical analysis) adapted from textbooks with locally recovered context via dependency-recovery pipeline. Evaluation of LLMs and formalization agents reveals failures in tracking local assumptions and long-horizon control in Lean.
QE-Catalytic-V2 is a unified graph-text multimodal LLM for catalytic materials. It integrates property prediction and inverse design in a single shared representation space, eliminating distribution shifts between decoupled models. Demonstrates superior performance on relaxed-energy prediction and inverse design tasks.
ChemVA framework advances LLM understanding of chemical reaction diagrams by addressing visual and semantic bottlenecks. Uses Visual Anchor mechanism for functional group detection and semantic alignment to activate chemical reasoning. Achieves 92.0% structural recognition accuracy on OCRD-Bench with ~20 percentage point gains across 9 diverse LLMs.
CAREBench is a benchmark evaluating LLMs' emotion understanding through cognitive appraisal reasoning. Tested on 6 models with complete inferential chain annotations (first/third-person perspectives), it shows stronger models match humans on some tasks but fall short on appraisal reasoning and positive emotion recognition.
Shallow neural network agents master the card game Schnapsen through reinforcement learning. RLBot, trained via asynchronous Monte Carlo updates, outperforms MLPBot (supervised imitation) and achieves statistically significant wins against RdeepBot, a search-based baseline. Combining learned value functions with deeper lookahead during gameplay improves performance.
MADP is a multi-agent architecture for enterprise document automation, combining deep learning classification and LLM extraction with human validation. Deployed on 955 real documents, it achieves 97% full-pipeline automation and reduces FTE requirements by 70%. 98.5% document-level accuracy with human-in-the-loop; 69% CO2 reduction vs manual processing.
Analysis of 130,882 images from 368 remix parties on Artbreeder (13 months). Images converged toward common thematic attractors (steampunk, alien architecture) while becoming simpler. Paradox: more novel parents produced more complex, liked children, yet users preferred remixing less novel images.
Automated heuristic discovery via continuous optimization in latent space. Encoder maps discrete programs to continuous embeddings, differentiable surrogate model predicts performance, invertible normalizing flow regularizes optimization trajectory. Evaluation on TSP, CVRP, KSP, and Online Bin Packing shows competitive results against evolutionary baselines.
ECC, a query clustering algorithm, calibrates semantic embeddings through model comparisons to align surface semantics with latent LLM capabilities. Using a Bradley-Terry model, it improves capability ranking by 17.64 points over human-labeled baselines and 18.02 points over embedding-based baselines, with applications to query routing.
Systematic investigation of logicality in LLM scientific reasoning. Authors develop a logicality-enriched methodology with assessment criteria and data sampling methods for logicality-guided training. Experiments on three backbone LLMs using physics problems extracted from academic literature. Code released.
RAGA is an LLM-based autonomous agent for knowledge graph construction and retrieval-augmented generation. It combines CRUD operations, a ReAct loop with Read-Search-Verify-Construct constraint, and KG-vector synchronization for hybrid retrieval. QASPER experiments show gains in answer and evidence quality.
Large Reasoning Models generate traces aligned with human reaction times, but this alignment persists regardless of inference-time reasoning budget. Study across GPT-OSS-20B and GPT-OSS-120B: three effort levels, six cognitive tasks. Token allocation tracks fine-grained human difficulty patterns and reflects a structure crystallized at training time, not modulated in real-time.
Critical study of LLM-based trading agents (FinCon, FinMem, TradingAgents, FinAgent, QuantAgent, FLAG-Trader). Reported Sharpe ratios do not constitute deployment evidence: temporal contamination, unmodeled frictions, and insufficient predictive calibration invalidate claims. Proposes P1-P6 protocol and modular architecture with LLM as audit interface.
Researchers present a framework for book-scale creative writing. Starting from public-domain novels, they build a multi-resolution scaffold (summary → chapters → scenes → full text) and train a long-context model on prompt-to-book trajectories. Goal: generate human literary prose rather than generic assistant-style text.
PersonaArena is a dynamic simulation framework for evaluating and improving persona-level role-playing in LLMs. It leverages a filtered corpus of user-generated social content, constructs a nuanced persona bank, and simulates multi-turn interactions in social environments. A multi-agent debating judge provides holistic and unbiased assessment.
DriveSafe is a framework for risk detection in autonomous driving scenarios. It generates spatially grounded captions enriched with motion and depth cues, then assesses risks using a fine-tuned adapter module on caption-risk pairs. Achieves SOTA on DRAMA benchmark.
PA-BDM improves Block Diffusion Models for document recognition by replacing bidirectional denoising with causal prefix-to-suffix denoising. Using Confidence-gated Structural Loss and Progressive Prefix Commitment, the 3B model achieves 71.6% higher inference throughput than MinerU-Diffusion 2.5B.
Study of autonomous AI agents in multi-echelon supply chains using MIT Beer Game. Reasoning models reduce costs by 67% vs human teams, but reveal an 'agent bullwhip effect': amplification of decision unreliability across echelons. A GRPO-based reinforcement-learning post-training framework using system-level rewards improves reliability and reduces tail events.
