LLaMA.cpp releases version b9784 with major optimizations for hexagon-based MM operations, including 32x32 tiled weight repack, improved dyn.quant handling, and unified kernel parameter management. The release includes new binaries for macOS (arm64 and x64), iOS, and multiple Linux architectures with support for Vulkan, ROCm, and OpenVINO.
The Buddy System uses a Rust entropy monitor to detect per-token uncertainty in local Gemma 3 4B inference, routing only uncertain tokens to Sonnet via NER-gated span extraction and semantic retrieval. Benchmarks show it achieves 71.4% accuracy at $0.21, outperforming the Anthropic Advisor pattern (62.9% at $0.44) across seven Hugging Face datasets, with a key improvement on SQuAD v2 by routing source passage chunks to the cloud model.
LRE is a CPU-only, language-model-free system that learns which interaction history units are load-bearing. It outperforms baselines in accuracy-cost balance, reducing peak context size by up to 52% and improving task completion by 37% in some cases. LRE achieves superior answer quality with 68% fewer tokens and requires no annotations or neural computation for training.
Beaver is an agent harness that extracts structured information from scientific papers by integrating multimodal evidence tooling, task scaffolding, and artifact-grounded autoresearch. It achieves 81.0 on the Gold-Referenced Attribute Score, outperforming frontier agents by over 23 points, with key gains on high-value attributes requiring cross-modal reasoning.
UltraQuant enables 4-bit KV caching for context-heavy agents, reducing P50 time-to-first-token by 3.47x in late rounds and boosting output throughput by 1.63x over FP8 KV baseline. It achieves this using FP8 queries, FP4 KV tensors, UE8M0 group scales, and native scaled-MFMA on AMD CDNA4 GPUs, with optimizations for decode-attention kernels and robust design choices like asymmetric K/V treatment and Walsh-Hadamard rotation.
UltraQuant introduces a 4-bit KV caching method tailored for context-heavy agent workloads. It achieves 3.47x reduction in P50 time-to-first-token in late rounds and 1.63x higher output throughput compared to FP8 KV caching, using FP8 queries, FP4 KV tensors, and native AMD CDNA4 scaled-MFMA support.
H-RePlan introduces a hierarchical replanning framework that separates device-local strategy recovery from global orchestrator replanning. It outperforms existing baselines by achieving higher completion and instruction adherence, with reduced token cost, through scope-aware recovery in multi-device agent systems.
UAV-VLN-FOV isolates the see-and-reach stage for precise evaluation of UAV navigation. 3DG-VLN enhances visual grounding and spatial alignment using dynamic 3D direction cues, achieving a 13.82% success rate improvement over baselines and validated in real-world trials.
JAMER introduces JamSet and JamBench, the first project-level game code dataset and benchmark on a professional game engine. Built from 8,133 verified Game Jam projects, it enables deterministic evaluation and reveals a capability cliff in AI models as project scale increases, with runtime pass rates dropping from 80.4% to 5.7%.
OneCanvas enables 3D scene understanding in Vision-Language Models by aggregating patch features onto a single panoramic canvas using 3D world coordinates. It achieves state-of-the-art performance on SQA3D and VSI-Bench, and generalizes to out-of-distribution data on SPBench, using significantly less training compute than existing methods.
User as Engram proposes storing per-user facts as surgical, hash-keyed edits to a memory table, leaving reasoning in a shared adapter. This design achieves 5.6x higher indirect-reasoning accuracy and maintains base-level reasoning performance, with a memory footprint 33,000x smaller than per-user LoRA. The approach enables disjoint user edits that compose losslessly, outperforming retrieval pipelines beyond 100 facts.
Data Intelligence Agents (DIA) deploy autonomous coding agents to streamline enterprise data workflows. The Query Generator matches or exceeds top published results on seven SQL benchmarks across four dialects, showing generalization through natural-language instructions and execution-based architecture.
NoiseTilt introduces NTRK, a reward-guided diffusion sampler that injects reward gradients via the noise term without altering the reverse kernel. By using a whitening operator, NTRK safely biases noise toward high reward, preserving sample quality while maintaining strong guidance. On aesthetic generation, NTRK achieves superior reward performance with 25 NFEs, reducing compute by 20× compared to state-of-the-art baselines.
BinTrack is a fully open-source spatial question answering agent that uses binary search over a robot's trajectory to locate answers. It achieves up to 22.8% higher accuracy than other open-source methods and matches closed-source model performance on the most challenging global category of the SpaceLocQA benchmark. The system also offers over 1.5x faster inference and introduces GangnamLoop, a real-world outdoor benchmark collected with a quadruped robot.
The llama.cpp project has released version b9786, introducing support for non-contiguous rows in normalization via OpenCL. This update is part of the ongoing development by the ggml-org team to enhance hardware compatibility and performance across various platforms. The release provides binaries for macOS Apple Silicon, Intel Macs, and iOS XCFrameworks. Linux users can access builds for Ubuntu x64, arm64, and s390x architectures using CPU, Vulkan, ROCm 7.2, OpenVINO, and SYCL backends. Android support is available for arm64 CPU devices, while Windows offers extensive options including CPU, CUDA 12 and 13, Vulkan, OpenVINO, SYCL, and HIP. The release also lists disabled builds for KleidiAI on macOS and openEuler platforms.
A user has reported a critical bug in the Hugging Face text-embeddings-inference library affecting Qwen3 and Gemma3 models. The issue arises when running inference on CPUs with concurrent requests, leading to significant accuracy degradation. Specifically, the Candle backend incorrectly skips attention masks for batches where all input sequences have equal lengths. This defect compromises the reliability of embeddings generated under these specific conditions. To address the problem, the author submitted a pull request containing a fix that was thoroughly tested on their local machines. The bug highlights potential stability risks in CPU-based embedding services handling batched inputs.
A new study decomposes agent memory into four core modules and evaluates 12 systems across five benchmark workloads. It finds no single architecture dominates, with performance dependent on alignment with workload bottlenecks, and reveals that localized maintenance is more cost-efficient than global reorganization.
A controlled study evaluates OpenPangu 1B and 7B models on Huawei Ascend 910B1 NPUs using weight-only and weight-activation quantization methods. Results show 8-bit weight-only quantization is lossless for both models, while 4-bit quantization is practical for 7B but harmful for 1B on reasoning, math, and code tasks. Ultra-low precision methods like 2-bit and binary fail, and W4A4 SmoothQuant produces non-finite perplexity, indicating extreme low-bit compression remains challenging.
The Hugging Face page for Llama 3.1 405B lists Featherless AI as a provider, but the test widget shows 'Failed to fetch' and featherless.ai does not list it as available. A similar issue is reported for Baidu's ERNIE-4.5-300B model.
Users report that Spaces tokens no longer function after a recent update. No generated files are being saved, disrupting workflow and model execution.