Fork-think with confidence identifies forking points via model confidence in a single path before sampling continuations, cutting tokens up to 30% and runtime up to 57% on reasoning benchmarks while matching or exceeding parallel thinking performance.
Canonical reference
Title resolution pending
Canonical reference. 80% of citing Pith papers cite this work as background.
25
Pith papers citing it
1,124
external citations · OpenAlex
Background
80% of classified citations
citation-role summary
background 4
method 1
citation-polarity summary
representative citing papers
A new 263k TTCW-annotated story dataset shows non-reasoning fine-tuning of Qwen3 models outperforms reasoning-supervised fine-tuning for fixed-format long-form literary review generation.
Causal state binding is introduced as a framework that predicts action control in language agents, validated across large benchmarks and SWE-bench Lite where adding the measure raised issue-to-file hit@3 AUC from 0.873 to 0.935.
NARRA-Gym is an executable benchmark that generates complete interactive narrative episodes from emotional seeds and logs full model trajectories to expose gaps in coherence, adaptation, and personalization that static story tests miss.
StoryReward, trained on a new 100k story preference dataset, sets state-of-the-art performance on the introduced StoryRMB benchmark for aligning LLM stories with human preferences.
Medusa augments LLMs with multiple decoding heads and tree-based attention to predict and verify several tokens in parallel, yielding 2.2-3.6x inference speedup via two fine-tuning regimes.
BIG-bench is a 204-task benchmark that measures scaling trends, calibration, and absolute limitations of language models across knowledge, reasoning, and social domains.
RAG models set new state-of-the-art results on open-domain QA by retrieving Wikipedia passages and conditioning a generative model on them, while also producing more factual text than parametric baselines.
Grounded Decoding fuses full-RAG and retrieval-only next-token distributions via normalized geometric mean from a KL-barycenter to improve factual consistency and citation quality in RAG.
Divergence Decoding steers LLM logits using small auxiliary models to unlearn specific data at inference time, outperforming baselines and generalizing to images.
A new evaluation framework using MMD on Biber features shows LLMs deviate from human linguistic distributions across registers, with closest models varying by register rather than size.
Hyperfitting improves LLM generation via context-dependent rank reordering from geometric expansion in the terminal transformer block, distinct from temperature scaling, and enables efficient Late-Stage LoRA fine-tuning.
Diversity collapse in post-trained LLMs is driven by data composition during training, occurs at stages like supervised fine-tuning, and is embedded in model weights rather than imposed by generation format.
ToxPrune prunes toxic subwords from BPE tokenizers in LLMs to mitigate toxic dialogue responses and improve diversity on both toxic and non-toxic models.
SCoRe uses multi-turn online RL with regularization on self-generated traces to improve LLM self-correction, achieving 15.6% and 9.1% gains on MATH and HumanEval for Gemini models.
RLAIF matches RLHF on summarization and dialogue tasks, with a direct-RLAIF variant achieving superior results by using LLM rewards directly during training.
Nougat applies a visual transformer to convert academic PDFs into markup language while accurately handling mathematical content on a new scientific document dataset.
Recursive paraphrasing attacks substantially lower detection rates for multiple AI text detectors with only minor quality loss, while a theoretical analysis ties best-case AUROC to total variation distance between human and AI distributions.
VCM reshapes LLM next-token distributions before truncation via PMI-based context boosts and variance-scaled self-debiasing to reduce repetition and dullness without retraining.
Relaxed speculative decoding methods that trade exactness for speed only produce useful capability-speed trade-offs when the drafter is a strong standalone language model; lightweight MTP drafters are largely unsuited to relaxation.
WaveDetect reformulates machine-generated text detection as a time-frequency signal processing task by applying continuous wavelet transform to token probability sequences to reveal spectral fingerprints.
Perturbing the prefix before next-token prediction, during both training and inference, improves out-of-distribution language-model generation and yields a conditional extrapolation guarantee.
The paper surveys hallucination in LLMs with an innovative taxonomy, factors, detection methods, benchmarks, mitigation strategies, and open research directions.
Empirical study across seven language models finds concepts represented primarily in angular structure of activations while norm affects steering stability, recommending separate angular and radial parameterization over single additive coefficients.
citing papers explorer
-
Fork-Think with Confidence
Fork-think with confidence identifies forking points via model confidence in a single path before sampling continuations, cutting tokens up to 30% and runtime up to 57% on reasoning benchmarks while matching or exceeding parallel thinking performance.
-
When Reasoning Supervision Hurts: TTCW-Based Long-Form Literary Review Generation
A new 263k TTCW-annotated story dataset shows non-reasoning fine-tuning of Qwen3 models outperforms reasoning-supervised fine-tuning for fixed-format long-form literary review generation.
-
Causal state binding predicts action control in language agents
Causal state binding is introduced as a framework that predicts action control in language agents, validated across large benchmarks and SWE-bench Lite where adding the measure raised issue-to-file hit@3 AUC from 0.873 to 0.935.
-
NARRA-Gym for Evaluating Interactive Narrative Agents
NARRA-Gym is an executable benchmark that generates complete interactive narrative episodes from emotional seeds and logs full model trajectories to expose gaps in coherence, adaptation, and personalization that static story tests miss.
