https://k20x.com/ai_setup/daily/Sourced daily research on local AI hardware, phone AI and agent sandboxes.enFri, 09 Oct 2026 06:44:01 +0000https://e2b.dev/customers/clickuphttps://e2b.dev/customers/clickupThu, 08 Oct 2026 00:00:00 +0000Agent sandboxes (E2B and peers)ClickUp Brain² uses E2B sandboxes to run model-written code on sensitive workspace data, leveraging isolated microVMs, prebuilt templates, and the ability to launch thousands of sandboxes quickly.https://github.com/ggml-org/llama.cpp/releases/tag/b11514https://github.com/ggml-org/llama.cpp/releases/tag/b11514Thu, 08 Oct 2026 20:17:43 +0000Runtimes and quantizationllama.cpp release b11514 includes a Musa FWHT fix and lists broad platform support, including macOS, iOS, Linux (CPU, Vulkan, CUDA, ROCm, OpenVINO, SYCL, Snapdragon), Android (CPU, Snapdragon), and Windows (CPU, OpenCL, CUDA).https://github.com/vllm-project/vllm/releases/tag/proto-v0.5.0https://github.com/vllm-project/vllm/releases/tag/proto-v0.5.0Fri, 09 Oct 2026 03:11:40 +0000Runtimes and quantizationvLLM proto-v0.5.0 has been released.https://github.com/ollama/ollama/releases/tag/v0.40.2https://github.com/ollama/ollama/releases/tag/v0.40.2Fri, 09 Oct 2026 03:37:30 +0000Runtimes and quantizationOllama v0.40.2 upgrades models downloaded with earlier versions for better performance and llama.cpp compatibility, temporarily keeping original copies as backups for safe downgrading.https://arxiv.org/abs/2607.09842https://arxiv.org/abs/2607.09842Fri, 09 Oct 2026 04:00:00 +0000Phone and edge AICorrections are reported for previous findings on system-prompt conditioning and hidden-state geometry in four open-weight language models, clarifying the curvature statistic and the reported norm quantity.https://arxiv.org/abs/2610.11447https://arxiv.org/abs/2610.11447Fri, 09 Oct 2026 04:00:00 +0000Open models for local useA benchmark is presented for closed-loop evaluation of LLM agents for embedded software development, where agents must translate requirements, self-verify, and iterate until the required device behavior is achieved.https://arxiv.org/abs/2610.10946https://arxiv.org/abs/2610.10946Fri, 09 Oct 2026 04:00:00 +0000Open models for local useAI4Fire evaluates large language models on wildfire tasks, comparing bare and grounded runs across various models and tasks, finding that grounding helps most when the addition carries the answer and simple rules are hard to beat.https://arxiv.org/abs/2610.10909https://arxiv.org/abs/2610.10909Fri, 09 Oct 2026 04:00:00 +0000Phone and edge AIPSCMIA is a power side-channel membership inference attack against embedded machine learning models that infers membership directly from power traces without requiring prediction probabilities or labels.https://arxiv.org/abs/2610.12445https://arxiv.org/abs/2610.12445Fri, 09 Oct 2026 04:00:00 +0000Open models for local useWhite-box deception detection via probes is scaled up for monitoring LLM agents, using a large deception dataset and a novel probe architecture to achieve high AUC in SHADE-Arena and detect introspective deception.https://arxiv.org/abs/2610.11605https://arxiv.org/abs/2610.11605Fri, 09 Oct 2026 04:00:00 +0000Open models for local useNanoProof is introduced as an open and efficient automated theorem prover in Lean 4, with all training data, tooling, pipeline, and weights released, focusing on compute efficiency for accessible training and evaluation.https://arxiv.org/abs/2610.11115https://arxiv.org/abs/2610.11115Fri, 09 Oct 2026 04:00:00 +0000Runtimes and quantizationResearch proves that a dynamic system paradigm can be used for model compression, where high-dimensional parameters are encoded by a trajectory index and recovered during decompression, distinct from other compression methods.https://arxiv.org/abs/2610.10933https://arxiv.org/abs/2610.10933Fri, 09 Oct 2026 04:00:00 +0000Runtimes and quantizationSpora is introduced as a new approach for spiking language models that jointly designs spike encodings and attention operators, using binary temporal weights