Research map / Large language models / Retrieval and knowledge

HYVE: Hybrid Views for LLM Context Engineering over Machine Data

NeurIPS 2026 · Large language models › Retrieval and knowledge · approach: Knowledge-augmented question answering · 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.

Proposes HYVE, a database-inspired preprocessing/postprocessing framework that converts machine-data JSON/AST payloads into hybrid column/row views backed by a request-scoped DuckDB datastore, cutting LLM token use by 50-90% while maintaining or improving quality.

Problem

LLMs are brittle on long, deeply nested machine-data payloads (logs, telemetry, configs) embedded in prompts, suffering token explosion, context rot, and weak numeric/sequence reasoning.

Approach

Wrap a single LLM call with a preprocessor that parses embedded JSON/Python literals, builds a hybrid column/row table view with BM25-based truncation, stores full data in an in-memory datastore, and a postprocessor that either returns output, backfills hidden data into a template via SQL, or runs one SQL-augmented synthesis call.

Key results

• On TOON-QA, GPT-4.1 + HYVE raises exact-match accuracy from 0.47 to 0.93 while cutting tokens from 1.3M to 153.8K (-88%).
• On Line chart generation with GPT-5, similarity rises from 0.68 to 0.97 and tokens fall from 4.2M to 0.75M (-82%).
• Disabling postprocessing drops Bar chart similarity from 0.99 to 0.06 for GPT-4.1, showing backfilling is essential for chart outputs.

Surprising findings

• Disabling truncation degrades Bar chart and Hard reasoning quality, suggesting overly long contexts induce context rot.
• TOON encoding, though more token-efficient, degraded structured-generation quality, attributed to JSON-heavy training data.

Limitations

• Most evaluation datasets are proprietary, limiting reproducibility.
• Mode 3 SQL-augmented synthesis adds an LLM call that can increase latency on fast baselines (e.g., GPT-4.1 on TOON-QA and RB-Text).
• Relies on fixed thresholds rather than an explicit token budget; ranking and truncation parameters are heuristic.

Experimental setup

Datasets: Cert-QA Runbook Line Bar Anom Sum Canvas RB-Text RB-JSON Hard TOON-QA

Benchmarks: Cert-QA Runbook Line Bar Anom Sum Canvas RB-Text RB-JSON Hard TOON-QA

Metrics: GenericJudge score similarity (nDTW/cosine/Pearson) ReasoningJudge score ExactMatch total token usage latency

Baselines: GPT-4.1 GPT-5 TOON encoding LLMLingua

Builds on: BM25 TOON JSONPath DuckDB

Closest papers by meaning

• CRAFT: Customizing LLMs by Creating and Retrieving from Specialized Toolsets (ICLR 2024 · similarity 0.87)
• ChatQA 2: Bridging the Gap to Proprietary LLMs in Long Context and RAG Capabilities (ICLR 2025 · similarity 0.86)
• LooGLE v2: Are LLMs Ready for Real World Long Dependency Challenges? (NeurIPS 2025 · similarity 0.86)
• Scaling Instruction-tuned LLMs to Million-token Contexts via Hierarchical Synthetic Data Generation (ICLR 2025 · similarity 0.85)
• Large Language Models as Tool Makers (ICLR 2024 · similarity 0.85)
• WorkflowLLM: Enhancing Workflow Orchestration Capability of Large Language Models (ICLR 2025 · similarity 0.85)