Mayank Singh retweetledi
Mayank Singh
8.9K posts
Mayank Singh
@geekmarcus
PhD student at @UArizona | Interested in #NLProc, History, #Astronomy, and more Part-time @GameOfThrones fan account
Katılım Haziran 2010
192 Takip Edilen404 Takipçiler
Our paper "Latent Agents" was accepted to #ACL2026 Main!
We distill multi-agent debate into a single LLM, matching debate performance at a fraction of the cost. We also show that internalized agents are discoverable and controllable.
Huge thanks to @amuuueller and Dokyun Lee!
English
This was accepted to ACL Main Conference! Thanks and congrats to my collaborators! 🎉🎊🎇 #acl2026
Mayank Singh@geekmarcus
🚨 Preprint alert: Grammar Search for Multi-Agent Systems 🧵 Instead of asking LLMs to invent multi-agent code from scratch, we define a grammar of simple, composable components and search over valid system compositions. arxiv.org/abs/2512.14079
English
Mayank Singh retweetledi
Mayank Singh retweetledi
@subhavsamarth जब जब शहर में गंदगी दिखता है तो शहर वाला कहता है साफ कर ले नहीं तो कोई पगला इंटर्न आ जाएगा।
हिन्दी
Mayank Singh retweetledi
Finally finished!
If you're interested in an overview of recent methods in reinforcement learning for reasoning LLMs, check out this blog post: aweers.de/blog/2026/rl-f…
It summarizes ten methods, tries to highlight differences and trends, and has a collection of open problems
English
Mayank Singh retweetledi
Top 50 fantasy series according to an internet poll from 2001. I love looking at lists like this and figuring out why each series fell off and why some stayed relevant.
English
Mayank Singh retweetledi
Claude Code: "You've hit your limit · resets 7pm"
Me from 5-6.59pm
English
Mayank Singh retweetledi
One of the cool pieces of news to emerge during those last, dark weeks is that nearly 30 inscriptions in Tamil Brahmi were found in the Egyptian Valley of Kings. This discovery means that apperently Indian travellers, likely merchants, had been coming to Egypt in the ancient era.
English
Mayank Singh retweetledi
Mayank Singh retweetledi
Mayank Singh retweetledi
斯坦福大学竟然开了一门 Vibe Coding 课程
课程编号 CS146S，全名叫《The Modern Software Developer》，持续 10 周，依次讲解了提示词工程、Agent 架构、MCP、上下文工程、安全攻防、Code Review、自动做 App 和上线运维。
themodernsoftware.dev
中文
Mayank Singh retweetledi
The hammer vs the anvil... such a cool shot! 👌
aki is mourning@alphalicent
maekar coming back for round two to slime dunk before baelor stopped him is kinda tickling me
English
Mayank Singh retweetledi
Mayank Singh retweetledi
📊 Across math, QA, and logical reasoning benchmarks, this approach matches or improves over prior automatic prompt optimization methods while using substantially fewer optimization-phase tokens.
📄 arxiv.org/abs/2602.00997
English
We propose a top-down alternative.
🧠Key idea: focus on the global failure landscape rather than reacting to individual examples or mini-batches. Aggregate errors, categorize them, and update the prompt to target recurring failure modes. This avoids overfitting to local failures.
English
🧪 New preprint: Error Taxonomy-Guided Prompt Optimization 🧵
Many feedback-driven automatic prompt optimization methods iteratively update prompts using feedback from individual examples or small mini-batches, which can overfit to local failures and miss the global picture.
English