Hitachi, Ltd.
# Hitachi, Ltd.’s LLM optimization does not settle on a single best answer
Optimization with LLMs can become a technology that lays out different trade-offs for people in the field to choose from, rather than one that guesses the only correct answer. Hitachi, Ltd. has developed a training technology that generates candidates for multiple KPIs, raising candidate coverage by up to 13.3% on logic-circuit tasks. If the candidates meet real-world constraints and safety conditions, logistics staff could choose vehicle-dispatch plans with different speeds and energy savings on the same screen.
In railway operations, delivery, and the allocation of power and computing resources, increasing speed can raise power consumption and costs, while prioritizing stability can mean sacrificing other performance. People in the field often combine multiple conditions into a single score and receive the plan with the highest score. They may also need to retrain the AI every time the conditions change, making it difficult to compare which trade-off to choose.
A single score cannot compare everything
An LLM is an AI that can generate not only text but also design and planning proposals. KPI (key performance indicator) is the yardstick used to measure those proposals. For a circuit, for example, power consumption, processing speed, and circuit size can serve as KPIs.
These cannot always all be improved at once. A fast circuit may use more power, while a smaller circuit may be slower. In multi-objective optimization, such proposals are not forced into a single ranking. Instead, they are compared as a group of candidates whose relative superiority is difficult to determine outright.
Teaching the model how candidates differ
Hitachi, Ltd. has developed an LLM training technology that creates multiple proposals with different balances, without collapsing multiple KPIs into a single overall score. It combines GRPO, a reinforcement-learning method for LLMs, with a metric called hypervolume contribution. This measures how broad a range the group of candidates covers across multiple yardsticks.
The AI creates its next candidate while looking at the candidates it has generated so far. It learns by evaluating whether the new proposal improves the performance of the candidate group as a whole and adds a proposal that is not too similar to the others. The aim is to produce not only proposals biased toward a particular weighting, but also proposals that prioritize speed, save energy, or keep size down. In logic-circuit generation, candidate coverage improved by up to 13.3% compared with the model before training, the company says. However, this result comes from circuit tasks whose performance is relatively easy to score automatically.
Toward proposals people can choose from in the field
If this technology proves effective while accounting for the constraints and uncertainty of real-world operations, as well as safety conditions, dispatchers at logistics centers could compare on the same screen a plan that meets deadlines, one that charges outside periods of high electricity prices, and one that preserves spare vehicle capacity. In a design room, they could choose among circuits with the same functionality: one that extends battery life, one that is faster, and one that fits on a smaller circuit board.
However, it will be necessary to verify that candidates meet the constraints of real vehicles, roads, contracts, and labor conditions, and to connect them to existing optimization software and operational systems. Responsibilities in the event of failure must also be decided. The value lies not in automatically adopting a single optimal solution, but in leaving people working in changing circumstances with trade-offs they can choose from.