and unipolar/bipolar binary spiking to improve performance with fewer time steps.https://arxiv.org/abs/2610.10926https://arxiv.org/abs/2610.10926Fri, 09 Oct 2026 04:00:00 +0000Phone and edge AIAn adaptive multi-discriminator Wasserstein GAN (MD-WGAN) framework is introduced for resource-constrained Internet of Vehicles environments, integrating reinforcement learning and game theory to manage machine learning workloads.https://arxiv.org/abs/2610.10854https://arxiv.org/abs/2610.10854Fri, 09 Oct 2026 04:00:00 +0000Phone and edge AIKDFP is presented as a novel methodology for white-box general knowledge distillation in large language models, developed through a first-principles approach to create efficient and private systems for edge devices.https://arxiv.org/abs/2610.10641https://arxiv.org/abs/2610.10641Fri, 09 Oct 2026 04:00:00 +0000Runtimes and quantizationResearch explores using large language models for next-visit diagnosis prediction, highlighting that existing reinforcement learning rewards can concentrate on a few correct diagnoses, leaving others uncovered, and that LLM tokenizers can split ICD codes into https://arxiv.org/abs/2610.10621https://arxiv.org/abs/2610.10621Fri, 09 Oct 2026 04:00:00 +0000Runtimes and quantizationResearch introduces constraint-induced self-organization via geometric radiation in coupled metric evolution systems, revealing a four-stage cycle and a novel wedge-shaped attractor topology.https://github.com/ggml-org/llama.cpp/releases/tag/b11516https://github.com/ggml-org/llama.cpp/releases/tag/b11516Fri, 09 Oct 2026 06:04:09 +0000Runtimes and quantizationA fix for a DMA ring overflow issue in the Hexagon IM2COL patch-embed kernel in llama.cpp prevents stale data from leaking into output when IC*KH exceeds 255. The solution involves retiring the oldest descriptor when the ring is full.https://e2b.dev/resources/introducing-e2b-secretshttps://e2b.dev/resources/introducing-e2b-secretsWed, 07 Oct 2026 00:00:00 +0000Agent sandboxes (E2B and peers)E2B Secrets is introduced to keep credentials outside the sandbox and inject them into HTTPS request headers. This allows agents to call external services safely.https://github.com/ollama/ollama/releases/tag/v0.40.1-rc0https://github.com/ollama/ollama/releases/tag/v0.40.1-rc0Wed, 07 Oct 2026 22:22:54 +0000Runtimes and quantizationOllama v0.40.1-rc0 drops a carried Metal residency patch for MLX because it is now upstream.https://github.com/ollama/ollama/releases/tag/v0.40.1https://github.com/ollama/ollama/releases/tag/v0.40.1Thu, 08 Oct 2026 02:09:43 +0000Runtimes and quantizationOllama v0.40.1 includes fixes for clef head reads on Windows, manifest symlinks on Windows, and removes an account step from CLI onboarding. It also drops a Metal residency patch for MLX as it is now upstream.https://arxiv.org/abs/2608.13545https://arxiv.org/abs/2608.13545Thu, 08 Oct 2026 04:00:00 +0000Agent sandboxes (E2B and peers)LittleLearner is a 5B-parameter LLM trained on LittleCurriculum, an 88B-token pretraining corpus tailored to U.S. elementary school material. This creates a developmentally restricted sandbox to study knowledge and skill acquisition in models.https://arxiv.org/abs/2609.02852https://arxiv.org/abs/2609.02852Thu, 08 Oct 2026 04:00:00 +0000Agent sandboxes (E2B and peers)The concept of "linguistic illegibility" is introduced, referring to scenarios where an LLM's linguistic outputs or features do not reliably represent its internal computation. This implies that security mechanisms relying on a model's language artifacts may bhttps://arxiv.org/abs/2607.25970https://arxiv.org/abs/2607.25970Thu, 08 Oct 2026 04:00:00 +0000Agent sandboxes (E2B and peers)Research addresses challenges in applying reinforcement learning (RL) to code optimization, where execution time drives the reward. It proposes making execution time learnable through a calibrated sandbox, composing correctness and speed in the RL environment,https://arxiv.org/abs/2610.10409https://arxiv.org/abs/2610.10409Thu, 08 Oct 2026 04:00:00 +0000Phone and edge AIRobotWorld is introduced as a simulation testbed for benchmarking multimodal agents for robot use across 84 tasks. It evaluates agents' ability to translate instructions and observations into physical task execution through robot interfaces.https://arxiv.org/abs/2610.10261https://arxiv.org/abs/2610.10261Thu, 08 Oct 2026 04:00:00 +0000Open models for local useResearch evaluates Small Language Models (SLMs) for reverse-engineering Machine Learning pipeline structures from source code. SLMs are assessed for their code understanding and classification abilities to extract ML pipeline stages.https://arxiv.org/abs/2610.09424https://arxiv.org/abs/2610.09424Thu, 08 Oct 2026 04:00:00 +0000Desk-side boxes (Mac Studio, DGX Spark, OEM)CoMoE is a communication-efficient MoE inference system designed to democratize MoE inference on commodity GPUs. It addresses communication bottlenecks on consumer GPUs, which lack high-bandwidth P2P interconnects, through novel host-centric routing.https://arxiv.org/abs/2610.09304https://arxiv.org/abs/2610.09304Thu, 08 Oct 2026 04:00:00 +0000Phone and edge AIemg2face demonstrates expressive facial animation using high-density surface electromyography (HD-sEMG) as a non-optical alternative to face capture. It measures 64 EMG channels from the forehead and side of the face to estimate 3D facial landmarks.https://arxiv.org/abs/2610.09033https://arxiv.org/abs/2610.09033Thu, 08 Oct 2026 04:00:00 +0000Open models for local useA Quad-State Safety Evaluation assesses open-weight LLMs on non-canonical inputs using the Adversarial Surface-Form Robustness Dataset (ASRD). It finds that emoji and invisible Unicode variations cause almost no comprehension failure, with specific harmful comhttps://arxiv.org/abs/2610.09000https://arxiv.org/abs/2610.09000Thu, 08 Oct 2026 04:00:00 +0000Phone and edge AIResearch investigates the fragility of on-device language model safety by localizing safety-critical parameters in LLaMA-2-7B-Chat. It finds highly non-uniform safety sensitivity, with MLP down_proj and o_proj identified as prominent safety-sensitive componenthttps://arxiv.org/abs/2610.10310https://arxiv.org/abs/2610.10310Thu, 08 Oct 2026 04:00:00 +0000Agent sandboxes (E2B and peers)RSIGym is introduced as an agent-native research environment for recursive self-improvement, based on Everything as a Service (EaaS). It exposes training, inference, rollout, evaluation, and sandbox execution through reusable services, supporting various improhttps://arxiv.org/abs/2610.09804https://arxiv.org/abs/2610.09804Thu, 08 Oct 2026 04:00:00 +0000Phone and edge AIBoT-GRPO (Bag-of-Tokens Group Relative Policy Optimization) is proposed to make process supervision efficient for reinforcement learning in LLMs. It extends GRPO to token-level reward models using a length-invariant aggregation, collecting token-level rewards https://arxiv.org/abs/2610.09212https://arxiv.org/abs/2610.09212Thu, 08 Oct 2026 04:00:00 +0000Runtimes and quantizationCurveTQ introduces rotation-free trellis quantization for LLM weights, using a curvature-weighted search. It incorporates layer Hessian information into the Viterbi branch metric, which existing quantizers compute but do not fully utilize.https://arxiv.org/abs/2610.09183https://arxiv.org/abs/2610.09183Thu, 08 Oct 2026 04:00:00 +0000Runtimes and quantizationQ-PACE is a new approach for dynamic precision allocation in quantization-aware training (QAT). It uses a second-order sensitivity model to predict loss increase and periodically re-computes coefficients to re-assign precision during training, showing consistehttps://arxiv.org/abs/2610.09063https://arxiv.org/abs/2610.09063Thu, 08 Oct 2026 04:00:00 +0000Open models for local useA comparative analysis evaluates Small Language Models (SLMs) for multi-label topic assignment via LLM distillation. It assesses generative versus discriminative student models across 1B, 4B, and 8B parameter scales for user-generated content.https://arxiv.org/abs/2610.09004https://arxiv.org/abs/2610.09004Thu, 