Developer fine-tuned Cohere Transcribe to add diarization (speaker identification) and timestamps. Model outputs parsable format with average temporal precision of ±0.097s. Supports up to 4 speakers per 30s, extensible to 32 with diarize_long.py script. Available free on Hugging Face.
Vector Policy Optimization (VPO) is an RL algorithm training language models to produce diverse solutions by anticipating multiple vector-valued reward functions. VPO replaces the GRPO advantage estimator and matches or beats scalar RL baselines across four tasks, with widening gaps as search budget grows.
OpenAI launches a ChatGPT PowerPoint plugin in beta, able to create presentations from notes and documents, and edit existing slides. Available worldwide across all tiers. OpenAI recommends saving important decks before use due to risks of accidental deletion.
Trained prompt injection detector using ml-intern and DeepSeek v4 Flash. DistilBERT achieves F1 99%, compressed to ONNX int8 (~65 MB), runs in browser via Transformers.js v3. Total API cost under $5 with DeepSeek.
ByteShape's CPU-5 quant for Qwen3.6-35B-A3B achieves 30% faster token generation than Unsloth UD-IQ4_XS on 6GB VRAM laptop GPU, with slightly slower prefill speed. Tested on RTX 3060 with 65536 token context.
Quantization benchmark on Qwen3-Coder-Next using 3× R9700 PRO. UD-Q5_K_M outperforms MXFP4_MOE on all quality metrics (94% vs 89.4% top-1 accuracy, KL divergence 0.0217 vs 0.0746) with negligible speed penalty (~10% decode). Unsloth's dynamic precision approach exponentially reduces cumulative errors on long outputs.
New Qwen-27B-IQ4_KS quantization optimized for 16GB NVIDIA GPUs via ik_llama.cpp. 14.1GB model delivers performance comparable to previous IQ4_XS, 1.5-1.75x faster, 105k token context window. Tests: Needle In Haystack 100k passed, perplexity 71.10.
Rocky's voice (Project Hail Mary) cloned in two days via open-source pipeline. Audio extraction (ffmpeg + demucs), transcription (Whisper), diarization (pyannote), then RVC v2 training on 2:10 min audio. Trained .pth model (55MB) and code public. Tested XTTS v2 / YourTTS / RVC v2 / OpenVoice v2.
Neural method to estimate pairwise conditional mutual information in masked diffusion models (MDMs). Framework uses hidden states from pretrained MDMs with supervision from ground-truth MI computed from model's conditional distributions. Applied to Sudoku and protein sequence generation (ESM-C), reduces inference forward passes by 3-5x via MI-guided parallel decoding while outperforming entropy-based methods.
DivSkill-SQL optimizes Text-to-SQL ensembles through residual skill learning on failed examples, without fine-tuning. On Spider2-Lite, gains of +11.1 pts (Snowflake) and +8.3 pts (BigQuery) vs baseline. Skills transfer across SQL dialects without retraining.
New ECUAS_n metric family for evaluating uncertainty-augmented systems that output predictions and uncertainty scores. Formalized as proper scoring rules, they enable tuning trade-offs between prediction errors and uncertainty imprecision per use-case. Validated on classification, generation, and TriviaQA.
Insights Generator is a multi-agent system for diagnosing LLM agent failures at corpus scale. It formulates and tests hypotheses across execution traces to produce evidence-backed insight reports. Human experts using IG improve performance by 30.4pp; coding agents show consistent gains.
AgentCo-op is a retrieval-based synthesis framework that composes reusable skills, tools, and external agents into executable workflows through typed artifact handoffs. Applied to genomics and coding/math benchmarks, it coordinates specialized agents without global topology search and reduces per-task cost versus multi-agent baselines.
ML study on 1,375 patients: predicting myocardial ischemia from non-contrast CT calcium scans. XGBoost+SHAP model combining Agatston score, 8 calcium-omics features, and age. Results: 98.9% precision, 79.2% sensitivity, 87.7% F1. Calcium-omics features significantly improve performance vs clinical variables alone (p<0.05).
Novel Pseudo-Siamese architecture (FF-BPSN) for planning dialogue paths toward predefined targets. Uses two bidirectional transformer decoders with forward-focused module. Tested on DuRecDial and DuRecDial 2.0, significantly improves target-oriented proactive dialogue systems.
Study on reducing sycophancy (model agreement even when user is wrong) using off-the-shelf persona vectors. Vectors steered toward doubt/scrutiny reduce sycophancy to 68–98% of CAA's (Contrastive Activation Addition) effect while maintaining accuracy. Sycophancy is a persona-level property, not a single steerable direction.
Ishigaki-IDS-Bench is a benchmark for evaluating generation of Information Delivery Specification (IDS) XML files from BIM requirements. On 166 expert-validated examples in English/Japanese, the 10 best LLMs reach 65.6% macro F1 for content agreement, but only 27.7% pass the IDS Content audit. Models struggle to generate XML conforming to IDS standards and IFC vocabulary constraints.
Comparative study of RAG systems for Khmer. BGE-M3 outperforms Jina-Embeddings-v3 and Qwen3-Embedding in dense retrieval (Hit Rate@3: 0.285). Evaluation of 5 generators (Qwen3, Qwen3.5, Sailor2, SeaLLMs-v3, Llama-SEA-LION-v2) on 200 QA pairs using 6 RAGAS metrics. No single model dominates all criteria; retriever selection remains the bottleneck.
Study on quantization of LLaMA-3.1 (8B) for qualitative analysis. 8-bit models maintain best precision; 4-bit, 3-bit, and 2-bit models suffer hallucinations. A multi-pass verification method reduces errors and stabilizes results, making low-bit models viable for qualitative research.
