Search

What are you looking for?

60 results for nn

AI·LLM-as-judge scores fluctuate across evaluators even with fixed candidate responses, treating the issue as measurement validity. Across four datasets, scaling Qwen3 from 1.7B to 32B parameters yields only limited adjacent gains, while MiniMax M2-to-M2.7 API upgrades show none; stronger judges reduce but do not eliminate position and verbosity biases, with repeated juries offering little benefit under correlated errors.

arXiv cs.CL·arxiv.org··1.9Researchllm-as-judgeevaluation-reliabilitybias

AI·This position paper reviews LLM-driven formal theorem provers and argues that current systems function mainly as solvers for well-defined problems, not as research agents capable of discovering new theorems or resolving open conjectures at the frontier. It identifies key limitations in datasets, exploration, tools, and collaboration and proposes a roadmap for AI4Math systems to support genuine mathematical research.

arXiv cs.CL·arxiv.org··1.7Researchllmformal-methodsai4math

AI·Hidden Decoding expands each token into n independent streams during continued pretraining of a fixed Transformer backbone, using separate embedding tables and retaining intermediate KV caches for latent computation; Stream-Factorized Attention limits cross-stream mixing to reduce costs. Experiments show frontier 80B and 617B MoE models improve on all benchmarks over matched baselines.

arXiv cs.CL·arxiv.org··1.7Researchllmscalingcontinued-pretraining

AI·LLM-based ASR in regulated domains like banking is limited by privacy and real-speech collection costs; synthetic TTS data is a cost-effective substitute, but acoustic mismatch hinders supervised fine-tuning (SFT). Group Relative Policy Optimization (GRPO) applied solely to synthetic speech reduces WER by 40% relative to SFT (36.71% to 22.09%) and by 45% in the SFT-then-GRPO sequence, by improving behavioral calibration and audio attention rather than representations.

arXiv cs.CL·arxiv.org··1.7Researchspeech-recognitionasrllm

AI·Across five instruction-tuned LLMs, ordinary answer-confidence separates correct from wrong answers on answerable questions but fails to distinguish unanswerable questions (e.g., false-premise ones in CREPE), while a linear probe on hidden states performs the opposite, revealing two distinct abstention axes that remain separable even at 14B scale.

arXiv cs.CL·arxiv.org··1.7Researchllmabstentionllm-safety

AI·DominoTree is a training-free best-first tree draft builder for speculative decoding that scores each candidate node by re-applying Domino's GRU-based causal correction along its specific root-to-node path. It restricts per-node correction to top-M candidates for efficiency and delivers up to 6.6x speedup with the highest mean accept length of any tested method.

arXiv cs.CL·arxiv.org··1.7Researchspeculative-decodingdraftingllm-inference

AI·ICDAR 2026 HIPE-OCRepair competition evaluated LLM-assisted post-correction of noisy OCR from 17th-20th century multilingual (EN/FR/DE) historical newspapers and books. Four teams used zero-shot to fine-tuning approaches; results show significant error reduction but recurring over-correction on low-noise inputs, with a public dataset and evaluation framework released.

arXiv cs.CL·arxiv.org··1.7Researchllmocrpost-correction

AI·XALPHA is a memory-driven AI quant researcher that uses multi-source research memory integrating external financial reports and prior discovery feedback, with Macro Brain for theme planning and archetype selection, Micro Brain for hypothesis-to-code translation and tri-alignment verification, and Cross Brain for feedback consolidation, enabling closed-loop continuous alpha discovery that outperforms baselines on CSI300.

arXiv cs.CL·arxiv.org··1.7Researchllmquantai-research

AI·Researchers developed SQZ-Qwen-ASR-1.7B for MLC-SLM 2026 Task 1, combining a modular speaker diarization front-end (VAD, CAMPPlus embeddings, spectral clustering, RTTM segmentation) with Qwen3-ASR-1.7B adapted via full supervised fine-tuning, LoRA on TTS-generated synthetic speech, and GRPO reinforcement learning for lower tcpMER (23.70 on dev set, 17.97 on eval set).