iaa01/CIA-1.7B

TEXT GENERATIONPricing:Input $0.32 / Cached $0.064 / Output $1.6Concurrent Unit Cost:1Model Size:2BQuant:BF16Context Size:32kTool Calling:SupportedPublished:Jan 11, 2026License:mitArchitecture:Transformer0.0K Open Weights Featherless Exclusive Cold

The iaa01/CIA-1.7B is a 2 billion parameter language model post-trained with Reinforcement Learning using the ∆Belief reward signal. This optimization method rewards the model for reducing its own belief uncertainty, providing dense feedback for long-horizon, open-ended information-seeking tasks. It learns general information-seeking strategies, particularly demonstrated in the Twenty Questions environment.

Loading preview...

Overview

iaa01/CIA-1.7B is a 2 billion parameter language model that has undergone reinforcement learning (RL) post-training. Its key differentiator is the use of the ∆Belief reward signal, a novel approach that rewards the model for actively reducing its own uncertainty over time, rather than relying solely on sparse success signals.

Key Capabilities

  • Dense Feedback: The ∆Belief reward signal provides continuous, turn-level credit assignment, which is crucial for complex, multi-step tasks.
  • Information-Seeking Optimization: It is specifically optimized for long-horizon, open-ended information-seeking tasks.
  • Generalization: The model learns general strategies for information acquisition, demonstrated by its training in the Twenty Questions environment, which can generalize beyond the specific training context.
  • Context Length: Supports a substantial context length of 32768 tokens.

Good For

  • Applications requiring models to actively seek information.
  • Scenarios where dense, continuous feedback is more effective than sparse rewards.
  • Developing agents that can navigate and resolve uncertainty in complex environments.

For more technical details, refer to the ∆Belief-RL project page and its code repository.

Finetunes
1 models