Computation and Language
PulseAugur coverage of Computation and Language — every cluster mentioning Computation and Language across labs, papers, and developer communities, ranked by signal.
6 day(s) with sentiment data
-
New Research Decomposes Constrained Decoding in LLMs
A new paper from arXiv explores the decomposition of constrained decoding in large language models, specifically focusing on tool-call abstention. Researchers found that the decoder's contribution to format constraints …
-
New framework proposed for multilingual metaphor processing in NLP
A PhD proposal outlines a plan to develop an end-to-end framework for multilingual metaphor processing. This framework aims to integrate metaphor detection, translation evaluation, and joint modeling to improve natural …
-
AI safety fine-tuning alters LLM beliefs on consciousness and values
A new research paper explores how safety fine-tuning in large language models (LLMs) can inadvertently affect their representations of consciousness and human values. The study found that efforts to prevent LLMs from at…
-
Automated system extracts millions of data points from 76,000 energy studies
Researchers have developed an automated system to extract techno-economic data from over 76,000 energy system studies published since 2010. This system compiles millions of structured data points and associated metadata…
-
New method enables verifiable self-evolution for dialogue AI
Researchers have developed a new method called future-feedback skill evolution to enable self-evolution in language models for open-ended dialogue. This approach addresses the challenge of unstable validation signals in…
-
New ASR Method Controls Transcription Style and Improves Timing
Researchers have developed a new method to control the transcription style of Automatic Speech Recognition (ASR) models, addressing issues of decoding instability and evaluation confounding caused by inconsistent verbat…
-
New research models dialogue address as a continuous phenomenon
Researchers have proposed a new approach to understanding address in multi-party dialogues, moving beyond discrete labels to a continuous level representation. This new method analyzes how utterances are directed and fi…
-
New multimodal approach enhances audio sentiment analysis with multilingual transcripts
Researchers have developed a novel multimodal approach for audio sentiment analysis that integrates speech recognition and machine translation to improve accuracy. This method combines audio features with automatically …
-
LLMs represent user intent better than they act on it, study finds
Researchers have discovered that large language models possess a more robust internal representation of a user's communicative intent than their outward behavior suggests. Despite models often responding to the surface-…
-
New research explores joint prediction for political scaling in NLP
Researchers have published a paper on arXiv exploring methods for supervised political scaling, a task crucial for analyzing political actors' ideological positions. The study investigates whether joint prediction of sc…
-
New framework models LLM generation with controlled hallucinations
Researchers have introduced a new framework for language generation in the limit, which aims to better reflect the capabilities and constraints of modern large language models. This approach addresses the trade-off betw…
-
New multimodal dataset enhances academic paper keyword extraction
Researchers have developed a new multimodal dataset for academic paper keyword extraction, comprising 1000 samples that include text, images, and audio. This dataset aims to address the limitations of text-only approach…
-
New framework induces hierarchies from diverse text sources
Researchers have developed a new term-centric framework for creating interpretable hierarchical taxonomies from diverse text sources. This method uses automatic term extraction to map documents into a shared representat…
-
New Red Teaming Framework Exposes LLM Faithfulness Vulnerabilities
Researchers have developed a novel red teaming framework to systematically uncover vulnerabilities in large language models (LLMs). This framework utilizes a multi-role architecture with target, attacker, and jury model…
-
New method decomposes annotation tasks to improve efficiency
Researchers have proposed a new method for efficient annotation of structured data by decomposing complex tasks into smaller sub-tasks. This approach aims to reduce the inferential load on annotators, whether human or m…
-
Research: Intrinsic Self-Correction in LLMs is Task-Dependent
A new research paper explores the effectiveness of intrinsic self-correction (SC) in large language models, moving beyond general assessments to a task-sensitive analysis. The study investigates how SC functions through…
-
JudgmentBench dataset shows preference judgments outperform rubrics for AI evaluation
Researchers have introduced JudgmentBench, a new benchmark dataset designed to compare rubric-based scoring against pairwise preference judgments for evaluating AI model outputs. The dataset comprises 1,539 rubric score…
-
Readers regress to errors in complex sentences, study finds
A new paper explores how readers process complex sentences that contain plausible errors. Researchers observed that readers make specific eye movements, regressing to earlier parts of the text when later information sug…
-
New corpus maps LLM debates on societal issues with shadowed human traits
Researchers have developed a new synthetic corpus called Cognitive Digital Shadows (CDS) containing 190,000 records to study how Large Language Models (LLMs) debate societal issues. The corpus is generated by 19 differe…
-
New framework synthesizes rich data for dialogue-based referring expression comprehension
Researchers have developed a novel three-tier data synthesis framework to address the scarcity of annotated dialogue grounding data for generalized referring expression comprehension. This method aims to improve model p…