News without the noise
Benchmark
Reported in 102 papers on Zotpaper. The latest, each as the paper states it:
66 results for “GSM8K” · page 1 of 4
Research · Language models
Sep-Latent: Separator Tokens as Natural Compression Anchors for Implicit Reasoning
Sep-Latent concentrates reasoning information into punctuation-like anchors, distills explicit chains into latent states, and learns when to stop internal computation.
GSM8k-Aug: 53.5% accuracy (as reported)
Zhejiang University, Alibaba Group
Research · Language models
HiLo-RD: Cross-Budget Self-Distillation for Masked Diffusion Language Models
HiLo-RD transfers a model’s own large-window representations into small-window decoding, improving GSM8K accuracy while reducing readable output length.
GSM8K: 78.24% accuracy (as reported)
The Hong Kong University of Science and Technology (Guangzhou), Zhejiang University, Nanyang Technological University
Research · Language models
Clock Diffusion: Efficient Semi-Autoregressive Continuous Diffusion Language Models
The authors' Clock Diffusion framework uses position-dependent noise schedules to produce state-of-the-art continuous diffusion language models on OpenWebText and GSM8K.
GSM8K: 67.2% pass@1 accuracy (as reported)
NVIDIA, Cornell University
Research · Language models
Hinted Scaling Laws
The authors report that using assisted answers from smaller models reduced median error in forecasts of larger models’ unaided accuracy from 21 to 4.3 percentage points.
GSM8K: 25.9% accuracy (as reported)
Stanford University
Research · Language models
Loopholing All the Way: Language Modeling with pure Self-Conditioning
LAW propagates the expected embedding across denoising steps and decodes tokens once at the end, matching or surpassing discrete and continuous diffusion baselines.
GSM8K: 26.4% accuracy (as reported)
University of Basel, Microsoft AI, ELLIS Institute Tübingen
Research · Theory
Does Random Search Discover Task Experts? Revisiting Neural Thickets
The apparent advantage of best-of-N selected perturbations mostly reflects selection noise, and top experts perform no better than the base model on held-out questions.
Countdown, GSM8K, MBPP: 0.16 pts mean held-out rank-1 accuracy gain (as reported)
University of Southern California, University of Oxford
Research · Safety & security
TRACER: Trace-Regularized Training Against Indirect Prompt Injection
TRACER training reduces attack success from 65.5% to 6.1% on RAG-framed attacks while preserving 78–82% clean accuracy
GSM8K: 6.1% ASR (as reported)
Mohamed bin Zayed University of Artificial Intelligence
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 · Language models
Beyond Data Schedule: Fast LLM Pretraining on Data Mixtures by Loss-Reweighting
A new method, LARE, automatically adjusts domain loss weights based on running statistics, yielding gains of up to 17 points on math benchmarks.
GSM8K: 17.06 pts gain (as reported)
Research · Evaluation & benchmarks
Models Verify Sufficiency, Not Validity: Usable Falsehoods Evade the Checks that Gaps Trigger
Controlled tests found that making an unsupported premise usable reduced models’ warnings, although accepting question-supplied premises can also reflect conditional reasoning.
GSM8K: 0.67 flagged unanswerable (as reported)
Research · Language models
LoOPSD: On-Policy Self-Distillation with Looped Teachers
LoOPSD searches for repeated layer blocks, then distills the looped model's token-level predictions back into the original student, improving reasoning accuracy across model families.
GSM8K: 299.98 tokens average total generated tokens (as reported)
Anhui Province Key Laboratory of Digital Security, Token Foundry, Alibaba Group, Independent Researcher
Research · Language models
How Depth Growth and Looping Shape In-Context Learning in Language Models
Across language models and controlled tasks, the authors find that growing or looping layers strengthens context-based learning and increases reliance on particular attention mechanisms.
GSM8K: 7 pts Accuracy (as reported)
Technical University of Munich, Helmholtz AI, Munich Center for Machine Learning (MCML)
Research · Reinforcement learning
RLVR Sharpens the Pencil, Not the Mind
Tracking one small model through arithmetic training, the authors find modest accuracy gains but no measurable improvement in internal signals distinguishing flawed reasoning.
GSM8K: 61.3% pass@1 (as reported)
Indian Institute of Technology Roorkee
Research · Language models
Where RL Breaks: Localizing and Repairing Failures via Targeted Continued Pretraining
Diagnostic phrases in failed reasoning traces retrieved a compact training corpus that improved subsequent reinforcement learning across model families.
GSM8K: 60.6% greedy accuracy (as reported)
Princeton Language and Intelligence
Research · Reinforcement learning
Think Less, Locally: Local Preference Learning for Reasoning Compression
A new method teaches large language models when to conclude reasoning rather than continue, cutting output length by up to 45% on math and 36% on code without losing accuracy.
AIME24: 46.82% AUC (as reported)
Xi’an Jiaotong University, Tencent Inc.
Research · Language models
Compile as Hypothesis: A Headroom Typology for Prompt Optimization on Real Reasoning Models
Automated prompt compilers show no statistically significant accuracy improvements across four optimizer families and eight tasks, but structured genomes prevent catastrophic search failures.
AIME 2024+25: 60 problems correct (as reported)
Hunan Agricultural University
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: 48.4% coverage retention (as reported)
Tencent
Research · Language models
Soft Teacher Forcing
Soft Teacher Forcing mixes ground-truth embeddings with model-derived continuous tokens, improving results across four mathematics benchmarks and seven language models.
GSM8K: 52.66% accuracy (as reported)
Google Deepmind
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: 92.04% accuracy (as reported)
Research · Language models
REWIND: Uncertainty-Guided Trajectory Repair for Test-Time Reasoning
Ding et al. introduce REWIND, a training-free method that uses token entropy dynamics to select optimal points for regenerating suffixes, improving reasoning accuracy across multiple models and benchmarks.
GSM8K: 90.9% accuracy (as reported)
The Hong Kong University of Science and Technology (Guangzhou), University of Edinburgh, Huawei