推論系統

推論 runtime、量化、 serving、KV cache、成本與效能。

Muse Code Loads Personal Claude and Codex Instructions by Default, Raising Data-Boundary Concerns Over Cross-Provider Compatibility

Independent testing found that Meta’s Muse Code looks for `CLAUDE.md` and `AGENTS.md` files in the user’s home directory at startup and incorporates their contents into the context sent to the model endpoint. This compatibility design does not amount to arbitrary file exfiltration, but it could send personal rules, internal paths, or sensitive prompts originally intended only for another provider to Meta.

2026-08-10

AgentPatch Repairs Merged Multimodal Agents Without Training, Raising the Six-Benchmark Average From 54.5 to 56.6

AgentPatch first identifies weaknesses introduced by model merging, then restores only the expert models’ unique residuals and critical FFN neurons. It produces a single static checkpoint with no routing or inference-time agents, but the complete evaluation adapters and model weights have yet to be released.

2026-08-10

AI 安全

OpenAI Pauses Some Astra Development: Initial Tests Cannot Rule Out That the Model Has Reached “Critical” Cyberattack Capability

OpenAI says advances in Astra’s agentic coding and cyber offense and defense capabilities mean it cannot rule out that the model has crossed the highest-risk threshold. The company has isolated testing environments, restricted tool and network access, and deployed monitors to inspect agents’ reasoning traces and high-risk actions.

2026-08-10

代理工具

DiDPO Breaks Code Diffs Into Credit Units, Raising Qwen2.5-Coder-7B’s Average on Key Benchmarks to 48.4%

Rather than distributing the outcome of an entire agent run evenly across all tokens, DiDPO compares similar sub-diffs across multiple trajectories. The public verl-code repository includes training entry points and data recipes, but no author-released model checkpoint yet, and the experiments do not cover full repository-level software repair.

2026-08-10

推論系統

Ling-3.0-flash INT4 Hits 38.7 tokens/s in Single-System Testing; Using Mainline vLLM May Silently Produce Incorrect Output

A community deployment recipe running on a single DGX Spark uses CUDA Graph and a one-layer MTP speculative decoder to increase Ling-3.0-flash INT4 generation speed from 20.8 to 38.7 tokens/s. More importantly, it finds that mainline vLLM does not yet support the V3 architecture: forcing the model through an older attention path may not raise an error but can produce plausible-looking yet incorrect content.

2026-08-10

時間序列與檢索增強

TS-RAG Fuses Historical Sequences via Reference Tokens, Reducing Average MSE to 0.310 Across Six Forecasting Datasets

Baidu researchers have adapted RAG from text generation to multivariate time-series forecasting, using a TCN to retrieve similar segments and cross-attention to fuse them. A channel-dependent variant outperformed the listed baselines across six datasets, but the code has not been released, and maintainers have warned that the conventional benchmarks used may no longer be sufficiently discriminative.

2026-08-09

DASH Dynamically Allocates Distillation Weights Based on Reasoning Divergence, Raising Qwen3-1.7B’s Math Average by 3.2 Points

Rather than treating teacher–student differences at every token equally, DASH adjusts the scope of supervision based on how divergence changes across an entire reasoning sequence. All three Qwen3 sizes outperform OPSD under the same conditions, but selecting the best checkpoint by average test-benchmark score may overestimate generalization gains.

2026-08-09

推論系統

PaDoc Parses Documents in Parallel with Layout Branches, Boosting Throughput by Up to 118% on a Single A800

PaDoc decodes the layout sequence and each region’s content in parallel while sharing the page prefix, avoiding serialization of the entire document into a single path. Its 2.1B model maintains a 94.24 Overall score on OmniDocBench, though the performance figures come from a single GPU under specific scheduling settings.

2026-08-07

硬體

esp32-ai Fits a 28.9M-Parameter Model on the ESP32-S3, Reaching 9.88 Tokens/s Offline

The open-source esp32-ai project uses 4-bit quantization and tiered memory allocation to keep roughly 25 million embedding parameters in flash, reading only the rows needed for each token. The model can generate text offline on a microcontroller costing under $10, though its capabilities remain limited to short stories and single-domain question answering.

2026-07-26

推論系統

PhyAI Unifies Inference for VLAs and World-Action Models, Delivering 1.40× to 4.65× Speedups over Official Implementations

PhyAI uses model adapters to preserve each architecture’s conditioning, solver, and action semantics while sharing graph execution, kernels, caching, and multi-GPU serving. Public benchmarks cover π0, π0.5, GR00T N1.7, Cosmos3, and MiniCPM-Robot, although some comparisons do not use exactly the same numerical precision.

2026-08-06

代理工具

TutorMoments Replays Real Tutoring Decisions, Finds Seven LLMs Default to Over-Scaffolding

Ai2 built a replayable evaluation from 462 real math tutoring transcripts, asking models to choose between providing scaffolding and requiring deeper student reasoning. Explicitly prompting for this tradeoff improves the performance of all seven models, but simulated students and model-based grading still cannot demonstrate real learning outcomes.

2026-08-09

Liquid AI Releases LFM2.5-2.6B Weights, Bringing 128K Context and Tool Calling to a 2.69B On-Device Model

LFM2.5-2.6B uses a hybrid architecture combining short convolutions with GQA to bring agentic reasoning, tool use, and long-context capabilities to smartphones and personal computers. Liquid AI reports decoding speeds of 220 tokens/s on an M5 Max, though the performance and agent benchmarks are based primarily on the vendor’s own evaluations.

2026-08-08