News without the noise
Benchmark
Reported in 102 papers on Zotpaper. The latest, each as the paper states it:
16 results for “GSM8K”
Research · Efficiency & systems
Integer-Exact Microscaling: Running MXFP4 Language Models on INT8 Matrix Units
The authors show MXFP4 weights can be folded into a single INT8 matrix multiply when their block exponents span at most three, a condition met by over 97.9% of rows in real models.
GSM8K: 94.8% accuracy (as reported)
UT Southwestern Medical Center, Emory University, Duke UniversityCode
Research · Efficiency & systems
Rethinking LLM Pruning via Test-Time Scaling
Experiments on math and coding benchmarks show that structured pruning preserves the range of solvable problems even as single-attempt success declines, but obtaining those solutions requires more sampling effort.
four tasks: 75.8% accuracy retention (as reported)
Tencent
Research · Efficiency & systems
Tetra: Serving Leech-Lattice Quantized LLMs at 2.7 Bits per Parameter
A fused GPU kernel decodes compact Leech-lattice weights during multiplication, reducing memory traffic while retaining usable accuracy on Qwen3 models.
GSM8K: 82.49% accuracy (as reported)
Research · Efficiency & systems
Distributed Subliminal Learning: Replacing Model Updates with Random-Carrier Outputs
The authors distill task-unrelated model outputs to combine specialized language models and train federated classifiers with substantially less uplink communication.
GSM8K: 22 pts accuracy gain over base model (as reported)
Università della Svizzera italiana (USI), MIT Media Lab, Massachusetts Institute of Technology
Research · Efficiency & systems
DuoRate: Two-Rate Per-Expert Precision Allocation for Mixture-of-Experts Serving
DuoRate stores weights once and adapts read precision per workload, matching 4-bit accuracy at 2.9 bpw
GSM8K: 88.25% accuracy (as reported)
Huawei Technologies Ltd.
Research · Evaluation & benchmarks
Predictable performance is suboptimal, and what it means for model routing
Zhang and Hardt connect prediction gains beyond a model’s own confidence to its room for improvement, then test the relationship across 16 language models.
GSM8K: 0.93 Pearson correlation between empirical suboptimality and (as reported)
Research · Efficiency & systems
ValuSpec: Plug-and-Play Candidate Valuation before Target Verification for Tree-Based Speculative Decoding
The plug-and-play module uses transfer-based supervision to align filtering with the target model without training.
GSM8K: 98.25 tokens/s throughput (as reported)
Tongji University, Nanyang Technological University, Marquette University
Research · Efficiency & systems
Is 2:4 Sparsity Still Worth It? Semi-Structured Sparsity versus Quantization in LLM Inference
Across three NVIDIA GPU generations, Qin finds that 2:4 sparsity becomes useful chiefly when no lower dense precision remains available.
GSM8K: 0.31 accuracy (as reported)
Code
Research · Language models
PV-Decoding: Progressive Verification Decoding with Decode Gating for Diffusion Language Models
PV-Decoding uses confidence thresholds and progressive verification to reduce model forward passes by 8–37% while keeping accuracy drop under 1.3%.
GSM8K: 68.5 MFP (as reported)
Huawei Technologies, City University of Hong Kong
Research · Efficiency & systems
When Does Better Acceptance Pay? A Controlled Study of Speculative Decoding
A controlled study across six drafter recipes and ten benchmarks shows that aligning training and inference conditions narrows method gaps, and acceptance improvements only translate to higher throughput under specific task uncertainty and execution contexts.
GSM8K: 5.44 tokens/s speedup (as reported)
Tianjin University of Technology, Baidu Inc., Shandong University
Research · Language models
Prism Attention: Proposal-Refined Index Sharing Mechanism for Efficient LLMs Inference
Prism Attention replaces static cross-layer index sharing with dynamic refinement, achieving 3.4x speedup over FlashAttention at 128K tokens.
128K context: 3.36 x end-to-end speedup over FlashAttention (as reported)
Soochow University, Baidu Inc.
Research · Reinforcement learning
ReFlow: Replay-Driven Decoupling for Efficient RL Post-Training
ReFlow, a decoupled and adaptive system, achieves up to 2.06× throughput improvement over existing methods while maintaining model quality.
DAPO-Math-17K: 0.2% rollout worker idle ratio (as reported)
The University of Hong Kong, University of Science and Technology of China, Alibaba Group
Research · Efficiency & systems
Few Bits, One Law: Toward W2A4KV2
CanonQ uses fixed transforms and reusable codebooks, then jointly trains models to tolerate quantized weights, activations and attention caches.
GSM8K: 47.8% final-answer exact match (as reported)
Meta AI
Research · Efficiency & systems
DScale: Scaling Block-Diffusion Speculative Decoding with Adaptive Verification
DScale reallocates limited verification capacity across requests while retaining useful draft tokens and fixed-shape GPU graphs.
GSM8K, MATH500, HumanEval, MBPP: 43.9% geometric-mean throughput gain vs DFlash (as reported)
Shenzhen Institutes of Advanced Technology, Chinese Academy of Sciences, University of Chinese Academy of Sciences, Victoria University of Wellington
Research · Efficiency & systems
A Tilted Bowl Is Not a Slippery Slope: Compressing Looped Models
Kolawole et al. show that rounding error in looped models shifts the fixed point rather than accumulating, enabling predictable failure and recovery with few high-precision steps.
GSM8K: 40% accuracy (as reported)
Carnegie Mellon University, ML Collective
Research · Agents
Self-Compacting Language Model Agents
SelfCompact scaffold pairs a compaction tool with a lightweight rubric, enabling adaptive context summarization without fine-tuning.
competition math (IMO-Answerbench, HMMT Nov 25/Feb 26): 18.1% accuracy improvement over no-summarization baseline (as reported)
Johns Hopkins University, AppleCode