DISCO: Distributed Long Context Scaling with Grounding-Reasoning Disaggregation

Hugging Face Daily Papers Papers
long-context llm distributed-ai reasoning scaling cost-efficiency context-rot

Summary

DISCO is a distributed method for long-context LLM inference that separates grounding and reasoning to prevent context rot, achieving high accuracy on benchmarks while reducing inference costs by over 80%.

While Large Language Models (LLMs) advertise million-token context windows, reasoning quality often collapses as inputs grow -- a phenomenon termed context rot. This failure stems from a structural entanglement in monolithic architectures, where the massive search burden of contextual grounding exhausts the representational capacity needed for complex reasoning. To resolve this, we propose Grounding-Reasoning Disaggregation via DIStributed long COntext scaling (DISCO). Inspired by distributed computing frameworks like Apache Spark, DISCO partitions long context across a fleet of Worker LLMs dedicated exclusively to parallel, localized grounding. A central Driver LLM, trained via Reinforcement Learning (GRPO) to optimize planning, orchestrates execution by dynamically mapping queries into atomic extraction tasks and reducing the gathered evidence to synthesize a final answer. By isolating reasoning from raw context noise, DISCO effectively eliminates context rot. On RULER-QA (1M tokens), it maintains 78.4% accuracy where standard baselines collapse. Furthermore, it outperforms full-context models by up to 9.8 points on LongBench v2 and matches frontier models like Gemini-3-Pro-Preview while reducing inference costs by over 80%, establishing a highly efficient paradigm for robust long-context inference.
Original Article
View Cached Full Text

Cached at: 09/29/26, 04:12 PM

Paper page - DISCO: Distributed Long Context Scaling with Grounding-Reasoning Disaggregation

Source: https://huggingface.co/papers/2609.33485

Abstract

WhileLargeLanguageModels(LLMs)advertisemillion-tokencontextwindows,reasoningqualityoftencollapsesasinputsgrow--aphenomenontermedcontextrot.Thisfailurestemsfromastructuralentanglementinmonolithicarchitectures,wherethemassivesearchburdenofcontextualgroundingexhauststherepresentationalcapacityneededforcomplexreasoning.Toresolvethis,weproposeGrounding-ReasoningDisaggregationviaDIStributedlongCOntextscaling(DISCO).InspiredbydistributedcomputingframeworkslikeApacheSpark,DISCOpartitionslongcontextacrossafleetofWorkerLLMsdedicatedexclusivelytoparallel,localizedgrounding.AcentralDriverLLM,trainedviaReinforcementLearning(GRPO)tooptimizeplanning,orchestratesexecutionbydynamicallymappingqueriesintoatomicextractiontasksandreducingthegatheredevidencetosynthesizeafinalanswer.Byisolatingreasoningfromrawcontextnoise,DISCOeffectivelyeliminatescontextrot.OnRULER-QA(1Mtokens),itmaintains78.4%accuracywherestandardbaselinescollapse.Furthermore,itoutperformsfull-contextmodelsbyupto9.8pointsonLongBenchv2andmatchesfrontiermodelslikeGemini-3-Pro-Previewwhilereducinginferencecostsbyover80%,establishingahighlyefficientparadigmforrobustlong-contextinference.

View arXiv pageView PDFAdd to collection

Get this paper in your agent:

hf papers read 2609\.33485

Don’t have the latest CLI?curl \-LsSf https://hf\.co/cli/install\.sh \| bash

Models citing this paper0

No model linking this paper

Cite arxiv.org/abs/2609.33485 in a model README.md to link it from this page.

Datasets citing this paper0

No dataset linking this paper

Cite arxiv.org/abs/2609.33485 in a dataset README.md to link it from this page.

Spaces citing this paper0

No Space linking this paper

Cite arxiv.org/abs/2609.33485 in a Space README.md to link it from this page.

Collections including this paper0

No Collection including this paper

Add this paper to acollectionto link it from this page.

Similar Articles

Autonomous Scientific Discovery via Iterative Meta-Reflection

Hugging Face Daily Papers

DiscoPER is an autonomous framework leveraging large language models and dynamic code generation for open-ended scientific research, using second-order meta-reflection to synthesize discoveries and statistical testing for rigor. Evaluated on a multimodal ecological benchmark, it outperforms baselines in recovering known patterns.

Scaling Scientific Discovery Environments for Turn-Level Agentic RL

arXiv cs.AI

This paper introduces SciDisco, a scalable framework for training scientific discovery agents via process-verifiable environments, DAG-grounded trajectory synthesis, and turn-level reinforcement learning. The proposed SciDisco-14B model achieves state-of-the-art on hypothesis-driven scientific data analysis benchmarks.

Disco-RAG: Discourse-Aware Retrieval-Augmented Generation

arXiv cs.CL

Disco-RAG proposes a discourse-aware retrieval-augmented generation framework that integrates discourse signals through intra-chunk discourse trees and inter-chunk rhetorical graphs to improve knowledge synthesis in LLMs. The method achieves state-of-the-art results on QA and summarization benchmarks without fine-tuning.