2-bit quantized reasoning models exhibit process failures like loops and delayed commitment that degrade end-to-end performance, but FP16 planning and loop rescue recover accuracy on MATH-500 from 17.2% to 74.2% for Qwen3-8B while retaining speed gains.
hub
Quantization hurts reasoning? an empirical study on quantized reasoning models.arXiv preprint arXiv:2504.04823
15 Pith papers cite this work. Polarity classification is still indexing.
15
Pith papers citing it
abstract
Recent advancements in reasoning language models have demonstrated remarkable performance in complex tasks, but their extended chain-of-thought reasoning process increases inference overhead. While quantization has been widely adopted to reduce the inference cost of large language models, its impact on reasoning models remains understudied. In this paper, we conduct the first systematic study on quantized reasoning models, evaluating the open-sourced DeepSeek-R1-Distilled Qwen and LLaMA families ranging from 1.5B to 70B parameters, QwQ-32B, and Qwen3-8B. Our investigation covers weight, KV cache, and activation quantization using state-of-the-art algorithms at varying bit-widths, with extensive evaluation across mathematical (AIME, MATH-500), scientific (GPQA), and programming (LiveCodeBench) reasoning benchmarks. Our findings reveal that while lossless quantization can be achieved with W8A8 or W4A16 quantization, lower bit-widths introduce significant accuracy risks. We further identify model size, model origin, and task difficulty as critical determinants of performance. Contrary to expectations, quantized models do not exhibit increased output lengths. In addition, strategically scaling the model sizes or reasoning steps can effectively enhance the performance. All quantized models and codes are open-sourced in https://github.com/ruikangliu/Quantized-Reasoning-Models.
hub tools
citation-role summary
background 3
citation-polarity summary
roles
background 3polarities
background 3representative citing papers
Across 13 quantization setups on six LLMs and two APR benchmarks, memory falls sharply while time and energy rise, nearly half of configs are Pareto-dominated, and repaired bug sets drift despite similar counts.
DynamicPTQ uses new metrics of residual-stream dynamics to apply 8-bit activation precision only to quantization-sensitive layers in W4A4KV4 LLM inference, improving perplexity and QA performance over static smoothing baselines.
Post-training quantization increases overthinking errors in reasoning models; a logit penalty on curated overthinking markers reduces CoT length 12-23% without accuracy loss.
ScaleSearch optimizes block floating point scales via fine-grained search to cut quantization error by 27% for NVFP4, improving PTQ by up to 15 points on MATH500 for Qwen3-8B and attention PPL by 0.77 on Llama 3.1 70B.
LAQuant improves long-decoding accuracy on quantized reasoning models like Qwen3-4B by 15pp on AIME25 via layer-wise lookahead loss, achieving 3.42x speedup over FP16.
QuantClaw dynamically routes precision in agent workflows to cut cost by up to 21.4% and latency by 15.7% while keeping or improving task performance.
LLM 2-bit quantization fails via either cumulative signal degradation or early computation collapse in key components.
FluxMoE decouples MoE expert weights from persistent GPU residency via on-demand paging, achieving up to 3x throughput gains over vLLM in memory-constrained inference without accuracy loss.
A pruning technique called Reasoning-Aware Compression (RAC) jointly reconstructs input and chain-of-thought activations to preserve reasoning performance better than standard methods when compressing models like DeepSeek-R1.
Cassandra is a self-speculative decoding system that builds a draft model via fine-grained data selection and optimized pruning/mantissa truncation, achieving up to 2.41x speedup over BF16 and 1.81x more tokens than Eagle-3 on Llama 3 8B without training.
PubSwap uses a small public dataset for selective off-policy response swapping in federated RLVR to improve coordination and performance over standard baselines on math and medical reasoning tasks.
ZipMoE delivers up to 72.77% lower inference latency and 6.76x higher throughput for on-device MoE models via lossless compression and cache-affinity scheduling with a claimed provable guarantee.
citing papers explorer
-
Extreme Low-Bit Inference in Reasoning Models: Failure Modes and Targeted Recovery
2-bit quantized reasoning models exhibit process failures like loops and delayed commitment that degrade end-to-end performance, but FP16 planning and loop rescue recover accuracy on MATH-500 from 17.2% to 74.2% for Qwen3-8B while retaining speed gains.
-
Smaller Models, Unexpected Costs: Trade-offs in LLM Quantization for Automated Program Repair
Across 13 quantization setups on six LLMs and two APR benchmarks, memory falls sharply while time and energy rise, nearly half of configs are Pareto-dominated, and repaired bug sets drift despite similar counts.
-
DynamicPTQ: Mitigating Activation Quantization Collapse via Residual-Stream Dynamics
DynamicPTQ uses new metrics of residual-stream dynamics to apply 8-bit activation precision only to quantization-sensitive layers in W4A4KV4 LLM inference, improving perplexity and QA performance over static smoothing baselines.
-
Quantized Reasoning Models Think They Need to Think Longer, but They Do Not
Post-training quantization increases overthinking errors in reasoning models; a logit penalty on curated overthinking markers reduces CoT length 12-23% without accuracy loss.
-
Search Your Block Floating Point Scales!
ScaleSearch optimizes block floating point scales via fine-grained search to cut quantization error by 27% for NVFP4, improving PTQ by up to 15 points on MATH500 for Qwen3-8B and attention PPL by 0.77 on Llama 3.1 70B.
-
LAQuant: A Simple Overhead-free Large Reasoning Model Quantization by Layer-wise Lookahead Loss
LAQuant improves long-decoding accuracy on quantized reasoning models like Qwen3-4B by 15pp on AIME25 via layer-wise lookahead loss, achieving 3.42x speedup over FP16.
-
QuantClaw: Precision Where It Matters for OpenClaw
QuantClaw dynamically routes precision in agent workflows to cut cost by up to 21.4% and latency by 15.7% while keeping or improving task performance.
-
From Signal Degradation to Computation Collapse: Uncovering the Two Failure Modes of LLM Quantization
LLM 2-bit quantization fails via either cumulative signal degradation or early computation collapse in key components.
-
FluxMoE: Decoupling Expert Residency for High-Performance MoE Serving
FluxMoE decouples MoE expert weights from persistent GPU residency via on-demand paging, achieving up to 3x throughput gains over vLLM in memory-constrained inference without accuracy loss.
-
Reasoning Models Can be Accurately Pruned Via Chain-of-Thought Reconstruction
A pruning technique called Reasoning-Aware Compression (RAC) jointly reconstructs input and chain-of-thought activations to preserve reasoning performance better than standard methods when compressing models like DeepSeek-R1.
-
Cassandra: Enabling Reasoning LLMs at Edge via Self-Speculative Decoding
Cassandra is a self-speculative decoding system that builds a draft model via fine-grained data selection and optimized pruning/mantissa truncation, achieving up to 2.41x speedup over BF16 and 1.81x more tokens than Eagle-3 on Llama 3 8B without training.
-
PubSwap: Public-Data Off-Policy Coordination for Federated RLVR
PubSwap uses a small public dataset for selective off-policy response swapping in federated RLVR to improve coordination and performance over standard baselines on math and medical reasoning tasks.
-
ZipMoE: Efficient On-Device MoE Serving via Lossless Compression and Cache-Affinity Scheduling
ZipMoE delivers up to 72.77% lower inference latency and 6.76x higher throughput for on-device MoE models via lossless compression and cache-affinity scheduling with a claimed provable guarantee.
- KronQ: LLM Quantization via Kronecker-Factored Hessian
- The Periodic Table of LLM Reasoning: A Structured Survey of Reasoning Paradigms, Methods, and Failure Modes