Research map / Large language models / Retrieval and knowledge

SLIDERS: Systematic Reviews via Automated Evidence Synthesis and Reconciliation

NeurIPS 2026 · Large language models › Retrieval and knowledge · approach: Domain-specific QA datasets · 0 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Introduces SLIDERS, an LLM pipeline that builds a reusable, typed evidence table from a document corpus via evidence-variable generation, chunk-level extraction, and a SQL-based reconciliation agent, then answers questions and follow-ups from that table.

Problem

Automating systematic-review-style evidence synthesis over large document corpora, where retrieval-first RAG misses distributed evidence and long-context prompting fails to scale or stay accurate.

Approach

Replace passage retrieval with a persistent question-specific evidence table: LLM-generated typed evidence variables, relevance-gated chunk extraction with source spans, primary-key grouping with canonicalization, and a coding agent that reconciles records via SQL before answer synthesis and follow-up questions.

Key results

• On BrowseComp+, SLIDERS reaches 90.00, about 10 points above the strongest baseline RLM (GPT-5).
• On WikiCeleb100 SLIDERS scores 78.91 vs 59.80 for RLM; on FinQ100 it scores 60.18 with Qwen3.5-122B-A10B vs 28.87 for the best completing baseline.
• On within-window benchmarks (FinanceBench, Loong, Oolong), SLIDERS with GPT-4.1 averages 77.64, versus 68.10 for the GPT-4.1 base LM.

Surprising findings

• A direct-table variant without reconciliation reaches 87.33 on BrowseComp+, only 2.67 points below the full system, when the table fits in context.
• On Loong, the Direct Table variant slightly exceeds the full system by 0.08 points.

Limitations

• Follow-up questions are assumed answerable from the evidence variables induced for the original question; missing variables cause failure.
• Exact counting and numerical questions remain hard, and reconciliation over the table can introduce reasoning errors (8 of 15 BrowseComp+ failures).
• Overhead is unnecessary for short contexts, such as Chinese classification subsets under 16K tokens.

Experimental setup

Datasets: BrowseComp+ FinanceBench Loong Oolong WikiCeleb100 FinQ100

Benchmarks: BrowseComp+ WikiCeleb100 FinQ100 FinanceBench Loong Oolong

Metrics: accuracy partial-credit score

Baselines: BM25 LongRAG GraphRAG DocETL RLM Chain of Agent BaseLM

Builds on: RAG DocETL Recursive Language Models Chain of Agents DocLing

Closest papers by meaning

• Evaluating AI-based Scientific Knowledge Synthesis with Epidemiological Systematic Reviews (NeurIPS 2026 · similarity 0.85)
• Format Matters: The Robustness of Multimodal LLMs in Reviewing Evidence from Tables and Charts (AAAI 2026 · similarity 0.83)
• CycleResearcher: Improving Automated Research via Automated Review (ICLR 2025 · similarity 0.83)
• EHRNoteQA: An LLM Benchmark for Real-World Clinical Practice Using Discharge Summaries (NeurIPS 2024 · similarity 0.83)
• PRISMM-Bench: A Benchmark of Peer-Review Grounded Multimodal Inconsistencies (ICLR 2026 · similarity 0.83)
• BooookScore: A systematic exploration of book-length summarization in the era of LLMs (ICLR 2024 · similarity 0.82)