Archive
GitHub 熱門專案、開源工具、框架、runtime 與重要 release。
AI 安全
A new study moves evaluation upstream to the initial-contact stage, before patients have clearly described their problems. Adding a “clarify first, advise later” system instruction improved response sequencing and handoff summaries, but still did not reliably elicit decisive medical history.
推論系統
Qualcomm AI Research adds per-layer fast weights to a frozen Transformer, absorbing context in 512-token segments so queries no longer need to read the source text. In 128K experiments, Qwen2.5-0.5B maintained fixed memory usage and outperformed ICL and RAG, though the evidence is currently limited to three synthetic retrieval tasks.
最新模型
TokEval incorporates numeric place value, program AST boundaries, UTF-8 integrity, and cross-lingual cost into tokenizer evaluation. Controlled experiments show that some metrics can predict specific downstream capabilities, although the results are currently limited to a single 1.27B-parameter architecture.
代理工具
Microsoft rewrote Agent Lightning to capture complete interactions from existing agent frameworks through a model API proxy layer, eliminating the need to move tool loops into the training engine. Using 6,000 training examples, the team reports that Qwen3.5-9B improved from 41.8% to 56.4% on SWE-bench Verified.
硬體
The open-source rl-triton library rewrites credit-assignment recurrences such as GAE, V-Trace, and Retrace as Triton associative scans, reducing sequential dependency depth from O(T) to O(log T). In tests with 4,096 environments and 128 steps, it was 1.6× to 5.70× faster than an already vectorized torch.compile baseline, although end-to-end PPO speedup for large policies was only about 1.02×.
推論系統
Rather than removing token connections, TileMix selects either the FP16 or INT8 Tensor Core path for each attention-score tile within a single fused kernel. It reaches 31.80K token/s on a 4K LLaMA 3.2 3B benchmark, though the results are currently limited to A100, static routing, and the prefill stage.
代理工具
Salesforce reevaluated two text-based memory methods and found that adding a self-improvement loop increased run-to-run variance in 71% of configurations. Once the fixed task order—which implicitly formed an easy-to-hard curriculum—was shuffled, ReasoningBank not only failed to improve the agent but reduced its success rate.
代理工具
Benchmark-as-Teacher turns stage-level evaluation scores directly into the curriculum for the next training round, while using isolated synthetic data to prevent test answers from leaking into updates. The paper reports that its 9B policy substantially outperforms standard GRPO, but the full training dataset has not yet been released, and cross-system rankings mix different execution frameworks.
FabriMAE 從 VLA 模型內部的視覺注意力熵產生可靠度分數，不需另訓練驗證器或重複執行機器人軌跡。PI0.5 在 LIBERO-Plus 的成功率增加 1.1 個百分點，但目前證據仍限於模擬環境，論文連結的程式庫亦尚未公開存取。
代理工具
Rather than requiring a language model to directly generate long-horizon action sequences, PDDLCoder first constructs a PDDL world model that can be verified by a symbolic planner. DeepSeek V4 Flash solved 95 of 106 problems across four held-out domains, but results for different base models ranged from 89.6% to zero, showing that the framework’s effectiveness depends heavily on a model’s ability to follow formatting requirements and use tool feedback.
推論系統
Rather than expanding ultra-low-bit weights back into FP16, FluxBin uses binary patterns to directly index precomputed partial sums. The paper reports up to a 5.92× speedup and a 10.19× improvement in energy efficiency on a single A100, but its public code link remains inaccessible.
代理工具
Fujitsu ported its fine-tuning-free Qwen3.5-27B setup, eight-stage repair workflow, and cross-agent test selector to Java unchanged. It solved 374 of 500 tasks on the Python SWE-bench Verified benchmark, but achieved only a 32.03% success rate on Java, showing that agent-framework portability does not imply equal language proficiency.