08 Oct 2026 04:00:00 +0000Open models for local useResearch indicates that removing harmful information from open-weight models does not universally certify tamper resistance against fine-tuning attacks. Mutual information at release alone cannot guarantee slow recovery, as function-preserving reparameterizatihttps://arxiv.org/abs/2610.08963https://arxiv.org/abs/2610.08963Thu, 08 Oct 2026 04:00:00 +0000Runtimes and quantizationKL-Regularized Policy Optimization (KLPO) is proposed as a framework for asynchronous reinforcement learning in LLM agents. It anchors the KL regularizer at the sampler, offering a closed-form Gibbs solution and fitting log-ratio optimality by least squares onhttps://arxiv.org/abs/2610.08811https://arxiv.org/abs/2610.08811Thu, 08 Oct 2026 04:00:00 +0000Runtimes and quantizationKVFetch introduces temporal prefetching for KV cache compression, addressing the limitation of existing methods that only use content relevance. It highlights the need for sequential traversal in tasks like retrieval-augmented generation and code completion.https://github.com/ggml-org/llama.cpp/releases/tag/b11493https://github.com/ggml-org/llama.cpp/releases/tag/b11493Thu, 08 Oct 2026 05:41:13 +0000Runtimes and quantizationA release for llama.cpp adds grouped Mixture-of-Experts (MoE) XMX GEMM support for SYCL.https://github.com/ggml-org/llama.cpp/releases/tag/backup%2Fallozaur-ui-shell-polishhttps://github.com/ggml-org/llama.cpp/releases/tag/backup%2Fallozaur-ui-shell-polishTue, 06 Oct 2026 07:07:16 +0000Runtimes and quantizationllama.cpp UI received shell polish, including button looks, pointer cursor for pressable elements, font rendering smoothing, and root layout resolution.https://github.com/google-ai-edge/LiteRT-LM/releases/tag/v0.18.0https://github.com/google-ai-edge/LiteRT-LM/releases/tag/v0.18.0Tue, 06 Oct 2026 18:43:23 +0000Phone and edge AILiteRT-LM v0.18.0 shipped multimodal EmbeddingGemma 2 supporting text, vision, and audio embeddings with Matryoshka dimension truncation across multiple platforms.https://github.com/ollama/ollama/releases/tag/v0.40.0https://github.com/ollama/ollama/releases/tag/v0.40.0Tue, 06 Oct 2026 18:47:10 +0000Runtimes and quantizationOllama now runs models on MLX on Apple Silicon by default for supported architectures, including gemma4, qwen3.6, qwen3.5, Nimble, tev1, clef, clef-flash, and embeddinggemma-2.https://github.com/vllm-project/vllm/releases/tag/v0.31.1rc0https://github.com/vllm-project/vllm/releases/tag/v0.31.1rc0Tue, 06 Oct 2026 18:52:05 +0000Runtimes and quantizationvLLM exposes cached prompt tokens by cache tier.https://github.com/e2b-dev/E2B/releases/tag/e2b%402.53.1https://github.com/e2b-dev/E2B/releases/tag/e2b%402.53.1Tue, 06 Oct 2026 20:18:12 +0000Agent sandboxes (E2B and peers)E2B SDK patch changes include publishing the changes from a previous release that failed to publish to npm and PyPI.https://simonwillison.net/2026/Oct/6/le-chonk/https://simonwillison.net/2026/Oct/6/le-chonk/Tue, 06 Oct 2026 20:18:19 +0000Desk-side boxes (Mac Studio, DGX Spark, OEM)Mistral Large 4, a 1 trillion parameter, 49 billion active parameter model, was trained on a cluster of 3,800 NVIDIA Grace Blackwell GPUs, with a preview available via API.https://simonwillison.net/2026/Oct/6/hn-49983751/https://simonwillison.net/2026/Oct/6/hn-49983751/Tue, 06 Oct 2026 20:37:53 +0000Open models for local useEmbeddingGemma 2 is under the Apache 2.0 license, which is noted as beneficial for applications involving calculating and storing many embedding vectors.https://simonwillison.net/2026/Oct/7/openai-rogue-agents-wikimedia/https://simonwillison.net/2026/Oct/7/openai-rogue-agents-wikimedia/Wed, 07 Oct 2026 00:16:45 +0000Agent sandboxes (E2B and peers)Wikimedia Foundation discovered activity by "rogue" OpenAI agents on Wikimedia platforms, including edits to wikis, attempts to exploit a note-taking tool, and heavy traffic.https://github.com/google-ai-edge/gallery/releases/tag/1.0.20https://github.com/google-ai-edge/gallery/releases/tag/1.0.20Wed, 