{"paper":{"title":"On-the-Fly Adaptation to Quantization: Configuration-Aware LoRA for Efficient Fine-Tuning of Quantized LLMs","license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","headline":"A single configuration-aware model generates effective LoRA adjustments for any quantization setting of an LLM without retraining per configuration.","cross_cats":["cs.AI"],"primary_cat":"cs.LG","authors_text":"Edith C. H. Ngai, Ming Tang, Rongguang Ye","submitted_at":"2025-09-22T11:07:50Z","abstract_excerpt":"As increasingly large pre-trained models are released, deploying them on edge devices for privacy-preserving applications requires effective compression. Recent works combine quantization with the fine-tuning of high-precision LoRA adapters, which can substantially reduce model size while mitigating the accuracy loss from quantization. However, edge devices have inherently heterogeneous capabilities, while performing configuration-wise fine-tuning for every quantization setting is computationally prohibitive. In this paper, we propose CoA-LoRA, a method that dynamically adjusts the LoRA adapte"},"claims":{"count":4,"items":[{"kind":"strongest_claim","text":"CoA-LoRA dynamically adjusts the LoRA adapter to arbitrary quantization configurations without requiring repeated fine-tuning and achieves comparable or superior performance to state-of-the-art methods that fine-tune a separate LoRA adapter for each configuration.","source":"verdict.strongest_claim","status":"machine_extracted","claim_id":"C1","attestation":"unclaimed"},{"kind":"weakest_assumption","text":"The configuration-aware model can accurately predict low-rank adjustments for unseen quantization configurations when trained only on a Pareto-selected subset of configurations that cover different total bit-width budgets.","source":"verdict.weakest_assumption","status":"machine_extracted","claim_id":"C2","attestation":"unclaimed"},{"kind":"one_line_summary","text":"CoA-LoRA trains a single configuration-aware model on a Pareto-optimized set of quantization configurations to enable dynamic LoRA adaptation to arbitrary bit-width assignments without per-configuration fine-tuning.","source":"verdict.one_line_summary","status":"machine_extracted","claim_id":"C3","attestation":"unclaimed"},{"kind":"headline","text":"A single configuration-aware model generates effective LoRA adjustments for any quantization setting of an LLM without retraining per configuration.","source":"verdict.pith_extraction.headline","status":"machine_extracted","claim_id":"C4","attestation":"unclaimed"}],"snapshot_sha256":"feac72112f5f4f2b596bfb1a9ff5e67f9fa16e4eb4d458372db1fcb8c227d842"},"source":{"id":"2509.25214","kind":"arxiv","version":4},"verdict":{"id":"005bcfd4-597f-4294-9493-a6a2f653299c","model_set":{"reader":"grok-4.3"},"created_at":"2026-05-18T13:49:52.475415Z","strongest_claim":"CoA-LoRA dynamically adjusts the LoRA adapter to arbitrary quantization configurations without requiring repeated fine-tuning and achieves comparable or superior performance to state-of-the-art methods that fine-tune a separate LoRA adapter for each configuration.","one_line_summary":"CoA-LoRA trains a single configuration-aware model on a Pareto-optimized set of quantization configurations to enable dynamic LoRA adaptation to arbitrary bit-width assignments without per-configuration fine-tuning.","pipeline_version":"pith-pipeline@v0.9.0","weakest_assumption":"The configuration-aware model can accurately predict low-rank adjustments for unseen quantization configurations when trained only on a Pareto-selected subset of configurations that cover different total bit-width budgets.","pith_extraction_headline":"A single configuration-aware model generates effective LoRA adjustments for any quantization setting of an LLM without retraining per configuration."},"integrity":{"clean":true,"summary":{"advisory":0,"critical":0,"by_detector":{},"informational":0},"endpoint":"/pith/2509.25214/integrity.json","findings":[],"available":true,"detectors_run":[],"snapshot_sha256":"c28c3603d3b5d939e8dc4c7e95fa8dfce3d595e45f758748cecf8e644a296938"},"references":{"count":0,"sample":[],"resolved_work":0,"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57","internal_anchors":0},"formal_canon":{"evidence_count":2,"snapshot_sha256":"d5a6c1ab72ef4a06457ff43482d5ad8e5646e8a7a417f930a2881632d3ba2ca1"},"author_claims":{"count":0,"strong_count":0,"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"builder_version":"pith-number-builder-2026-05-17-v1"}