Non-parametric estimators (KM-ARL and KM-ADD) for evaluating quickest changepoint detectors on finite and irregular sequence lengths. Uses survival analysis analogy to model detection probabilities under truncation. Python code provided.
HPML (Hodge-Projected Multi-agent Learning) stabilizes multi-agent learning by projecting the joint update field onto a metric-gradient component. The method uses Hodge-type projection in an L² space of vector fields, implemented via graph-based and amortized neural realizations. Results: improved stability and normalized returns on CTDE benchmarks.
Researchers use functional MRI (fMRI) to improve encoding models trained on ECoG (electrocorticography). By fine-tuning spoken language representations on fMRI, they achieve better ECoG predictions despite 100× lower temporal resolution. Performance improves with more fMRI training data.
arXiv study evaluating LLM integration in survey research for disaster preparedness. Five-stage framework tested on 946 Florida residents (Hurricane Milton 2024). Proposed A-TLM model outperforms classical imputation methods (RMSE 1.439 vs 1.496) under MNAR conditions with PMT theoretical structure.
Robustness study of Document Layout Analysis (DLA) pipelines used in RAG and long-document QA. Authors identify footprint bias and propose a lightweight auditing framework measuring block-level structural loss (B-SLR). On 1,000 pages with MinerU and PP-StructureV3, B-SLR correlates better with OCR instability (R²=0.727/0.916) than area-based metrics (R²=0.384/0.110).
IMLJD is a dataset of 3,613 Indian court judgments on matrimonial disputes (IPC Section 498A, Protection of Women from Domestic Violence Act, CrPC Section 482). Data from Supreme Court of India (2000-2024, 1,474 cases) and Karnataka High Court (2018-2024, 2,139 cases). Quashing petition success rates: 57.6% at Supreme Court vs 39.7% at Karnataka High Court. Dataset, code, and knowledge graph released open-source.
Conditional Entropy Shaping (CES) dynamically controls token-level entropy to balance reasoning conciseness and accuracy. Implemented on DeepSeek-R1-Distill-7B, CES penalizes high-entropy tokens on correct reasoning paths and rewards them on incorrect paths. Results: improved accuracy with reduced response length across 12 mathematical benchmarks.
Local inference of DeepSeek-V4-Flash (284B, 13B active) on 4x RTX 2080 Ti (~$2.5k). Custom Turing CUDA kernels + W8A8 quantization + heterogeneous offloading achieve 255 tok/s prefill. Open-sourced on GitHub.
Vercel AI SDK now includes native agent tools via createChatTools(chat). Chat SDK's read and write actions are wired directly into agents. Three presets (reader, messenger, moderator) scope available tools. Write operations require approval by default.
Vercel AI Chat SDK adds message.subject to access parent context (GitHub issues/PRs, Linear) and exposes native SDKs (GitHub, Linear, Slack) for direct API calls. Per-message caching optimizes repeated access.
Vercel Chat SDK now supports callback URLs on buttons and modals. Developers can pause a run on a card and resume it when someone clicks a button. Form data is included in the payload sent to the endpoint.
llm-gemini 0.32 adds support for gemini-3.5-flash model. Plugin updated to access Google's latest Gemini 3.5 Flash version.
Google AI Edge Gallery v1.0.13 and v1.0.14 add Gemma 4 with multi-token prediction, Pixel TPU support, experimental MCP, new skills, and chat history saving.
OpenAI integrates Google's SynthID watermark into DALL-E to mark AI-generated images. A verification tool detects these invisible markings, improving traceability of synthetic content.
AXON visualizes real-time concept activations in GPT-2 through a 3D force-directed graph. A Sparse Autoencoder decomposes the residual stream into interpretable features (geography, cities, languages) per generated token. Stack: TransformerLens + SAELens (backend), FastAPI WebSocket, Three.js (frontend). ~35ms/token on GPU.
PrivateScribe.ai, fully local open-source transcription platform (MIT license), announces v1 with signed macOS app. Stack: FasterWhisper, pyannote, Ollama, Vite/Flask/SQLite. 256-bit encryption, zero network calls, audit trail, speaker diarization. Built for clinics, law firms, therapists with HIPAA compliance.
Open-source tool to generate 3D objects with articulated, functional parts. Instead of diffusion (point-cloud blobs), the pipeline uses an LLM as a structured code compiler, generating native Blender Python code targeting specific scene graph nodes. Flutter/Three.js frontend, model-agnostic. Gemini recommended; local models still hallucinate on complex matrix transforms.
KV cache quantization benchmarks on RTX 3090 with Qwen 27B: TurboQuant overrated except TCQ (best at 2-3 bits), q5 underrated, asymmetric q4_0 beats symmetric q4_1. KLD exposes tail issues PPL hides, llama.cpp rotation matches turbo4 performance.
Codegraph, an open-source tool by Colbymchenry, reduces Claude/Cursor API calls by 94% using a pre-indexed knowledge graph (symbol relationships, call graphs). Tests show 3 calls vs 52 without on VS Code TypeScript, with 72-82% speedup.