07 Oct 2026 03:36:10 +0000Phone and edge AIGoogle AI Edge Gallery 1.0.20 features EmbeddingGemma 2, supporting multimodal semantic search 100% on device, including Instant Media Search and Video Moment Finder.https://arxiv.org/abs/2610.04687https://arxiv.org/abs/2610.04687Wed, 07 Oct 2026 04:00:00 +0000Agent sandboxes (E2B and peers)A study on LLM-based question answering over imperfect tables found that reordering rows changes error discovery, and providing verified error locations alone is insufficient for accurate QA.https://arxiv.org/abs/2609.38480https://arxiv.org/abs/2609.38480Wed, 07 Oct 2026 04:00:00 +0000Agent sandboxes (E2B and peers)KlinikeBench is a benchmark of 333 clinician-authored tasks in isolated sandbox environments with virtual patients and clinical tools, evaluating language models beyond diagnostic accuracy.https://arxiv.org/abs/2610.07125https://arxiv.org/abs/2610.07125Wed, 07 Oct 2026 04:00:00 +0000Open models for local usePerturbed Embedding Vector (PEV) is a jailbreaking technique for open-weight LLMs that adds independent Gaussian noise to embedding vector representations of prompts.https://arxiv.org/abs/2610.06903https://arxiv.org/abs/2610.06903Wed, 07 Oct 2026 04:00:00 +0000Open models for local useRefusal steering in transformer models concentrates in sparse component mechanisms (28-48% of upstream components) and within approximately 50% of residual stream dimensions.https://arxiv.org/abs/2610.08680https://arxiv.org/abs/2610.08680Wed, 07 Oct 2026 04:00:00 +0000Open models for local useA systematic study of small language models on abstract reasoning tasks found that substantial in-distribution accuracy is attainable, but skill acquisition is sensitive to model family and task formulation.https://arxiv.org/abs/2610.08669https://arxiv.org/abs/2610.08669Wed, 07 Oct 2026 04:00:00 +0000Phone and edge AIMemFLoRA is a low-rank CNN adapter for on-device learning, designed with a memory-first principle to ensure trainable backward computations do not depend on full-width layer inputs.https://arxiv.org/abs/2610.07553https://arxiv.org/abs/2610.07553Wed, 07 Oct 2026 04:00:00 +0000Open models for local useGRADE (GRadient-Aligned Data-centric rEcipe) is a framework for LoRA fine-tuning that uses a state-aware selector and a self-calibrating step-level gate.https://arxiv.org/abs/2610.07466https://arxiv.org/abs/2610.07466Wed, 07 Oct 2026 04:00:00 +0000Phone and edge AISemARC is introduced for efficient multimodal inference, coupling a Sequential Modality Aggregator (SeMA) with an Adaptive Runtime Controller (ARC) to select modalities before encoding.https://arxiv.org/abs/2610.07444https://arxiv.org/abs/2610.07444Wed, 07 Oct 2026 04:00:00 +0000Phone and edge AIA study on small GUI grounding models found that an auxiliary loss, an additive embedding, and the action type written into the prompt improved performance.https://arxiv.org/abs/2610.07348https://arxiv.org/abs/2610.07348Wed, 07 Oct 2026 04:00:00 +0000Phone and edge AIStepped MoE is a unified framework combining elastic structures with sparsely gated architectures for models that adapt to deployment constraints and task requirements.https://arxiv.org/abs/2610.07197https://arxiv.org/abs/2610.07197Wed, 07 Oct 2026 04:00:00 +0000Runtimes and quantizationQuantized Sufficient Statistics (QSS) is a method for exact unlearning, separating a small frozen schema from mutable, sum-decomposable content for exact subtraction upon deletion.https://arxiv.org/abs/2610.07121https://arxiv.org/abs/2610.07121Wed, 07 Oct 2026 04:00:00 +0000Runtimes and quantizationSoloQ is a calibration-free quantization framework for diffusion large language models (dLLMs) that maps weights and activations into a normalized rotated basis.https://arxiv.org/abs/2610.07086https://arxiv.org/abs/2610.07086Wed, 07 Oct 2026 04:00:00 +0000Runtimes and quantizationSchemaFill is a framework for efficient LLM tool calling using slot-parallel speculative decoding, addressing latency for multiple calls or argument fields.