ablation-study

This paper investigates whether complex transformer-based attention is necessary for global spatial information extraction in traffic forecasting, finding that simple global aggregation operators achieve comparable performance with lower computational complexity.

This paper systematically analyzes multiple pipeline extension components for Natural Language to SQL translation, including intermediate representations, synthetic data, preprocessing, and a reranker, using SmBoP and RASAT architectures, and finds that their interactions matter more than simply combining all components.

This paper proposes INTENT, an LSTM framework for predicting vehicle intentions (straight, left turn, right turn) at intersections 2 seconds ahead, achieving 99.71% accuracy on the InD dataset.

This paper investigates whether LLM-generated reusable skill files improve performance of AI data scientists across data science workflows. Through extensive ablation experiments involving over 9,000 runs, the authors find no significant improvement from generated skills over baseline prompting, cautioning against their default use.

This paper presents a candidate-constrained RAG system for the LongEval-RAG task at CLEF 2026, combining deterministic provenance tracking with passage retrieval, query expansion, pseudo-relevance feedback, reciprocal rank fusion, evidence reranking, and citation-aware aggregation. An ablation study of ten pipeline variants shows that a rule-based chunking pipeline with sentence-level neural selection achieves the best performance.

This paper presents a five-arm ablation methodology for diagnosing which component of retrieval-warmed energy-based reasoning (RW-EBR) drives performance gains, applied to structured reasoning tasks like graph reachability and Sudoku. The method separates effects of class-prior bias, stochastic warm-starting, and graph-aligned value reuse.

This paper presents NebulaExp, a transparent ablation-driven post-training pipeline for 8B-scale LLMs, covering SFT, GRPO RL, and multi-teacher distillation. It identifies key trade-offs between mathematical reasoning and code generation, and demonstrates that data correctness filtering is the first-order optimization factor.

This paper theoretically and empirically examines adaptive patching for time-series Transformers, deriving conditions under which content-adaptive tokenization should outperform tuned uniform patching. Controlled experiments on standard benchmarks show that a well-tuned uniform baseline is competitive with dynamic patching methods, challenging the assumed benefit of adaptive approaches.

A research report detailing controlled experiments on building an external memory architecture that enables persistent AI identity independent of model weights, finding that accumulated fragment history consistently dominates system prompts in shaping output across three topologies.

This paper analyzes 935 ablation experiments from 161 publications to show that AI architectural evolution follows the same statistical laws as biological evolution, including heavy-tailed fitness effect distributions and punctuated equilibria dynamics. The findings suggest that evolutionary statistical structure is substrate-independent, determined by fitness landscape topology rather than the mechanism of selection.