๐ป Code Intelligence¶
๐ฌ EMNLP2026 ยท 8 paper notes
๐ Same area in other venues: ๐ง NeurIPS2026 (43) ยท ๐๏ธ ECCV2026 (2) ยท ๐ฌ ICLR2026 (59) ยท ๐ฌ ACL2026 (49) ยท ๐งช ICML2026 (22) ยท ๐ค AAAI2026 (10)
๐ฅ Top topics: LLM ร4 ยท Code Intelligence ร2
- ChainSWE: Benchmarking Coding Agents on Multi-Bug Software Maintenance
-
Introduces ChainSWE, the first benchmark evaluating coding agents on sequential multi-bug software maintenance across shared codebases, demonstrating that accumulated repository state drift causes resolution rates to drop by up to 70% along the repair chain.
- DSEffi-Bench: Demystifying Large Language Modelsโ Capabilities in Data Science Programming
-
Introduces DSEffi-Bench, the first execution efficiency benchmark for data science code generation (comprising 1,000 human-validated instances from real Stack Overflow pain points paired with stress-testing suites), revealing significant decoupling between functional correctness and runtime efficiency, uncovering that 79.1% of performance deficits stem from domain-specific root causes, and demonstrating up to +14.7% efficiency improvement and 13.0ร cost reduction via taxonomy-guided optimization and library-conditioned routing.
- Fast and Focused: Interactive Repository-Level Fill-in-the-Middle Code Completion with Hybrid Structural Retrieval
-
To tackle coarse retrieval granularity and suffix duplication in interactive repository-level code completion, ACAR combines lexical sliding-window and AST/LSP structural retrieval via line-level reciprocal rank fusion (RRF) alongside syntax-aware post-processing, outperforming strong baselines with only 87โ95 ms non-decoding overhead.
- Functional Entropy: Predicting Functional Correctness in LLM-Generated Code with Uncertainty Quantification
-
Addressing the collapse of NLI-based semantic entropy caused by its inability to distinguish functional code differences, this paper introduces a family of functional equivalence methods powered by LLM-based functional equivalence assessment to accurately predict code correctness across multiple programming languages without code execution.
- PolySQL: Scaling Text-to-SQL Evaluation Across SQL Dialects via Automated Backend Isomorphism
-
PolySQL introduces a dual-execution evaluation methodology based on backend isomorphism and normalized comparison that eliminates the need for fragile query transpilation, uncovering a 25.2% relative performance degradation and predominant logical regression across SQL dialects with 100% execution coverage.
- SABER: Benchmarking Operational Safety of LLM Coding Agents in Stateful Project Workspaces
-
SABER establishes a stateful benchmark of 716 executable Docker-sandboxed project tasks to evaluate the operational safety of LLM coding agents from multi-step execution traces and persistent workspace state deltas, revealing that even frontier models exhibit over 54.7% harmful safety-violation rates.
- Verification Trap: Understanding Test-Time Selection Failures under False Premises in Code Generation
-
This paper demonstrates that test-time verifier selection in code generation fails systematically under misleading task premises because the generator and verifier become coupled through shared mistaken assumptions; verifier-written tests omit counterexamples and validate premise-consistent shortcuts over correct programs, a failure that can be predicted prior to hidden execution and mitigated via decoupled evidence channels.
- XRepoTest: Benchmarking Multilingual Repository-Level Unit Test Generation for Large Language Models
-
XRepoTest establishes the first repository-level unit test generation benchmark across 5 high-impact yet underexplored languages (Rust, Go, Julia, PHP, and Ruby) spanning 3,642 focal methods, introducing Invocation Rate (IR) to expose misleading indirect test coverage and revealing that LLMs face severe bottlenecks under real project constraints with test pass rates peaking below 27%.