# US 2025/0156682 A1 · Generating symbolic plans using transformer-based models

> Agents: this data is available as JSON from the patentgrep API and MCP server — see https://patentgrep.com/llms.txt

- Number: US 2025/0156682 A1
- Document ID: US20250156682A1
- Status: Published application
- Kind: A1
- Application number: 18/509,359
- Filed: 2023-11-15
- Published: 2025-05-15
- Earliest claimed: 2023-11-15
- Assignees: International Business Machines Corporation; University of South Carolina; University of Brescia Italy
- Applicants: International Business Machines Corporation; University of South Carolina; University of Brescia Italy
- Inventors: Francesca Rossi; Lior Horesh; Keerthiram Murugesan; Vishal Pallagani; Biplav Srivastava; Andrea Loreggia
- Web page: https://patentgrep.com/patent/US20250156682A1

## CPC

- G06N 3/0455: Auto-encoder networks; Encoder-decoder networks
- G06N 3/045: Combinations of networks
- G06N 3/09: Supervised learning
- G06N 3/096: Transfer learning
- G06N: COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS

## Abstract

Embodiments of the invention are directed to a computer-implemented method that includes inputting a first input into a plansformer that includes a transformer-based neural network (NN). The first input includes symbols and a problem. The computer-implemented method further includes, in response to the inputting, receiving as output from the plansformer a plan for solving the problem.

## Claims

### Claim 1

A computer-implemented method comprising:
inputting a first input into a plansformer comprising a transformer-based neural network (NN), the first input comprising symbols and a problem; and
in response to the inputting, receiving as output from the plansformer a plan for solving the problem.

### Claim 2 (depends on 1)

The computer-implemented method of claim 1, wherein the symbols comprise computer code.

### Claim 3 (depends on 1)

The computer-implemented method of claim 1, wherein the plan comprises more symbols.

### Claim 4 (depends on 1)

The computer-implemented method of claim 1, further comprising training the plansformer via submitting training data to a large language model (LLM).

### Claim 5 (depends on 4)

The computer-implemented method of claim 4, wherein the training data comprises planning problems and associated plans generated from a test domain.

### Claim 6 (depends on 5)

The computer-implemented method of claim 5, wherein:
the planning problems are received from a problem generator in response to a domain model being input into the problem generator; and
the associated plans are received from a progression planner in response to the planning problems being input into the progression planner.

### Claim 7 (depends on 4)

The computer-implemented method of claim 4, wherein the LLM comprises a code-aware encoder-decoder architecture.

### Claim 8 (depends on 7)

The computer-implemented method of claim 7, wherein the code-aware encoder-decoder architecture is pre-trained using one or more code-related tasks selected from a group consisting of code summarization, code generation, code translation, code refinement, code defect detection, code clone detection, and text-code matching.

### Claim 9 (depends on 4)

The computer-implemented method of claim 4, wherein the code-aware encoder-decoder architecture implements masked language modeling.

### Claim 10 (depends on 1)

The computer-implemented method of claim 1, further comprising evaluating the plan to produce a confidence score.

### Claim 11 (depends on 1)

The computer-implemented method of claim 1, further comprising evaluating the plan for validity and optimality.

### Claim 12 (depends on 11)

The computer-implemented method of claim 11, wherein the evaluating for validity and optimality comprises applying relaxation conditions to the plan and the first input.

### Claim 13 (depends on 1)

The computer-implemented method of claim 1, further comprising evaluating the plan via a natural language metric tool.

### Claim 14 (depends on 1)

The computer-implemented method of claim 1, wherein the plan comprises a series of time steps and one or more actions to instantiate for each of the time steps.

### Claim 15 (depends on 1)

The computer-implemented method of claim 1, wherein the plansformer comprises a tokenizer that produces planning-language specific tokens.

### Claim 16

A computer system comprising a processor system electronically coupled to a memory, wherein the processor system performs processor system operations comprising:
inputting a first input into a plansformer comprising a transformer-based neural network (NN), the first input comprising symbols and a problem; and
in response to the inputting, receiving as output from the plansformer a plan for solving the problem.

### Claim 17 (depends on 16)

The computer system of claim 16, wherein:
the symbols comprise computer code;
the plan comprises more symbols;
the processor system operation further comprise training the plansformer via submitting training data to a large language model (LLM); and
the training data comprises planning problems and associated plans generated from a test domain.

### Claim 18 (depends on 17)

The computer system of claim 17, wherein:
the planning problems are received from a problem generator in response to a domain model being input into the problem generator;
the associated plans are received from a progression planner in response to the planning problems being input into the progression planner; and
the LLM comprises a code-aware encoder-decoder architecture.

### Claim 19

A computer program product comprising a computer readable program stored on a computer readable storage medium, wherein the computer readable program, when executed on a processor system, causes the processor to perform processor system operations comprising:
inputting a first input into a plansformer comprising a transformer-based neural network (NN), the first input comprising symbols and a problem; and
in response to the inputting, receiving as output from the plansformer a plan for solving the problem,
wherein the symbols comprise computer code;
wherein the plan comprises more symbols; and
training the plansformer via submitting training data to a large language model (LLM);
wherein the training data comprises planning problems and associated plans generated from a test domain.

### Claim 20 (depends on 19)

The computer program product of claim 19, wherein:
the planning problems are received from a problem generator in response to a domain model being input into the problem generator;
the associated plans are received from a progression planner in response to the planning problems being input into the progression planner; and
the LLM comprises a code-aware encoder-decoder architecture.
