{"paper":{"title":"Rethinking Output Alignment For 1-bit Post-Training Quantization of Large Language Models","license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","headline":"Naive output alignment fails in 1-bit LLM quantization because errors accumulate across layers and distort the representation space unevenly.","cross_cats":[],"primary_cat":"cs.LG","authors_text":"Cuong Nguyen, Cuong Pham, Dung Anh Hoang, Jianfei Cai, Thanh-Toan Do, Trung Le","submitted_at":"2025-12-25T12:39:36Z","abstract_excerpt":"Large Language Models (LLMs) deliver strong performance across a wide range of NLP tasks, but their massive sizes hinder deployment on resource-constrained devices. To reduce their computational and memory burden, various compression techniques have been proposed, including quantization, pruning, and knowledge distillation. Among these, post-training quantization (PTQ) is widely adopted for its efficiency, as it requires no retraining and only a small dataset for calibration, enabling low-cost deployment. Recent advances for post-training quantization have demonstrated that even near 4-bit met"},"claims":{"count":4,"items":[{"kind":"strongest_claim","text":"we show that this failure arises from two fundamental issues: error accumulation across layers and, more critically, anisotropic distortion of the representation space. Based on these insights, we propose a novel PTQ method for 1-bit LLMs that explicitly addresses these issues while maintaining computational efficiency.","source":"verdict.strongest_claim","status":"machine_extracted","claim_id":"C1","attestation":"unclaimed"},{"kind":"weakest_assumption","text":"The assumption that correcting error accumulation and anisotropic distortion on a small calibration set will generalize to the full test distribution without introducing new distortions or requiring architecture-specific tuning.","source":"verdict.weakest_assumption","status":"machine_extracted","claim_id":"C2","attestation":"unclaimed"},{"kind":"one_line_summary","text":"A post-training 1-bit quantization method for LLMs that fixes error accumulation and anisotropic representation distortion to outperform prior weight-driven and naive output-driven baselines.","source":"verdict.one_line_summary","status":"machine_extracted","claim_id":"C3","attestation":"unclaimed"},{"kind":"headline","text":"Naive output alignment fails in 1-bit LLM quantization because errors accumulate across layers and distort the representation space unevenly.","source":"verdict.pith_extraction.headline","status":"machine_extracted","claim_id":"C4","attestation":"unclaimed"}],"snapshot_sha256":"224ff5bc432eafc18df120726120f86c4c3cd7b46daa7e979c46ced76a7fe856"},"source":{"id":"2512.21651","kind":"arxiv","version":2},"verdict":{"id":"8b19c048-9dd2-4c3b-ae39-a8155570b87f","model_set":{"reader":"grok-4.3"},"created_at":"2026-05-16T19:27:59.724906Z","strongest_claim":"we show that this failure arises from two fundamental issues: error accumulation across layers and, more critically, anisotropic distortion of the representation space. Based on these insights, we propose a novel PTQ method for 1-bit LLMs that explicitly addresses these issues while maintaining computational efficiency.","one_line_summary":"A post-training 1-bit quantization method for LLMs that fixes error accumulation and anisotropic representation distortion to outperform prior weight-driven and naive output-driven baselines.","pipeline_version":"pith-pipeline@v0.9.0","weakest_assumption":"The assumption that correcting error accumulation and anisotropic distortion on a small calibration set will generalize to the full test distribution without introducing new distortions or requiring architecture-specific tuning.","pith_extraction_headline":"Naive output alignment fails in 1-bit LLM quantization because errors accumulate across layers and distort the representation space unevenly."},"references":{"count":22,"sample":[{"doi":"10.1609/aaai.v34i05.6239","year":2005,"title":"Language Models are Few-Shot Learners","work_id":"214732c0-2edd-44a0-af9e-28184a2b8279","ref_index":1,"cited_arxiv_id":"2005.14165","is_internal_anchor":true},{"doi":"","year":null,"title":"Stbllm: Breaking the 1-bit barrier with structured binary llms","work_id":"f9cf4e8e-ce01-40e5-820a-ae805db27e12","ref_index":2,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":2017,"title":"Network sketching: Exploiting bi- nary structure in deep cnns.2017 IEEE Conference on Computer Vision and Pattern Recog- nition (CVPR), pp","work_id":"18dc22c6-7aeb-46ba-8eab-94ca65110314","ref_index":3,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":null,"title":"Deep Compression: Compressing Deep Neural Networks with Pruning, Trained Quantization and Huffman Coding","work_id":"d79f6a26-2d92-4f7c-aea0-8aabf3b668c8","ref_index":4,"cited_arxiv_id":"1510.00149","is_internal_anchor":true},{"doi":"","year":null,"title":"Distilling the Knowledge in a Neural Network","work_id":"d927ab1f-17b8-4002-9d09-c3d55764fbad","ref_index":5,"cited_arxiv_id":"1503.02531","is_internal_anchor":true}],"resolved_work":22,"snapshot_sha256":"1b2bbe69b439b266218239681008d63b393781439a83862d6e51b0eaac1b81b6","internal_anchors":10},"formal_canon":{"evidence_count":2,"snapshot_sha256":"0fa99115169d6e8517e200daa4364b1dcbb699592f7b4bc95a067dd42541b33e"},"author_claims":{"count":0,"strong_count":0,"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"builder_version":"pith-number-builder-2026-05-17-v1"}