Dingjie James Song

PhD student @LehighU | Intern @salesforce | Prev RA @cuhksz | MEng & BEng @NanjingUnivers1 | AI4Science,Agent

[Dingjie James Song @bbsngg1002]

Think deeper. Skip the reasoning text. 🧠

Meet 🌀 Lev: System 1.5 decision models with a depth dial. An independent, open-source take on OpenAI Decisions-style text decisions.

🔁 One checkpoint scores the options after every loop.

🔓 Paper, code and weights are open. 🧵👇

[Dingjie James Song @bbsngg1002]

@Max_Black_4411 I used Claude Code to make this

[Dingjie James Song @bbsngg1002]

Think deeper. Skip the reasoning text. 🧠

Meet 🌀 Lev: System 1.5 decision models with a depth dial. An independent, open-source take on OpenAI Decisions-style text decisions.

🔁 One checkpoint scores the options after every loop.

🔓 Paper, code and weights are open. 🧵👇

[Dingjie James Song @bbsngg1002]

Thanks for reading the table that closely! Agreed: on Lev-2.6B most of the gain is from depth 1 to 3 (63.2 → 75.3%), and step 4 adds 0.9 pt for another 50 ms. Typed-decisions is a good suggestion and per-depth Brier is exactly what we’d want to see there, so we’ll look into running it zero-shot.

[Dingjie James Song @bbsngg1002]

Thanks for the question. We don't catch it automatically. It's actually our most common false-alarm cause (21 of 85 in the failure analysis). Since we take the min over premises, an invented premise only causes a false alarm, never a missed error. Each flag also shows the premise behind it, so a person can spot it quickly. Keeping the decomposer closer to what the reviewer said is our future work.

[Dingjie James Song @bbsngg1002]

📄 PremiseCheck, efficient fact-checking of human and AI peer reviews 🧵

Review comments, by humans or AI, sometimes misstate what a paper says. 🤔 Can we flag them without a frontier LLM rereading the whole paper for every comment?

https://t.co/JOhQI8kZh2

[Dingjie James Song @bbsngg1002]

@solomonmithra99 Thanks! The shared readout + joint loss over steps is what makes every depth answerable. Train only the last step and the shallow readouts are useless. We haven’t measured drift in the residual stream itself, though that would be a nice follow-up.

[Dingjie James Song @bbsngg1002]

https://t.co/xNAiGHndtC

[Dingjie James Song @bbsngg1002]

📄 Paper: https://t.co/JOhQI8kZh2
💬 alphaXiv: https://t.co/YioESe2fxj
🤗 HF Papers: https://t.co/cdhWfOyaXg
💻 Code: https://t.co/zMQQdksSCW
📦 Data: https://t.co/X4VFKRsR9A
🌐 Page: https://t.co/m5Cg3qIlWj

🙏 With Zhiling Yan, Ziqian Zhang, Yuxuan Zhang, and Lichao Sun.

[Dingjie James Song @bbsngg1002]

🔓 Also in the paper: with the open-weight Kev-27B as decision model, NLP AUROC is the same as with Jev. Checking a whole review at once, fusing batched PremiseCheck and GPT-5.5 scores gives 71.5 AUROC at 28% of the reference cost. 🚩 Each flag names the premise found false.

[Dingjie James Song @bbsngg1002]

8/8
📄 Paper: https://t.co/g4SHCIObL2
🤗 HF Papers: https://t.co/xNAiGHndtC
💻 Code: https://t.co/xsX3NyThOA
📦 Models + data: https://t.co/Tsgnmz40xD
🌐 Project: https://t.co/bgAbAtSNwr

💬 What decision workload would you give adjustable depth?

[Dingjie James Song @bbsngg1002]

7/8 🙏 Built on Ouro and Nanbeige backbones, with Kev's decision-model foundation. @nanbeige @jaredpalmer Thanks also to JevBench. @airesearch12

Lev adds joint depth supervision and workload stopping analysis. An independent research implementation, with no OpenAI model or API equivalence claim.