Study using Transcranial Doppler (TCD) and MOCAIP algorithm to predict brain vascular age. 168 healthy and 277 diseased subjects (stroke, Alzheimer's, MCI) analyzed. Model predicts accelerated aging in patients, with 3.69-year overestimation in healthy subjects.
EEG study of 27 participants analyzing neural mechanisms for detecting AI hallucinations. Researchers recorded brain activity during verification of image descriptions generated by an MLLM. Results show that misjudged hallucinations fail to trigger standard fact-verification neural pathways.
Unified framework for disease trajectory modeling in clinical AI, integrating factual forecasting, counterfactual estimation, and policy evaluation. Addresses treatment assignment bias, time-varying confounding, and observation bias to transform static predictions into treatment-sensitive dynamic estimates.
NGM is a training-free memory module for LLMs using a Causal N-Gram Encoder and Cosine-Gated Memory Injector. Tested on Qwen3 (0.6B-14B), it improves average performance by 0.5-1.2 points, with notable gains on code generation (+3.0 LiveCodeBench) and knowledge-intensive tasks (+3.03 GPQA).
Reasoning models outperform base LLMs on complex benchmarks. Study shows the advantage stems from a small set of early decision tokens (~8% on Qwen3-0.6B), concentrated in planning phases. Selective intervention by the reasoning model at these critical tokens restores performance without major computational overhead.
New paradigm for experience-driven learning in multimodal settings: agents learn to dynamically construct and organize memory based on task requirements and interaction history, rather than relying on fixed memory schemas. Adaptive memory design improves performance and generalization across multimodal tasks.
HT-GRPO, a hierarchical reinforcement learning method for diffusion multi-modal models, organizes optimization into three stages (global, structure, refinement). It solves multiple unmasking sequences and assigns differentiated rewards based on token importance. Tests on MMaDA and Lumina-DiMOO show gains on GenEval and DPG benchmarks.
Study of an open-source interactive atlas mapping 200+ participatory AI initiatives. Reproducible protocol for discovery, vetting, and harmonization of cases. Findings: initiatives concentrated in few countries, participation mostly in problem formulation and evaluation, rarely in model development.
A hypothesis proposes that LLMs encode concept lattices in their embedding geometry. The framework unifies the Linear Representation Hypothesis with Formal Concept Analysis (FCA), showing that linear attribute directions induce lattices via half-space intersections. Experiments on WordNet validate that embeddings capture logical and hierarchical structures.
GRID is an end-to-end framework for constructing security knowledge graphs from cyber threat intelligence articles. Using Qwen3-4B-Instruct, it combines graph extraction, text revision, and a task bank (multi-choice questions + regex) to generate stable rewards. On 249 CTI articles, the Task-bank Reward model achieves 84.62% precision, 64.91% recall, and 68.53% Avg F1.
VGGT-CD is a training-free pipeline for 3D change detection from multi-view images. It decouples cross-temporal registration from dynamic-change interference via joint keyframe inference and dense reconstruction purification. On the World Across Time benchmark, it reduces Absolute Trajectory Error by 44% outdoors and 59% indoors, 6× faster.
LLM-based behavioral planning framework for autonomous vehicles to anticipate pedestrian behavior. Evaluated on SUMO: 68% collision-free success rate zero-shot (vs 17.7% deep RL), 96% with few-shot episodic memory. Interpretable decisions with cross-behavior transfer across scenarios.
On-device AI system for ecological monitoring in remote areas. Architecture separates visual perception from reasoning using dynamic knowledge base, eliminating cloud dependency and continuous retraining. Collaboration with biologists and Indigenous communities for ethical AI co-development.
Post-hoc multimodal alignment method using relative representations at token level to match separately pre-trained encoders with limited paired data. Learns learnable anchors in each modality space to induce consistent cross-modal similarity patterns. Outperforms existing methods on zero-shot classification, cross-modal retrieval, and zero-shot segmentation.
F³A is a training-free router for visual token pruning in vision-language models. It selects relevant visual tokens via question-conditioned cues without extra LLM forward passes, reducing inference costs while maintaining performance across model scales.
Decentralized building management framework integrating DAOs, digital twins, LLMs, and blockchain for transparent governance. System evaluated on cost efficiency, scalability, data security, and usability via System Usability Scale and expert interviews.
HASP converts textual skills for LLM agents into executable Program Functions that actively intervene in the agent loop at failure-prone states. The framework achieves 25% improvement on web-search vs ReAct and 30.4% gain on math/coding vs Search-R1 through inference-time intervention, post-training, or self-improvement mechanisms.
arXiv paper demonstrating that multi-turn conversations in low-resource African languages (Afrikaans, Kiswahili, isiXhosa, isiZulu) bypass safety mechanisms in commercial LLMs. Testing ChatGPT, Claude, DeepSeek, Gemini, and Grok shows jailbreak rates from 52.7% to 83.6% depending on model. Translation quality is the critical success factor.
SomaliWeb v1: Somali corpus of 819,322 documents (~303M tokens) with BPE-16K tokenizer and language-identification benchmark. Reveals critical defects in HPLT v2 (17.3% exact duplicates, 56.1% mojibake, 10.7% near-duplicates). Tokenizer 40.2% more efficient than cl100k_base on FLORES-200.