Research map / Architectures and efficiency / Quantization
RaBiT: Residual Aware Binarization Training for Accurate and Efficient LLMs
Youngcheon You, Banseok Lee, Minseop Choi, Seonyoung Kim, Hyochan Chong, Changdong Kim, Youngmin Kim, Dongkyu Kim
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use coupled QAT (shared FP weight, sequential residual paths) instead of standard QAT (ablation) | RaBiT 2-bit · WikiText-2, Llama2-7B | perplexity: 6.55 -> 5.84 | Improved |
| add iterative residual SVID and importance-scaled preconditioning to coupled QAT (ablation) | RaBiT · WikiText-2, Llama2-7B | perplexity: 5.84 -> 5.78 | Improved |
| add iterative SVID and preconditioning to standard QAT (ablation) | Standard QAT · WikiText-2, Llama2-7B | perplexity: 6.55 -> 6.18 | Improved |
| replace greedy SVID with iterative residual SVID init (ablation) | RaBiT init · Initial KL loss, Llama2-7B q_proj | KL loss: 17,152 -> 13,760 | Improved |
| add I/O channel importance-scaled preconditioning to init (ablation) | RaBiT init · Initial KL loss, Llama2-7B q_proj | KL loss: 13,760 -> 2,672 | Improved |
| add I/O channel importance scaling (weight reconstruction view) (ablation) | RaBiT init · Weight reconstruction, Llama2-7B q_proj | Avg MSE: 0.122 -> 0.302 | Hurt |
| use coupled residual training vs standard QAT on path behavior (ablation) | RaBiT · Llama2-7B layer 5 residual alignment | Corr(R1,y2): 0.4395 -> 0.6484 | Improved |
| freeze primary binary core (MBOK-style) instead of coupled training (ablation) | RaBiT variant · Llama2-7B training dynamics | inter-path correlation / KL loss | Hurt |
| freeze scales or binary cores in coupled training (ablation) | RaBiT variants (Scale-only, Scale-frozen) · Llama2-7B training loss | KL loss | Hurt |
| use RaBiT instead of QTIP at 2 bits | RaBiT vs QTIP · WikiText-2, Llama2-7B | perplexity: 5.86 -> 5.78 | Improved |
| use RaBiT instead of QTIP at 2 bits | RaBiT vs QTIP · WikiText-2, Llama2-13B | perplexity: 5.11 -> 5.15 | Hurt |
| use RaBiT instead of QTIP on hard benchmarks | RaBiT vs QTIP · Challenging benchmarks avg, Llama2-13B | accuracy: 25.38 -> 27.14 | Improved |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.
- QuIP# — QuIP: 2-Bit Quantization of Large Language Models With Guarantees (NeurIPS 2023) · better on 1 of 1 shared result
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- BiLLM: Pushing the Limit of Post-Training Quantization for LLMs (ICML 2024 · similarity 0.90)
- ARB-LLM: Alternating Refined Binarizations for Large Language Models (ICLR 2025 · similarity 0.89)
- PB-LLM: Partially Binarized Large Language Models (ICLR 2024 · similarity 0.89)
- HESTIA: A Hessian-Guided Differentiable Quantization-Aware Training Framework for Extremely Low-Bit LLMs (NeurIPS 2026 · similarity 0.88)
- LittleBit: Ultra Low-Bit Quantization via Latent Factorization (NeurIPS 2025 · similarity 0.88)
- ResQ: Mixed-Precision Quantization of Large Language Models with Low-Rank Residuals (ICML 2025 · similarity 0.88)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
