Samuel Jayden Shi

Researcher interested in AI Environment Protection and Social Benefits (e.g., Efficient AI, Deep Learning Theory, Optimization and etc.)

[SuSu_酥酥👅 @NFT_Chen]

💥DeepSeek-V4-Flash + StateM 仅用$15.2就干翻Fable 5（$552.67）！节省约36倍成本！

拆解核心信息：
🔹不是单纯模型对拼：Agent能力增长不能只靠模型scaling。训练几个月期间任务表现不该停滞，他们提出Harness Scaling新路径——StateM（基于状态的可配置显式控制层）。

🔹现有Agent三大痛点：
1️⃣Plan了不执行/验证不充分（Codex、Claude Code等软checklist，无强制性，人难介入）
2️⃣缺乏既往流程经验，部分plan和自验证失效
3️⃣Context管理不可细化控制

🔹StateM解决方案：YAML配置 + CLI运行，状态机式显式控制，中途可阻断失败���支持人机可读可干预。

🔹验证结果（Terminal-Bench 2.1）：用GPT-5.5 xhigh + Codex开发可泛化runbook（作者定方向，Agent自主迭代），迁移到DeepSeek-V4-Flash成本<$38。最终DeepSeek-V4-Flash + StateM仅$15.2就击败Fable 5（$552.67）。同runbook让GPT-5.5/5.6也起飞，反超更贵的Ultra版本。

🔹论文：https://t.co/rbbEbuFE8p（还有配套verification scaling工作互补）

Harness才是Agent落地的真正杠杆！低成本模型+强控制层，性价比直接起飞。欢迎讨论～
#DeepSeek #Agent #StateM #TerminalBench #DeepSeekHarness #Fable5

[EDDY VU @eddyvustg]

💥 DeepSeek-V4-Flash + StateM just CRUSHED Fable 5 for $15.20 vs. $552.67! 💥

That’s a ~36x cost reduction while taking the crown on Terminal-Bench 2.1! 📈

Here’s the breakdown on why this changes everything:

🔥 It’s Not Just Model Scaling

Scaling model size alone won't solve Agent capabilities. Harness Scaling is the new Meta—enter StateM, a state-machine explicit control layer that drives massive performance gains during training.

⚠️ 3 Massive Pain Points with Current AI Agents:
1️⃣ Unenforced Plans: Tools like Codex or Claude Code rely on soft checklists—no hard enforcement or easy human intervention.
2️⃣ No Memory / Experience: Lack of prior workflow context causes planning and self-verification to fail.
3️⃣ Coarse Context Management: Zero granular control over context window execution.

💡 The StateM Solution:

Built on YAML config + CLI execution. It provides explicit state-machine control, auto-halts on failure, and makes workflow fully human-readable and steerable in real time.

🏆 The Benchmark Proof (Terminal-Bench 2.1):

Developers used GPT-5.5 xhigh + Codex to build a generalizable runbook (human sets direction, Agent iterates), then ported it to DeepSeek-V4-Flash for <$38.

Final Result: DeepSeek-V4-Flash + StateM hit the top spot for $15.20 vs Fable 5’s $552.67. The same runbook even supercharged GPT-5.5/5.6, outperforming much costlier Ultra setups!

💡 Takeaway: Harnessing is the ultimate leverage for AI Agent deployment. Combine a low-cost model with a rigid control layer and watch ROI skyrocket.

https://t.co/ykIvAaTyCD

What are your thoughts on Harness Scaling vs Model Scaling? 👇

#DeepSeek #AIAgents #StateM #TerminalBench #DeepSeekHarness #Fable5 #LLM #AIEngineering #BuildInPublic #AIResearch

[DailyPapers @HuggingPapers]

StateM: harness scaling for reliable agents

A runtime that gives agents durable state, checked transitions, and recoverable runbooks. On Terminal-Bench 2.1, it lifts GPT-5.5 to 92.1% and DeepSeek-V4-Flash to 88.1% for under $15.

[Samuel Jayden Shi @Samuel_Jayd]

40X saving v.s. Jacky's 4X saving from STF？🧐

[Qin Ziheng @henryqin1997]

Is model scaling the only source of agent improvement?

We @henryqin1997 @YaxinLu1997 @VITAGroupUT @VictorKaiWang1 are glad to share our work: We reach 95.3% Raw Accuracy, or a $15 Frontier Run (88.8% with DeepSeek-v4-Flash, matching GPT-5.6 Sol Max), on Terminal-Bench 2.1 via Harness Scaling.

Human effort -> Non scalable ❌
Agent direction? Local optimization ❌

Human direction + agent action = better workflow control✅

StateM:
1. Open-source model + StateM -> Close-source frontier, everyone can try!
2. Training free, significantly improve frontier agents.
3. Zero-transfer-cost on same family of agents. [1/8]🧵

📄https://t.co/GjM1VOzWUN
🌐 https://t.co/KapEB3kjXr

[Qin Ziheng @henryqin1997]

Is model scaling the only source of agent improvement?

We @henryqin1997 @YaxinLu1997 @VITAGroupUT @VictorKaiWang1 are glad to share our work: We reach 95.3% Raw Accuracy, or a $15 Frontier Run (88.8% with DeepSeek-v4-Flash, matching GPT-5.6 Sol Max), on Terminal-Bench 2.1 via Harness Scaling.

Human effort -> Non scalable ❌
Agent direction? Local optimization ❌

Human direction + agent action = better workflow control✅

StateM:
1. Open-source model + StateM -> Close-source frontier, everyone can try!
2. Training free, significantly improve frontier agents.
3. Zero-transfer-cost on same family of agents. [1/8]🧵

📄https://t.co/GjM1VOzWUN
🌐 https://t.co/KapEB3kjXr

[Samuel Jayden Shi @Samuel_Jayd]

@TencentHunyuan ASTONISHING! Parametric Memory during Inference. 😃I have just read the work MemoryOS released last year, which focuses on token (input) memory. This work just broader the scope of another implicit memory research line, PARAMETRIC Memory.😆 Insane!

