The Solution? A Private Working Memory.
We propose an architecture with a "scratchpad" that:
✅ Persists across turns
✅ Is readable/writable by the agent
✅ Is NEVER shown to the user
This Generative-Retention Loop allows agents to finally play Hangman correctly.

We prove an Impossibility Theorem: Standard chat agents face a fundamental trade-off between Secrecy and Consistency.

1️⃣ If the secret is in the context, then it’s leaked (violates Secrecy).
2️⃣ If it’s NOT in the context, then the model forgets it (violates Consistency).

Can LLMs play Hangman? Spoiler alert: Not yet.
Check out “LLMs Can’t Play Hangman: On the Necessity of a Private Working Memory for Language Agents”, led by Davide Baldelli, Ali Parviz, AmalZouaq and Sarath Chandar.

Learn more at:

📝 Paper: arxiv.org/abs/2507.09792
💻 Code: github.com/chandar-lab/...
🤗 Models & Dataset: huggingface.co/collections/...
🌐 Website: chandar-lab.github.io/cadmium-webs...
📋 Blog post: t.co/c3b6U3bIWl

With human-written prompts, CADmium’s 14B fine-tuned model achieves:

⬆️ Higher F1 on lines, circles, extrusions
⬆️ Better Chamfer & curvature metrics
⬆️ Stronger structural similarity
We avoid extensive pre-training, customized embeddings, and other domain-specific strategies.

Compared to previous CAD annotations, CADmium’s expert-level descriptions are:

📈 Shorter, clearer, and more natural-sounding
📈 Richer vocabulary (2× more unique words)
📈 Preferred by humans for readability & accuracy

📊 Plots illustrating better lexical diversity and conciseness!

Highlights:
✅ GPT-4.1 generates concise, expert-level, human-like geometric descriptions of 176k models.
✅ Qwen2.5-Coder (fine-tuned with LoRA) translates them back into CAD sequences in JSON format.
✅ New metrics to evaluate the structural and topological characteristics.

Computer-Aided Design (CAD) models are built step-by-step with sketches and extrusions. What if we could generate them directly from natural language?

🤖 Meet CADmium - a new dataset and fine-tuning framework for solving the text-to-CAD problem.