Benchmark of small models for local meeting summarization on 6GB GPU. Qwen3.5:0.8b produces structured summary in 57s using 2.2GB VRAM. Granite 4 350M is faster (0.6-2.8s) but hallucinates (invents topics, confuses entities).
Bio-plausible backprop-free agent (Predictive Coding + distributional Hebbian plasticity) vs PPO on Pong: 57% vs 59%. The 2% gap stems from catastrophic forgetting under self-play dynamics, not lack of backprop. ~1500 lines of code released.
Graph spectral analysis (Fiedler value + Scheffer critical slowing down) predicts grokking 21k steps before loss convergence. Five reproducible CPU experiments: early detection, distinct structural fingerprints for grokking vs catastrophic forgetting, guided intervention preserves 91.7% vs 2.6%, 48x acceleration across sequential tasks. Limited to 2-layer MLPs and 1-layer transformers.
Nuxt MCP Toolkit now supports MCP apps with interactive HTML responses rendered inline by Claude and ChatGPT instead of plain text. Tools declared with the defineMcpApp macro can access pre-hydrated data and trigger follow-up prompts via the useMcpApp composable.
Odyssey releases Agora-1, a world model enabling up to four players to act simultaneously in an AI-generated world. Tested on GoldenEye (N64), the system uses two separate models for game state simulation and real-time rendering. Potential applications: collaborative robotics and AI agent training.
ByteDance releases Lance, an open-source multimodal model with 3B active parameters. Supports image/video generation and editing in a single framework. Trained from scratch on 128 A100-GPU budget.
rtk is a Rust CLI proxy that reduces LLM token consumption by 60-90% on common dev commands. Single binary, zero dependencies.
Shimmy: Rust inference server compatible with OpenAI API, Python-free. Supports GGUF and SafeTensors, hot model swap, auto-discovery, single binary. Free forever.
Code-Review-Graph builds a local knowledge graph for Claude Code. It creates a persistent map of your codebase, reducing tokens by 6.8× on reviews and up to 49× on daily coding tasks.
Mistral AI acquires Vienna-based Emmi AI, a physical AI startup, to expand its industrial client offerings across Europe.
Anthropic acquires Stainless, a startup specializing in SDKs and API tools. The acquisition strengthens Anthropic's control over generative AI infrastructure, amid competition from OpenAI and Google.
Sapient Intelligence releases HRM-Text 1B: 1B parameters trained on 40B tokens in 1.9 days for ~$1k budget. Outperforms Llama3.2 3B on MATH (56.2 vs 48.0) and DROP (82.2 vs 45.2), but lags on MMLU (60.7 vs 64.7 for Qwen3.5 2B). Specialized for multi-step reasoning tasks.
OpenAI deploys content provenance tools: Content Credentials, SynthID, and a verification tool to identify and validate AI-generated media.
Gemini 3.5 Flash now available on Vercel AI Gateway with improved coding proficiency, parallel agentic execution, and enhanced reasoning. Model defaults to medium thinking level for quality-cost balance. Integration via AI SDK with observability support and intelligent provider routing.
Study on non-uniform replay effectiveness in off-policy RL. Authors identify three key factors: replay volume, recency of transitions, and entropy of sampling distribution. They propose Truncated Geometric replay, which biases toward recent experience while maintaining high entropy, improving sample efficiency in low-volume regimes.
MemOCR is a multimodal memory agent that compresses long interaction histories into structured images with adaptive information density. Trained via RL with budget-aware objectives, it outperforms text-based baselines on multi-hop and single-hop QA benchmarks under extreme context constraints.
PersonaDual is an LLM framework that adaptively switches between objective and personalized reasoning based on context. Trained via SFT then optimized with RL using DualGRPO, it preserves personalization benefits while maintaining factuality and reducing interference.
Researchers propose UALM (Unified Agent Lifecycle Management), a governance framework for agentic AI systems in healthcare. The model addresses agent sprawl through five control-plane layers: identity registry, orchestration, PHI-bounded context, runtime policy enforcement with kill-switch triggers, and lifecycle management linked to credential revocation and audit logging.
R2M (Real-Time Aligned Reward Model) introduces a lightweight RLHF framework to mitigate reward overoptimization. Instead of relying solely on semantic representations, R2M leverages evolving hidden states from the policy model to align with real-time distribution shifts during reinforcement learning training.
New DDR-Bench benchmark evaluates investigatory intelligence of LLMs: autonomous ability to explore databases and extract insights without explicit queries. Frontier models show emerging agency but struggle with long-horizon exploration. Study distinguishes investigatory intelligence (setting own goals) from executional intelligence (completing assigned tasks).
Researchers show that combining acoustic and expectation-related ANN representations improves music identification from brain activity (EEG). Models pretrained on both representation types outperform baselines, with complementary gains exceeding seed ensembles. This approach enables general-purpose EEG models grounded in cortical encoding principles.