#instruction following

11 stories · 0 in 7 days · 0 in 30 days

AI safety 3 Gary Marcus 2 Qwen3 2 DeepSeek-V3.2 1 Gacha Decoding 1 Gemini-2.5-Pro 1 MLLMs 1 SciMIF 1

Stories, products and papers tagged "instruction following", newest first.

Oct 12

Industry Policy16:06

Chinese contentGary Marcus：生成式 AI 无法可靠遵循简单通用指令

Gary Marcus 转发并吐槽了一件事：给 AI 写条“别伤害人”的规则它都执行不稳，他说这个被低估的风险早晚会出事。

#Gary Marcus #generative AI #AI safety #instruction following

Gary Marcus@GaryMarcusSource

Oct 2

Papers Builders10:53

Chinese contentGacha解码：通过指令跟随实现多样化生成

Gacha解码方法让语言模型通过'掷骰子'实现多样化生成，模型能力越强效果越好。

#Gacha Decoding #language models #instruction following #Vendi

arXiv cs.AI@Scott Geng et al. (6)Source

Sep 29

Papers BuildersSelected18:34

ScopeIF Framework for LLM Instruction Following

ScopeIF addresses a critical gap in LLM training by providing dense supervision for complex constraints that existing binary reward models miss. This approach significantly improves performance on multi-faceted instructions without compromising general capabilities. Developers working with complex task automation can implement this framework to enhance their models' precision, particularly when dealing with segmented constraints across different response parts.

#ScopeIF #Qwen3 #Gemini-2.5-Pro #DeepSeek-V3.2

arXiv: DeepSeek@Bosi Wen et al. (6)Source

Sep 1

PapersSelected09:20

Reasoning Models Show Asymmetric Disclosure

The findings reveal vulnerabilities in AI oversight mechanisms that developers and policymakers should consider when implementing reasoning model monitoring systems.

#Qwen3 #reasoning models #chain-of-thought #AI safety

arXiv: DeepSeek@Zimo Shi et al. (3)Source

Aug 27

Chinese contentSciMIF：理解科学领域的多模态指令遵循

SciMIF基准为评估MLLMs在科学领域的指令遵循能力提供了新工具，揭示了不同学科间的性能差异，对科学研究和MLLMs发展具有重要意义。

#SciMIF #multimodal #MLLMs #instruction following

arXiv cs.LG@Ye Shen et al. (7)Source

Aug 18

Chinese contentGary Marcus：AI不遵循指令被忽视，LLM架构威胁安全

Gary Marcus直接点出AI不听指令比幻觉更危险，还引用了越具体越不听话的现象。想了解AI安全风险，看这条。

#Gary Marcus #AI safety #instruction following #hallucinations

Gary Marcus@GaryMarcusSource

Jul 30

Chinese contentCohere 发布 rerank-2.5-lite：支持指令跟随的自然语言重排序，每百万 token 仅 $0.02

Cohere 出了个新重排序模型，能照着你的自然语言描述来排序，价格还超低，每百万 token 才两分钱。

#rerank-2.5-lite #Cohere #OpenRouter #instruction following

OpenRouter@OpenRouterAISource

Chinese contentrerank-2.5：指令跟随重排序模型，检索更准，每百万token 0.05美元

OpenRouter 出了个新重排序模型 rerank-2.5，能听懂指令来排序，检索更准，价格还特别便宜，每百万token才5分钱。

#rerank-2.5 #OpenRouter #reranking #RAG

OpenRouter@OpenRouterAISource

Jul 1

Chinese content自生成问答训练语言模型的隐藏脆弱性

这篇论文告诉你，用模型自己问自己生成的问答数据来训练模型，可能会引入隐蔽的偏见和指令注入风险。他们用一个简单方法就能把问题率从88%压到13%，值得做数据合成的同学看看。

#fine-tuning #distillation #instruction following #data quality

arXiv cs.AI@Ekaterina Alimaskina et al. (6)Source

Jun 10

PapersSelected09:33

Chinese contentTREAD：用VLM重标注机器人数据，提升指令跟随鲁棒性

做机器人数据增强和指令跟随的团队，TREAD用VLM低成本提升数据集质量，直接增强策略泛化，值得在LIBERO等基准上试试。

#robot learning #data augmentation #VLM #instruction following

arXiv cs.LG@Artur Kuramshin et al. (4)Source

Jun 4

Chinese contentVentureTwins 评测：该模型对长短提示词都遵循得很好

对提示词长度和复杂度不敏感，做内容生成或创意设计的团队可以试试，能省去反复调优提示词的麻烦。

#model evaluation #instruction following #prompt engineering #content generation

Justine Moore@venturetwinsSource