CITA
CITA is a comparative reasoning framework for long-horizon tool-calling scenarios. It trains a Comparative Inference Model (CIM) to evaluate the relative value of candidate tool calls, combines a Bayesian tool-graph simulator with LLM semantic judgment to provide step-level supervision signals, and addresses the credit-assignment problem in agent tool use.
Timeline (last 90 days)
CITA (Comparative Inference for Tool-use Agents) is a comparative inference framework for tool-use agents, whose core is training a Comparative Inference Model (CIM) to estimate the probability that a candidate tool call supports ultimate task success.
The core idea of CITA is that an efficient tool-use agent should estimate the long-range value of a tool call before executing it — that is, 'selection before action.'
CIM training relies on pairwise signals, integrating three complementary supervision sources: observed tool behavior, a Bayesian tool graph simulator, and LLM semantic judgment.
The Bayesian tool graph simulator can estimate the impact on subsequent success probability of choosing tool X instead of Y at a given step without actual execution, thereby generating counterfactual contrastive samples.
The researchers validated CITA on three tool-use benchmarks in combination with multiple different backbone LLMs, and the results show consistent improvements on both Tool F1 and task success rate metrics.
CITA transforms the problem from hard-to-obtain absolute value supervision into more tractable comparative supervision, providing a new technical path for long-range decision-making tasks.
All Facts (6)
CITA (Comparative Inference for Tool-use Agents) is a comparative inference framework for tool-use agents, whose core is training a Comparative Inference Model (CIM) to estimate the probability that a candidate tool call supports ultimate task success.
50%UnverifiedThe core idea of CITA is that an efficient tool-use agent should estimate the long-range value of a tool call before executing it — that is, 'selection before action.'
50%UnverifiedCIM training relies on pairwise signals, integrating three complementary supervision sources: observed tool behavior, a Bayesian tool graph simulator, and LLM semantic judgment.
50%UnverifiedThe Bayesian tool graph simulator can estimate the impact on subsequent success probability of choosing tool X instead of Y at a given step without actual execution, thereby generating counterfactual contrastive samples.
50%UnverifiedThe researchers validated CITA on three tool-use benchmarks in combination with multiple different backbone LLMs, and the results show consistent improvements on both Tool F1 and task success rate metrics.
50%UnverifiedCITA transforms the problem from hard-to-obtain absolute value supervision into more tractable comparative supervision, providing a new technical path for long-range decision-making tasks.
50%