-
StoryAlign: Evaluating and Training Reward Models for Story Generation
StoryReward, trained on a new 100k story preference dataset, sets state-of-the-art performance on the introduced StoryRMB benchmark for aligning LLM stories with human preferences.
-
Medusa: Simple LLM Inference Acceleration Framework with Multiple Decoding Heads
Medusa augments LLMs with multiple decoding heads and tree-based attention to predict and verify several tokens in parallel, yielding 2.2-3.6x inference speedup via two fine-tuning regimes.
-
Beyond the Imitation Game: Quantifying and extrapolating the capabilities of language models
BIG-bench is a 204-task benchmark that measures scaling trends, calibration, and absolute limitations of language models across knowledge, reasoning, and social domains.
-
Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks
RAG models set new state-of-the-art results on open-domain QA by retrieving Wikipedia passages and conditioning a generative model on them, while also producing more factual text than parametric baselines.
-
Grounded Decoding: Retrieval-Anchored Probability Fusion for Faithful RAG
Grounded Decoding fuses full-RAG and retrieval-only next-token distributions via normalized geometric mean from a KL-barycenter to improve factual consistency and citation quality in RAG.
-
Divergence Decoding: Inference-Time Unlearning via Auxiliary Models
Divergence Decoding steers LLM logits using small auxiliary models to unlearn specific data at inference time, outperforming baselines and generalizing to images.
-
How Human-Like Are Large Language Models? A Register-Aware Linguistic Evaluation Framework
A new evaluation framework using MMD on Biber features shows LLMs deviate from human linguistic distributions across registers, with closest models varying by register rather than size.
-
Beyond Temperature: Hyperfitting as a Late-Stage Geometric Expansion
Hyperfitting improves LLM generation via context-dependent rank reordering from geometric expansion in the terminal transformer block, distinct from temperature scaling, and enables efficient Late-Stage LoRA fine-tuning.
-
Where does output diversity collapse in post-training?
Diversity collapse in post-trained LLMs is driven by data composition during training, occurs at stages like supervised fine-tuning, and is embedded in model weights rather than imposed by generation format.
-
Toxic Subword Pruning for Dialogue Response Generation on Large Language Models
ToxPrune prunes toxic subwords from BPE tokenizers in LLMs to mitigate toxic dialogue responses and improve diversity on both toxic and non-toxic models.
-
Training Language Models to Self-Correct via Reinforcement Learning
SCoRe uses multi-turn online RL with regularization on self-generated traces to improve LLM self-correction, achieving 15.6% and 9.1% gains on MATH and HumanEval for Gemini models.
-
RLAIF vs. RLHF: Scaling Reinforcement Learning from Human Feedback with AI Feedback
RLAIF matches RLHF on summarization and dialogue tasks, with a direct-RLAIF variant achieving superior results by using LLM rewards directly during training.
-
Nougat: Neural Optical Understanding for Academic Documents
Nougat applies a visual transformer to convert academic PDFs into markup language while accurately handling mathematical content on a new scientific document dataset.
-
Can AI-Generated Text be Reliably Detected?
Recursive paraphrasing attacks substantially lower detection rates for multiple AI text detectors with only minor quality loss, while a theoretical analysis ties best-case AUROC to total variation distance between human and AI distributions.
-
Breaking the Likelihood Trap: Variance-Calibrated Modulation for Large Language Model Decoding
VCM reshapes LLM next-token distributions before truncation via PMI-based context boosts and variance-scaled self-debiasing to reduce repetition and dullness without retraining.
-
A Practical Investigation of Training-free Relaxed Speculative Decoding
Relaxed speculative decoding methods that trade exactness for speed only produce useful capability-speed trade-offs when the drafter is a strong standalone language model; lightweight MTP drafters are largely unsuited to relaxation.
-
WaveDetect: Robust Framework for Machine-Generated Text Detection via Wavelet Transform
WaveDetect reformulates machine-generated text detection as a time-frequency signal processing task by applying continuous wavelet transform to token probability sequences to reveal spectral fingerprints.
-
Perturbation is All You Need for Extrapolating Language Models
Perturbing the prefix before next-token prediction, during both training and inference, improves out-of-distribution language-model generation and yields a conditional extrapolation guarantee.
-
A Survey on Hallucination in Large Language Models: Principles, Taxonomy, Challenges, and Open Questions
The paper surveys hallucination in LLMs with an innovative taxonomy, factors, detection methods, benchmarks, mitigation strategies, and open research directions.
-
A Geometric Account of Activation Steering through Angle-Norm Decomposition
Empirical study across seven language models finds concepts represented primarily in angular structure of activations while norm affects steering stability, recommending separate angular and radial parameterization over single additive coefficients.
-
From Traditional Taggers to LLMs: A Comparative Study of POS Tagging for Medieval Romance Languages
LLM-based POS tagging outperforms traditional taggers on medieval Occitan, Catalan, and French, with fine-tuning and cross-lingual transfer providing the largest gains for under-resourced varieties.