A new arXiv paper (2610.09186v1) argues that step-by-step reasoning and pattern recognition in large language models lie on a spectrum defined by how much preceding context each next token depends on,…
A new arXiv paper (2610.09250v1) proposes a sample-only framework for evaluating and selecting Best-of-N (BoN) inference-time alignment policies without access to response likelihoods, which standard …
A developer built OracleBench, a benchmark that grades two small open LLM judges (Qwen2.5-3B and Qwen2.5-0.5B) against deterministic oracles — the FAR/DFARS clause registry, GSM8K arithmetic, and 25 r…
Condition-anchored generative distillation (CAGD) reduced four-task final held-out loss in a 219M masked diffusion language model from 2.927 to 1.114 in one task order and from 2.168 to 0.891 in exact…
DART-ES, a new Evolution Strategies fine-tuning method from researchers publishing on arXiv as 2610.06993v1, improves average accuracy from 72.07% to 73.53% over standard ES and exceeds GRPO's 73.26% …
Researchers introduced Sigma, described as the first large-scale continuous diffusion language model at 3B and 8B parameters, built on steerable low-dimensional ODE/SDE latent trajectories and trained…
Training Qwen3-4B with counterexample-only supervised fine-tuning collapsed true-theorem recognition from 0.27 to 0.00, while reinforcement learning with a sparse outcome-only reward repaired the coll…
A working implementation of reinforcement learning for calibrated decisions (RLCD) for reasoning models, OpenJev-RLCD, matches or beats supervised fine-tuning, RFT/STaR and GRPO in accuracy and beats …
GoldiMask, a fine-tuning method for discrete diffusion language models introduced in arXiv paper 2609.38385v1, selects which response tokens to reveal as context by approximately maximizing a submodul…
A neurosymbolic router that learns a deterministic finite automaton via the L* grammatical inference algorithm achieved 100% routing accuracy and 98.3% overall accuracy on 100 untested prompts from De…
A sweep of eight agent orchestration architectures across five instruction-tuned 7-9B models and six benchmarks found that scaling a team from three to thirty calls raises accuracy by up to 17 points …
A February 14, 2026 arXiv paper (2602.13639) proposes an Entropy-Based Adaptive Guidance Framework to fix cognitive mismatching in heterogeneous LLM multi-agent systems, after experiments showed stron…
ELF-REG, a method that scales Embedded Language Flows (ELF) continuous diffusion language models to reasoning tasks, reached 55.96% pass@1 on GSM8K at 64 network function evaluations and 13.39% on MAT…
Attention-Aware Routing (AAR), a method that augments a Mixture-of-Experts router with temporal and spectral features from a sliding window of attention weights, improved GSM8K accuracy by 3.37 percen…
A new arXiv paper (2609.19499v1) reports that generation schedule, not candidate count alone, determines the energy and latency cost of LLM test-time scaling: on A100 GPUs, eight serial generation cal…
Applying reinforcement learning with verifiable rewards to the GLM-4-Voice speech model raised free-form accuracy on the GSM8K math benchmark to 74.8%, a new state of the art for speech-native models,…
A study measuring 54 configurations of Qwen2.5-7B-Instruct on vLLM 0.12 across L4, A100, and H100 GPUs found that 18 of 36 configurations reach the cost, quality, and latency Pareto frontier, with com…
A training-free, CPU-only lexical prompt-compression pipeline reduced prompt tokens by a mean of 40.3% (sigma = 9.2) while holding BERTScore-F1 at 0.876 against original-prompt output, according to an…
A developer outlined best practices for evaluating AI models, emphasizing a multi-layered framework combining standardized benchmarks like MMLU and HumanEval, adversarial red teaming for prompt inject…
A Rust-based tool called Spanda computes LLM epistemic uncertainty in 652.1 nanoseconds (0.65 µs) per evaluation, roughly 90,000× faster than neural Semantic Entropy methods that add about 92.4 ms of …