Archive
GitHub 熱門專案、開源工具、框架、runtime 與重要 release。
最新模型
A new study turns NoThink, Short, and Long into the model-generated first token, allowing it to allocate a reasoning budget for each problem without a separate router. On MATH-500, the method maintains accuracy close to the base model while cutting tokens by 76% on the easier GSM8K benchmark.
代理工具
A study spanning nine models and three long-horizon benchmarks finds that skills induced from complete tasks usually make agents perform worse, while those induced from subtasks deliver average gains. Natural-language process notes are also more robust than executable Python skills, though results remain highly dependent on the model and benchmark.
代理工具
MemTrapBench does not test whether memories can be retrieved. Instead, it examines whether correct, relevant past experiences can lock a model into the wrong strategy or belief. Across two models, all five memory methods underperformed the no-memory baseline, while a simple prompt-based guardrail recovered only part of the loss.
代理工具
SAPO uses two causal boundaries in a generated sequence to read out state and action values simultaneously, eliminating the need for a separate critic or multiple rollouts of the same task. Qwen2.5-7B outperforms PPO and GRPO on the overall metrics for both ALFWorld and WebShop, although no public implementation is yet available for verification.
最新模型
DECOWAM adds base, arm, and camera egomotion interfaces to the frozen 6.02-billion-parameter FastWAM, using only the current observation at deployment to predict 48 action steps. Its overall completion rate in physical robot tests is only slightly higher than FastWAM’s, but its robustness success rate under base disturbances rises from 12.7% to 30.4%.
推論系統
Axon expresses model architectures as functional specifications with symbolic tensor shapes, then compiles them to PyTorch, Triton, JAX, MLX, and native vLLM. Across 467 inference tests, it accelerated most generative models, though some BF16, encoder, and sequence-to-sequence models still showed accuracy or performance regressions.
代理工具
Four modules sharing Qwen2.5-0.5B and the same LoRA learned radically different composition strategies solely because of different attention masks. Limiting each module to its own evidence fragment made it easier to develop intermediate-value communication that transferred across examples, but the experiment still missed its preregistered overall success threshold by 0.0012.
AI 安全
New research finds that single-run greedy decoding and unstable thresholds can make a completely untrained model appear to gain or lose capabilities. After switching to exact per-item tests, three forms of self-training only consistently strengthened existing capabilities, with no evidence that they acquired new capabilities never demonstrated by the base model.
代理工具
CentaurBench separately evaluates models’ automation and assistance capabilities across seven work tasks, finding a correlation of only about 0.48 between the two rankings. Unassisted GPT-3.5-Turbo outperformed every guided version on three tasks, showing that adding a powerful model does not necessarily improve an agent workflow.
代理工具
TrueFoundry has released TrueForge under the MIT License, combining agent loops, MCP tools, on-demand sandboxes, subagents, and approval workflows into a self-hostable execution framework. In its Enterprise-Bench report, the company claims savings of about 30% using the same model and about 75% after switching to GLM-5.2, although the comparison covers only 14 tasks.
代理工具
Salesforce has launched Slack Code, allowing teams to create project-specific Code Channels directly from conversations and collaboratively monitor, revise, and approve agent-generated code. Initial integrations include Claude, Devin, GitHub Copilot, ChatGPT, and Vercel agents, but details about security isolation, permission granularity, and APIs have yet to be fully disclosed.
代理工具
AFANet converts agent conversations into graphs with temporal and same-agent edges, then uses a two-layer GCN to identify faulty agents and error types. It outperforms several LLM baselines on AEGIS-Bench, but its agent–error pair F1 remains low across datasets.