arXiv subjects

Yusen Su

Publications and source records attributed to Yusen Su.

4 recordsLinked to original sources

arXiv2025

Deliberate Planning in Language Models with Symbolic Representation

Planning remains a core challenge for large language models (LLMs), particularly in domains that require coherent multi-step action sequences grounded in external constraints. We introduce SymPlanner, a novel framework that equips LLMs with structured planning capabilities by interfacing them with a symbolic environment that serves as an explicit world model. Rather than relying purely on natural language reasoning, SymPlanner grounds the planning process in a symbolic state space, where a policy model proposes actions and a symbolic environment deterministically executes and verifies their effects. To enhance exploration and improve robustness, we introduce Iterative Correction (IC), which refines previously proposed actions by leveraging feedback from the symbolic environment to eliminate invalid decisions and guide the model toward valid alternatives. Additionally, Contrastive Ranking (CR) enables fine-grained comparison of candidate plans by evaluating them jointly. Conceptually, SymPlanner operationalizes two cognitive faculties: (i) error monitoring and repair via externalized feedback (IC) and (ii) preference formation among alternatives via pairwise comparison (CR), advancing cognitively plausible, symbol-grounded planning aligned with the rich structure in intelligent systems. We evaluate SymPlanner on PlanBench, demonstrating that it produces more coherent, diverse, and verifiable plans than pure natural language baselines.

cs.CL↗

arXiv2024

Automatic Inference of Relational Object Invariants

Relational object invariants (or representation invariants) are relational properties held by the fields of a (memory) object throughout its lifetime. For example, the length of a buffer never exceeds its capacity. Automatic inference of these invariants is particularly challenging because they are often broken temporarily during field updates. In this paper, we present an Abstract Interpretation-based solution to infer object invariants. Our key insight is a new object abstraction for memory objects, where memory is divided into multiple memory banks, each containing several objects. Within each bank, the objects are further abstracted by separating the most recently used (MRU) object, represented precisely with strong updates, while the rest are summarized. For an effective implementation of this approach, we introduce a new composite abstract domain, which forms a reduced product of numerical and equality sub-domains. This design efficiently expresses relationships between a small number of variables (e.g., fields of the same abstract object). We implement the new domain in the CRAB abstract interpreter and evaluate it on several benchmarks for memory safety. We show that our approach is significantly more scalable for relational properties than the existing implementation of CRAB. For evaluating precision, we have integrated our analysis as a pre-processing step to SEABMC bounded model checker, and show that it is effective at both discharging assertions during pre-processing, and significantly improving the run-time of SEABMC.

cs.PL↗

arXiv2021

Verifying Verified Code

A recent case study from AWS by Chong et al. proposes an effective methodology for Bounded Model Checking in industry. In this paper, we report on a follow up case study that explores the methodology from the perspective of three research questions: (a) can proof artifacts be used across verification tools; (b) are there bugs in verified code; and (c) can specifications be improved. To study these questions, we port the verification tasks for $\texttt{aws-c-common}$ library to SEAHORN and KLEE. We show the benefits of using compiler semantics and cross-checking specifications with different verification techniques, and call for standardizing proof library extensions to increase specification reuse. The verification tasks discussed are publicly available online.

cs.SE↗

arXiv2020

Data Flow Refinement Type Inference

Refinement types enable lightweight verification of functional programs. Algorithms for statically inferring refinement types typically work by reduction to solving systems of constrained Horn clauses extracted from typing derivations. An example is Liquid type inference, which solves the extracted constraints using predicate abstraction. However, the reduction to constraint solving in itself already signifies an abstraction of the program semantics that affects the precision of the overall static analysis. To better understand this issue, we study the type inference problem in its entirety through the lens of abstract interpretation. We propose a new refinement type system that is parametric with the choice of the abstract domain of type refinements as well as the degree to which it tracks context-sensitive control flow information. We then derive an accompanying parametric inference algorithm as an abstract interpretation of a novel data flow semantics of functional programs. We further show that the type system is sound and complete with respect to the constructed abstract semantics. Our theoretical development reveals the key abstraction steps inherent in refinement type inference algorithms. The trade-off between precision and efficiency of these abstraction steps is controlled by the parameters of the type system. Existing refinement type systems and their respective inference algorithms, such as Liquid types, are captured by concrete parameter instantiations. We have implemented our framework in a prototype tool and evaluated it for a range of new parameter instantiations (e.g., using octagons and polyhedra for expressing type refinements). The tool compares favorably against other existing tools. Our evaluation indicates that our approach can be used to systematically construct new refinement type inference algorithms that are both robust and precise.

cs.PL↗