Comparative study of four chunking strategies (Recursive, Khmer-Aware, Sentence-Based, LLM-Based) for RAG on Khmer agricultural documents. Recursive chunking with 300 characters achieves best performance: L2 distance 0.4295, Answer Relevance 0.8663, Khmer IoU 0.6441. Statistically significant improvement over Sentence-Based (p=0.0121).
Hyper-Align introduces a hypergraph-native framework for LLMs. The method compiles hypergraph structures into tokens via HIDT-O (hybrid template) and HIP (incidence projector), preserving high-order associations. Evaluated on HyperAlign-Bench, it outperforms existing methods on vertex and hyperedge tasks.
arXiv study on implicit Chinese toxicity attacks (CITA). Three-stage red-teaming framework (harmful intent learning, implicit toxicity enhancement, obfuscation rewriting) generating evaluation data. Seven tested detectors show 69.48% average miss-detection rate. Defense model CITD fine-tuned on CITA data improves robustness.
New arXiv paper introduces the 'Representation Gap', a metric related to neural network generalization error with better asymptotic dynamics. Authors derive precise asymptotic equivalence governed by task intrinsic dimension, validated on synthetic and realistic datasets.
Mean-field theory of dropout as perturbation of critical signal propagation at edge of chaos. Authors derive scaling laws and show smooth activations and ReLU-like activations form distinct universality classes. Front-loaded dropout schedules reduce test loss at no extra computational cost.
AiraXiv is an AI-driven open-access preprint platform designed for human and AI authors. It integrates AI-augmented analysis, reader feedback, and Model Context Protocol (MCP) for automated interactions. Deployed as submission platform for ICAIS 2025.
BlockFormer, a transformer architecture, infers parameters from interaction maps (Hi-C) by formulating the problem as a generic inverse problem. The method uses synthetically generated data from a custom simulator to accurately localize centromeres across multiple species.
A framework uses an LLM to translate natural language queries into deterministic spatial operations against a PostGIS database. Tested on Massachusetts transportation safety data (crash records, roadway attributes, schools, bus stops), the system validates 29% of erroneous queries through a rule-based layer, preserving reproducibility while democratizing data access.
PeakFocus is a unified framework for electricity load peak forecasting (ELPF), simultaneously predicting peak timing and intensity. It combines a peak-aware pipeline with triple hybrid loss, a multi-scale peak locator, and a location-aware decoder to overcome two-stage approach limitations. Evaluated on ELC and WLEL datasets.
Study of harness engineering for inference-time alignment of LLM agents. Authors decompose harnesses into task decomposition and guided execution mechanisms. They identify failure modes (over-decomposition, over-pruning) and show partial harnesses specifying only initial steps can outperform fully structured workflows.
Extension of Equilibrium Propagation framework to skew-gradient systems with demonstrated equivalence between deep Energy-Based Models and Hamiltonian neural networks. Applied to diffusively coupled Fitzhugh-Nagumo neuron networks, showing stationary solutions admit spatial Hamiltonian structure enabling Hamiltonian Echo Backpropagation methods.
HES (High-Entropy Sum) is a training-free metric for selecting high-quality reasoning data in LLMs. Tested across SFT, RFT, and RL paradigms, it achieves full-dataset performance using only the top 20% of samples, significantly reducing computational overhead.
VBFDD-Agent is a vehicle battery fault detection and diagnosis agent for electric vehicles using large language models. The system converts lithium-ion battery signals into natural language descriptions, integrates historical case retrieval and local maintenance manuals to generate structured, interpretable diagnostic results and maintenance recommendations.
Hierarchical variational framework for adapting pretrained diffusion models to reward-aligned objectives. Formulates test-time adaptation as a lightweight stochastic policy that amortizes per-step control. On 4x super-resolution: better perceptual quality with 5x faster inference than best baseline.
CUGA introduces a modular governance system for generalist LLM agents in enterprise settings. Through five enforcement checkpoints (Intent Guard, Playbook, Tool Guide, Tool Approvals, Output Formatter), the platform enforces policies without model fine-tuning, ensuring compliance and auditability across compound workflows.
Computational model of Arabic inflectional morphology focused on broken plurals. Reverses traditional root-and-pattern paradigm into pattern-and-root. Applied to 3,200 noun entries with 160 inflectional classes (22 triliteral patterns, 3 quadriliteral). Formal separation of inflection, derivation, and semantics.
New inference-time method for flow models: Conflict-Aware Additive Guidance (CAR) corrects off-manifold drift when composing multiple constraints. Dynamically detects and resolves gradient conflicts. Validated on image editing, planning, and control tasks.
DDS (Declarative Data Services) is an architecture for structured agentic discovery of data-system compositions. Addressing unbounded agentic discovery failures, the framework decomposes search into typed sub-searches via four contracts (intent, operator DAG, skills, runtime attribution). Tested on a trading-backend workload, DDS converges where unbounded approaches fail.
I-SAFE is a post-hoc auditing framework for scientific AI models based on the Wasserstein Coherence Metric (WCM). It evaluates whether model predictions reflect domain structure or exploit statistical shortcuts. Tested on drug-target interaction prediction (DeepConvDTI, DeepDTA, TAPB), it reveals distinct distributional response profiles invisible to accuracy metrics.
Method for creating high-quality embeddings for Horn logic reasoning. Authors use triplet loss with three innovations: anchors with repeated terms, balanced easy/medium/hard examples, and periodic emphasis of hardest cases. Evaluation across multiple knowledge bases.
GHI is a Graphormer-based framework for aspect-based sentiment analysis (ABSA). It uses a bipartite hypergraph structure to represent token-hyperedge incidence relations, integrating linguistic and semantic signals. With 247M parameters, GHI outperforms DeBERTa on six SemEval benchmarks and approaches Flan-T5 11B performance on ISE.