SR
S.M. Rasing
info
Please Note
<p>This page displays the records of the person named above and is not linked to a unique person identifier. This record may need to be merged to a profile.</p>
2 records found
1
Existing Inductive Program Synthesis (IPS) techniques generate programs from
examples and are restricted to specific task classes to exploit domain knowledge.
To compete with general-purpose LLM-based code generation, a single IPS system
must handle tasks with different types of inputs, outputs, and instructions without
relying on domain-specific knowledge. This thesis introduces task-level adapta-
tion, a framework in which a synthesizer must learn how to search only from the in-
dividual task it is solving rather than from a predefined domain. We focus specific-
ally on search guided by automatically generated properties that capture useful
aspects of a program’s output. We introduce PHALCON to demonstrate the frame-
work’s viability. PHALCON repeatedly samples programs, selects properties that
distinguish incorrect outputs from intended behavior, and uses them to constrain
subsequent sampling. PHALCON outperforms state-of-the-art techniques special-
ized for string and bit-vector transformations, and solves tasks from the challenging
Abstraction and Reasoning Corpus. These results establish task-level adaptation as
a promising foundation for efficient, general program synthesis. ...
examples and are restricted to specific task classes to exploit domain knowledge.
To compete with general-purpose LLM-based code generation, a single IPS system
must handle tasks with different types of inputs, outputs, and instructions without
relying on domain-specific knowledge. This thesis introduces task-level adapta-
tion, a framework in which a synthesizer must learn how to search only from the in-
dividual task it is solving rather than from a predefined domain. We focus specific-
ally on search guided by automatically generated properties that capture useful
aspects of a program’s output. We introduce PHALCON to demonstrate the frame-
work’s viability. PHALCON repeatedly samples programs, selects properties that
distinguish incorrect outputs from intended behavior, and uses them to constrain
subsequent sampling. PHALCON outperforms state-of-the-art techniques special-
ized for string and bit-vector transformations, and solves tasks from the challenging
Abstraction and Reasoning Corpus. These results establish task-level adaptation as
a promising foundation for efficient, general program synthesis. ...
Existing Inductive Program Synthesis (IPS) techniques generate programs from
examples and are restricted to specific task classes to exploit domain knowledge.
To compete with general-purpose LLM-based code generation, a single IPS system
must handle tasks with different types of inputs, outputs, and instructions without
relying on domain-specific knowledge. This thesis introduces task-level adapta-
tion, a framework in which a synthesizer must learn how to search only from the in-
dividual task it is solving rather than from a predefined domain. We focus specific-
ally on search guided by automatically generated properties that capture useful
aspects of a program’s output. We introduce PHALCON to demonstrate the frame-
work’s viability. PHALCON repeatedly samples programs, selects properties that
distinguish incorrect outputs from intended behavior, and uses them to constrain
subsequent sampling. PHALCON outperforms state-of-the-art techniques special-
ized for string and bit-vector transformations, and solves tasks from the challenging
Abstraction and Reasoning Corpus. These results establish task-level adaptation as
a promising foundation for efficient, general program synthesis.
examples and are restricted to specific task classes to exploit domain knowledge.
To compete with general-purpose LLM-based code generation, a single IPS system
must handle tasks with different types of inputs, outputs, and instructions without
relying on domain-specific knowledge. This thesis introduces task-level adapta-
tion, a framework in which a synthesizer must learn how to search only from the in-
dividual task it is solving rather than from a predefined domain. We focus specific-
ally on search guided by automatically generated properties that capture useful
aspects of a program’s output. We introduce PHALCON to demonstrate the frame-
work’s viability. PHALCON repeatedly samples programs, selects properties that
distinguish incorrect outputs from intended behavior, and uses them to constrain
subsequent sampling. PHALCON outperforms state-of-the-art techniques special-
ized for string and bit-vector transformations, and solves tasks from the challenging
Abstraction and Reasoning Corpus. These results establish task-level adaptation as
a promising foundation for efficient, general program synthesis.
Brute, A state-of-the-art inductive program synthesis (IPS) system, introduced a two-phase algorithm; first, complex pro- gram instructions are invented from basic instructions. Sec- ond, a best-first search algorithm finds a sequence of invented instructions to solve an IPS task. This method is limited because invented instructions are always of the same com- plexity, also when less or more complexity is needed. Also, best-first search falls into local optima easily. In this paper, I describe Vlute, an IPS system using Large Neighborhood Search (LNS), in which a solution is gradually improved by exploring neighboring solutions, and Variable-Depth Invent (VDI), in which instruction complexity is increased dynami- cally. Vlute is tested on three IPS domains (robot-planning, string transformations, and drawing ASCII-art). Results show that using VDI improves Vlute’s performance only for string transformation. Vlute can outperform Brute and escape local optima encountered by Brute also only for string transforma- tion. A limitation of Vlute is finding large programs.
...
Brute, A state-of-the-art inductive program synthesis (IPS) system, introduced a two-phase algorithm; first, complex pro- gram instructions are invented from basic instructions. Sec- ond, a best-first search algorithm finds a sequence of invented instructions to solve an IPS task. This method is limited because invented instructions are always of the same com- plexity, also when less or more complexity is needed. Also, best-first search falls into local optima easily. In this paper, I describe Vlute, an IPS system using Large Neighborhood Search (LNS), in which a solution is gradually improved by exploring neighboring solutions, and Variable-Depth Invent (VDI), in which instruction complexity is increased dynami- cally. Vlute is tested on three IPS domains (robot-planning, string transformations, and drawing ASCII-art). Results show that using VDI improves Vlute’s performance only for string transformation. Vlute can outperform Brute and escape local optima encountered by Brute also only for string transforma- tion. A limitation of Vlute is finding large programs.