[Tencent Hy @TencentHunyuan]

One static model does not fit all😭

We just dropped our latest work: Functional Neural Memory. Instead of static models, we generate custom "parameters" for every single input.

✅Prompt your model anytime
✅Instant personalization
✅Better instruction following
✅Flexible & dynamic memory (w/o memory bank✌️)

(🧵1/6)

[AK @_akhaliq]

Drag-and-Drop LLMs

Zero-Shot Prompt-to-Weights

[Victor.Kai Wang @VictorKaiWang1]

Customizing Your LLMs in seconds using prompts🥳!
Excited to share our latest work with @HPCAILab, @VITAGroupUT, @k_schuerholt, @YangYou1991, @mmbronstein, @damianborth : Drag-and-Drop LLMs(DnD). 2 features: tuning-free, comparable or even better than full-shot tuning.(🧵1/8)

[Ziqiao Wang @ZiqiaoWang63428]

Representation Alignment (REPA) is NOT ALWAYS helpful for diffusion training!🤷
Sharing latest work w/ @HPCAILab and @VITAGroupUT: "REPA Works Until It Doesn't: Early-Stopped, Holistic Alignment Supercharges Diffusion Training".
Acceleration up to 28x w/o performance drop.(🧵1/7)

[Samuel Jayden Shi @Samuel_Jayd]

Data distillation's new benchmark!! Tackling issues of current DD research.🧐🧐🧐
https://t.co/6cEF5hd70B

[Zekai Li @Richard91316073]

🚀 We’re thrilled to introduce DD-Ranking: Rethinking the Evaluation of Dataset Distillation—a unified, user-friendly, and long-term maintained benchmark for dataset distillation (DD)!
📢 Together with 20 institutions worldwide, we’re releasing our Repo, API Documentation, and Leaderboard:
🔗 Repo: https://t.co/n2WT5ziaJJ
📖 Documentation: https://t.co/9nqogEM350
🏆 Leaderboard: https://t.co/tAHZKXOLH1 🧵1/7

[Zekai Li @Richard91316073]

🚀 We’re thrilled to introduce DD-Ranking: Rethinking the Evaluation of Dataset Distillation—a unified, user-friendly, and long-term maintained benchmark for dataset distillation (DD)!
📢 Together with 20 institutions worldwide, we’re releasing our Repo, API Documentation, and Leaderboard:
🔗 Repo: https://t.co/n2WT5ziaJJ
📖 Documentation: https://t.co/9nqogEM350
🏆 Leaderboard: https://t.co/tAHZKXOLH1 🧵1/7

[Samuel Jayden Shi @Samuel_Jayd]

🔥 ViTs are generated within minutes, not far from 175B LLMs!
Congrats to RPG 🚀！
https://t.co/0vuyc9TKJv

[Victor.Kai Wang @VictorKaiWang1]

Generating ~200 million parameters in just minutes! 🥳
Excited to share our work with @MTDovent , @heisejiasuo96 , and @YangYou1991: 'Recurrent Diffusion for Large-Scale Parameter Generation' (RPG for short).
Example: Obtain customized models using prompts (see below). (🧵1/8)

[Victor.Kai Wang @VictorKaiWang1]

Generating ~200 million parameters in just minutes! 🥳
Excited to share our work with @MTDovent , @heisejiasuo96 , and @YangYou1991: 'Recurrent Diffusion for Large-Scale Parameter Generation' (RPG for short).
Example: Obtain customized models using prompts (see below). (🧵1/8)

[DeepSeek @deepseek_ai]

🚀 Introducing DeepSeek-V3!

Biggest leap forward yet:
⚡ 60 tokens/second (3x faster than V2!)
💪 Enhanced capabilities
🛠 API compatibility intact
🌍 Fully open-source models &amp; papers

🐋 1/n

[Zhenyu (Allen) Zhang @KyriectionZhang]

❓ How much optimization states memory do we need for LLM training ? 🧐Almost zero.

📢 Introducing APOLLO! 🚀 A revolutionary optimizer with SGD-like memory cost, yet AdamW-level performance (or better!).

📜 Paper: https://t.co/jIzU6nyXAy
🔗 GitHub: https://t.co/SKxEC9ZdXp

[Yang Luo @YangL_7]

Training-free Video Enhancement: Achieved 🎉

Nice work with @oahzxl @shaowenqi126301 @VictorKaiWang1 @VitaGroupUT @YangYou1991 et al.

Non-trivial enhancement, training-free, and plug-and-play 🥳

Blog: https://t.co/8Cz78M0L7v (🧵1/6)

[Samuel Jayden Shi @Samuel_Jayd]

[2412.12496] Faster Vision Mamba is Rebuilt in Minutes via Merged Token Re-training

Arxiv: https://t.co/nxe7CNdxCw

[Samuel Jayden Shi @Samuel_Jayd]

Today we release R-MeeTo (Re-training Merged Token) the first Token Merging method for Mamba.🥳Faster Mamba can be achieved in Minutes-Level with minimal degradation of performance.

PDF:https://t.co/oXLu8ViQ9N
Home:https://t.co/SAwPxGVknc

[Samuel Jayden Shi @Samuel_Jayd]

Modification: key knowledge is consist of general and specific knowledge. General Knowledge includes the common partterns shared among tokens, and Specific Knowledge indicates the specific partterns in particular tokens. They look like this in Mamba: