AITOPAI热报

Latest

Every published item, in order · 1438 items
BlogsNewsPostsPapersPapers →
Sep 25
Sep 24
PapersBuildersMulti-sourceSelected73°23:14
NVIDIA's Skill2Env Turns Agent Skills into RL Environments

Rubric-augmented rewards trade raw benchmark score for trajectories a judge rates far higher, a hint that outcome-only RL over-optimises for narrow metrics while behavioural rubrics shape more genuinely skilled agents. Teams training terminal agents can reuse the open pipeline to convert existing Agent Skills corpora into RL environments instead of hand-crafting tasks. The US$90,000 generation cost also sets a concrete budget baseline for anyone planning similar corpus builds.

#NVIDIA#Skill2Env#RL environments#Qwen3.8-27B
Event
DAIR.AI@dair_ai9 sourcesSource
PapersBuildersMulti-sourceSelected09:35
SkillGym Turns Human Agent Skills into Training Environments

The key finding is that skills can be internalized as model weights rather than injected at inference time, since the fine-tuned agent beats skill-assisted base models even without the skills provided. Agent engineers and fine-tuning practitioners can reuse the released environments and trajectories for SFT or RL with outcome-based rewards. The result also shows a 35B open Qwen-based model outscoring much larger closed models on SkillsBench, a useful cost comparison for teams building coding and tool-use agents.

#SkillGym#Qwen3.5#agents#fine-tuning
Event
arXiv: DeepSeek@Zhilong Ge et al. (9)6 sourcesSource
Sep 23
ModelsUS vs ChinaMulti-sourceSelected19:27
Step 5 Preview tested on two real coding tasks

Benchmark scores can hide whether a model survives long, multi-step builds, so the value here is watching Step 5 Preview handle mechanical synchronisation and real tooling rather than canned puzzles. Developers choosing between Step 5, GLM-5.3, Kimi K3 and Claude Opus 5 get a practical cost-versus-reliability comparison for agentic coding workflows. The stated price gap makes it a concrete candidate for teams in Southeast Asia and beyond cutting per-task inference spend.

#Step 5 Preview#StepFun#coding assistants#GLM-5.3
Event
掘金本周最热@猫猫头啊6 sourcesSource
PapersBuildersSelected10:28
FeatLens: Feature-Guided Code Graph Retrieval for Repo-Level Code Generation

The key shift is replacing expensive persistent repository graphs and LLM-based exploration with a deterministic, feature-indexed retrieval pipeline, making repo-level context cheap enough for routine use. Teams building coding assistants or agent frameworks can adopt the approach to cut context costs without sacrificing dependency recall. For cost-sensitive deployments, pairing it with DeepSeek-V3.2 gives near-identical DIR@1 to GPT-5-mini at lower inference cost.

#FeatLens#DeepSeek-V3.2#GPT-5-mini#RAG
arXiv: DeepSeek@Xutian Li et al. (9)Source
PapersBuildersSelected09:09
Post-Training Delivery Benchmark Tests LLM Agents as Forward-Deployed Engineers

The benchmark reframes agent evaluation around deliverable outcomes and payment-gated acceptance, a shift from metric-chasing benchmarks that miss corrupted training runs. Teams building or buying post-training-as-a-service can use the acceptance gate and corruption detector to screen automated fine-tuning pipelines. The human FDE baseline gives a concrete performance reference against all four frontier agents.

#PTaaS#benchmark#fine-tuning#agents
arXiv: DeepSeek@Weihang Ding, Junfei ZhanSource
PapersBuildersSelected73°04:40
ReFigBench shows harness choice changes coding agent results

The same model can score differently depending on the harness around it, so benchmark numbers reflect the full stack rather than the model alone. Teams evaluating coding agents should rerun comparisons under their own harness before drawing conclusions. ReFigBench adds a new evaluation axis, and details are on arXiv (2609.18844) alongside a chat interface from dair.ai.

#ReFigBench#GPT-5.5#Claude Code#Codex
elvis@omarsar0Source
PapersBuildersSelected73°04:04
Jev tested across three Zilliz open-source RAG projects

The benchmarks show where cheap small models can replace larger ones in RAG pipelines: stopping decisions and relationship filtering see large latency and cost savings with matched recall. Engineers building retrieval systems can use the published code to pick where to route between Jev, GPT-5-mini and Voyage rerank-3. The evaluation includes Chinese-language queries, useful for teams serving bilingual users.

#Jev#Zilliz#RAG#reranking
Milvus@milvusioSource
ProductsBuildersMulti-sourceSelected73°03:58
DocJev OSS Library Classifies and Splits Document Packets

The non-obvious angle is that DocJev combines small-model prediction with pluggable OCR backends, letting teams trade latency for accuracy per document type rather than defaulting to a heavy VLM pipeline. Developers building RAG ingestion or document-processing workflows can wire it into existing pipelines via the GitHub repo, with the default path fully free. Compared with gpt-5.6-luna, it targets teams where throughput on large document packets matters more than marginal accuracy gains.

#DocJev#LlamaIndex#LlamaParse#RAG
Event
Jerry Liu@jerryjliu03 sourcesSource