PhysicsWallahAI

/

Aryabhata-1.0

like 110

Follow

PhysicsWallah 141

Hi team,

First of all, thank you for open-sourcing Aryabhata-1.0. We have been evaluating the model for educational math-solution generation workflows and found its mathematical reasoning capabilities to be quite strong.

During our experiments, we observed a few architectural behaviors that we wanted to discuss with the community and maintainers to better understand whether these are expected characteristics of the model, or whether there are recommended mitigation strategies.

Observations

Aryabhata appears to be strongly optimized toward detailed reasoning and step-by-step derivations. While this is valuable for deep mathematical explanation, we faced challenges when attempting to generate concise or short-step solutions consistently.

We specifically observed:

1. The model tends to generate long Chain-of-Thought (CoT) reasoning by default.
2. Prompt tightening alone does not reliably reduce verbosity.
3. The reasoning trajectory often expands into exploratory derivations even when concise solutions are requested.

Continuation / Context Window Challenges

In our current inference environment, we are operating under relatively constrained generation windows (~1024 tokens total).
When attempting continuation-based generation for longer solutions, we consistently observed:

• Loss of latent reasoning continuity between generations
• Shift in reasoning trajectory after continuation
• Duplicate or conflicting derivation paths
• Coherence degradation across merged outputs

In many cases:

• the initial generation follows one reasoning path,
• while the continuation begins exploring a different path altogether.

This makes deterministic multi-stage generation difficult for production workflows.

Questions

We wanted to understand from the maintainers/community:

1. Is Aryabhata intentionally optimized toward verbose reasoning traces?
2. Are there recommended decoding parameters or prompting patterns for obtaining concise solutions?
3. Has the team experimented with:

• controllable reasoning depth,
• abbreviated derivation modes,
• or answer-first generation strategies?

4. Are continuation-instability behaviors expected for this architecture under constrained context settings?
5. Would larger context support significantly improve reasoning continuity?

We would really appreciate any guidance, recommended inference settings, or architectural insights from the team.

Thanks again for releasing the model and making these experiments possible.