Archive
推論 runtime、量化、 serving、KV cache、成本與效能。
推論系統
Liquid AI has released three DSpark draft models with roughly 300 million parameters, enabling LFM2.5 to verify multiple candidate tokens at once through speculative decoding. Tests with batch size 1 and greedy decoding showed speedups of up to 3.18×, but SGLang support remains in an open PR, leaving deployment maturity yet to be validated.
最新模型
Open-MOPD finds that the main bottleneck in multi-teacher online policy distillation is not conflicting teacher gradients, but a severe imbalance in the effective number of token updates allocated across domains. By combining token-share balancing, gap-aware allocation, and student reward refresh, it recovers 83.4% of the available capability gains in a single 3B model.
科學機器學習
Monroe combines sparse graph attention, stereochemical edges, and 1,152 pretraining objectives, then uses TabPFN for in-context prediction on frozen molecular embeddings. It ranks in the statistically best-performing group across all 28 Polaris tasks, although its “100% win rate” includes ties in which it is not significantly different from the best method.
最新模型
Researchers repeatedly update hidden states through the same set of Transformer layers before generating each token, comparing recurrent and conventional models using identical data and LoRA configurations. Gains are concentrated in parallel calls and cross-call dependencies, while API-Bank, which primarily involves single API calls, shows no consistent advantage.
Cartesia 推出支援逾40種語言的 Sonic-3.6，並在 Artificial Analysis 的供應商語音及受控語音榜取得首位。產品頁宣稱首段音訊延遲低於90毫秒，但公開資料尚不足以分離模型計算、網路及播放緩衝成本。
代理工具
FACET binds instructions, containers, reference solutions, and verifiers to the same execution state, synthesizing 6,078 verified tasks from more than 71,000 agent skills. Fine-tuned on just 1,200 successful trajectories, Qwen3.5-9B improved its Terminal-Bench 2.1 score from 27.34 to 35.58.
代理工具
Researchers applied control-flow rewrites, identifier renaming, and dead-code insertion to the same repair tasks, reducing solve rates by as much as 6.7 percentage points. Model robustness changed across mini-SWE agent, OpenCode, and datasets, meaning a single model leaderboard cannot predict deployment performance.
AI 基礎設施
Rather than moving to a new process node, the CS-4 increases the WSE-3’s clock speed, power delivery, and I/O, then combines three wafer-scale processors in a Nexus rack. Cerebras claims inference speeds of up to 30 times those of GPU-based systems, but its PFLOPS and throughput-per-watt figures still lack comparable independent end-to-end testing.
推論系統
A preregistered study attempted to train expert locality directly into an MoE router, but found a clear trade-off between cache performance and model quality. Combining training-based and training-free rerouting reduced misses further, yet still did not demonstrate a proportional end-to-end speedup.
代理工具
SkillGate finds that when long-horizon agents are trained with outcome rewards, the small number of tokens used to select a skill receive almost no effective credit. The new method separates two gradient channels within the same GRPO update and outperforms the outcome-reward baseline across five agent benchmarks.
推論系統
Google DeepMind proposes mixing the deep residual stream from one token into shallow layers for subsequent tokens at inference time, creating cross-token recurrence in off-the-shelf Gemma 3 models. The method adds no model weights, but turns parallel prefill into sequential processing, and its real-world latency has yet to be fully measured.
硬體
The 28 nm ETHEREAL chip separates irregular event-graph access from regular spline convolution and supports per-layer 4/8-bit precision. DAGr-GNN measurements reached 25.6 microseconds and 1.6 microjoules per event, although the final three detection-head layers were still evaluated off-chip.