LatentSkill — A New Method That Turns Agents' Textual Skills into Weights
A joint team from Shanghai Jiao Tong University, Sun Yat-sen University, Shanghai Innovation Institute, and OPPO Research Institute has published "LatentSkill," a method that converts the textual skills LLM agents carry in their prompts (in-context textual skills) into a form that can be **directly embedded into the model's weights (in-weight latent skills)**.
Traditionally, the only way to teach an agent a new skill was to append long text to the prompt, which crowded the context window and inflated inference costs. LatentSkill solves this problem by distilling textual skills into plug-and-play weight parameters.
In agent development, "adding skills" and "saving context" have long been a fundamental trade-off. If this approach works at a practical level, it could contribute significantly to making agents lighter and faster.
RAG Rewriting Improvements — "Answer Leakage" Was the Real Culprit
Another paper that has been drawing attention on Hugging Face has identified a fundamental problem in how query rewriting in RAG (Retrieval-Augmented Generation) is evaluated.
Many rewriting methods have claimed to "improve F1 scores," but this paper, through rigorous intervention experiments, showed that **the majority of the score gains come from gold answer leakage**. Specifically, removing gold answers from the texts caused F1 scores to plummet by 28 to 64 points. Conversely, injecting gold answers into texts that originally contained none raised F1 in 10 out of 12 cases.
The study also found that the conventional single-probe leakage check is extremely fragile. An audit using five sentinels (monitoring probes) confirmed cases where F1 improvements previously judged "non-leaky" actually flipped negative (+4.12 → -7.81).
This paper does not propose a new rewriting method; its goal is **making evaluation more rigorous**. The authors have open-sourced their intervention runner and sentinel panel, laying the groundwork for uniformly auditing claims made by future rewriting methods.
Is Open Source LLM "Good Enough" Already? — A Heated Debate on Reddit
On Reddit's LocalLLaMA community, the question "Have open source LLMs reached a 'good enough' level, satisfying 95% of requirements?" has sparked a lively debate.
The poster asks where the added value of the remaining 5% comes from, examining cost-effectiveness from several angles: (a) improved answer quality, (b) cleanliness of automation loops, (c) the risk of criticism that "it's not the best," (d) further productivity gains, and (e) risk management.
What makes this discussion interesting is that while the community is genuinely feeling the performance gains of open source models, it also sheds light on the organizational and psychological biases that keep teams choosing commercial APIs (OpenAI, Anthropic, Google, etc.). The effort to frame the question through the practical lens of "cost versus effectiveness" offers a perspective that is also useful for LLM adoption decisions inside enterprises.
A Discussion Seeking Practical Tips on Cutting AI Token Costs
On Hacker News, a post complaining that "LLM token fees hurt" is also gaining traction. It asks for **ideas to reduce costs in real-world applications** beyond simply switching to cheaper models.
As LLM usage moves into serious production, inference cost optimization is becoming a shared challenge for development teams. It's a thread where concrete techniques — prompt caching, semantic caching, reducing output tokens, optimizing batch processing, and more — are expected to be shared.
The Cheapest Setup to Run a 120B Dense LLM at 10+ tok/sec
Also on Reddit LocalLLaMA, someone asked about the cheapest hardware configuration for running a 120B-parameter dense model (Q5–Q6 quantization, 64k context) at 10 tokens per second or more.
The poster lays out three patterns: CPUOnly inference (8-channel RAM workstation class), GPUOnly inference (requiring 120GB VRAM), and hybrid inference, and is specifically asking the community about the optimal hybrid configuration, which remains unclear. For scaling up local LLMs, a clear picture of hardware options and cost-performance is valuable information for many practitioners.
---