THPL
A generative decision-making framework for precision feeding of rainbow trout in recirculating aquaculture systems. It integrates fish-school kinematic information with LLM reasoning capabilities through a vision-language model fusion to enable interpretable, automated feeding decision support.
Timeline (last 90 days)
THPL is a generative feeding decision-making framework for rainbow trout in recirculating aquaculture systems, combining continuous fish population kinematics information with the reasoning capabilities of large language models.
The THPL framework adopts a three-stage technical approach, decomposing the process from observing fish populations to feeding decisions into three progressive steps: trajectory-to-activity coefficient, hierarchical behavior encoder, and LLM fine-tuning.
The Spearman correlation coefficient between the Activity Coefficient AC and expert-annotated feeding intensity reached ρ=0.925, p<0.001, showing a statistically significant monotonic positive correlation.
THPL fine-tunes the large language model via LoRA and introduces counterfactual multimodal Direct Preference Optimization (counterfactual mDPO) to reinforce the model's causal reasoning capability.
The text-only baseline model achieved a decision accuracy of 33.33%, which improved to 93.33% after introducing dual evidence tokens.
Continuous spatiotemporal tokens provide large language models with the necessary physical grounding, as textual descriptions alone are insufficient to support feeding decisions.
The validation results of THPL were obtained under relatively controlled experimental conditions, and whether they can remain robust in actual aquaculture environments still awaits larger-scale, long-term validation.
All Facts (7)
THPL is a generative feeding decision-making framework for rainbow trout in recirculating aquaculture systems, combining continuous fish population kinematics information with the reasoning capabilities of large language models.
50%UnverifiedThe THPL framework adopts a three-stage technical approach, decomposing the process from observing fish populations to feeding decisions into three progressive steps: trajectory-to-activity coefficient, hierarchical behavior encoder, and LLM fine-tuning.
50%UnverifiedThe Spearman correlation coefficient between the Activity Coefficient AC and expert-annotated feeding intensity reached ρ=0.925, p<0.001, showing a statistically significant monotonic positive correlation.
50%UnverifiedTHPL fine-tunes the large language model via LoRA and introduces counterfactual multimodal Direct Preference Optimization (counterfactual mDPO) to reinforce the model's causal reasoning capability.
50%UnverifiedThe text-only baseline model achieved a decision accuracy of 33.33%, which improved to 93.33% after introducing dual evidence tokens.
50%UnverifiedContinuous spatiotemporal tokens provide large language models with the necessary physical grounding, as textual descriptions alone are insufficient to support feeding decisions.
50%UnverifiedThe validation results of THPL were obtained under relatively controlled experimental conditions, and whether they can remain robust in actual aquaculture environments still awaits larger-scale, long-term validation.
50%