Structured Output Collapses Answer Diversity Across 44 Language Models
Research finds requesting JSON output from 44 LLMs significantly reduces answer diversity on open-ended prompts, collapsing options.
arXiv cs.CL — Computation and Language·22 Jul
Search signals, briefings, company results, benchmarks and glossary terms.
Search signals, briefings, company results, benchmarks and glossary terms.
Use this view to inspect the underlying evidence corpus. For ranked developments, decision posture and interpretation, use Signals.
Raw feed or Signals?
Raw feed is chronological evidence. Signals ranks and interprets material change.
Research finds requesting JSON output from 44 LLMs significantly reduces answer diversity on open-ended prompts, collapsing options.
Research proposes RLAES, an LLM framework using reinforcement learning with rubric rewards for automated essay scoring and feedback generation.
New research proposes Causal Alignment and Structural Enforcement (CASE) to improve chain-of-thought faithfulness in LLMs, ensuring reasoning supports answers.
Research introduces PARSE, a prompt injection defense for LLM agents, demonstrating improved resilience on a new benchmark of real enterprise documents.
New research proposes ScAffolded Generative models for Explanation (SAGE) framework for computational pragmatic reasoning in language models.
Research explores adapting LLM-based classification, developed on US police data, to identify vulnerability indicators in UK police incident logs.
RF-Agent is a new framework using textbook-driven knowledge distillation and a multi-agent QTSA pipeline for RFIC design, addressing data scarcity.
XCOMPS, a new multilingual benchmark for conceptual minimal pairs across 17 languages, reveals weaker LLM understanding for low-resource languages.
Research finds LLM prompt robustness differs between objective questions with fixed answers and subjective opinion-based queries.
Research identifies "repetitive copying" as a critical failure mode in long-context LLM reasoning, proposing evidence-aware reinforcement learning.
Research on Dice Loss for natural language processing tasks addresses data imbalance where negative examples significantly outnumber positive ones.
SIFT introduces a self-improving classifier for dynamic document classification that reduces reliance on upfront labeling and continuous retraining.
Research identifies 'Lifelong Normalization' as key to stable, continuous LLM updates without catastrophic forgetting, critical for model editing.
Guard Vector introduces a method for LLM guardrails using task-vector composition and streaming-aware prefix SFT, moving beyond English-centric models.
ChineseBERT incorporates glyph and pinyin information into pretraining models for enhanced Chinese language understanding.
Research introduces 'Node-as-Agent' framework, combining Graph Neural Networks with agentic capabilities to address node informativeness and local structural limitations.
AutoJourn is a research system demonstrating multi-perspective summarization, bias detection, and bias neutralization for LLM-generated news.
A new research framework, MEDIC, introduces leading indicators beyond licensing exams to evaluate LLM competence for clinical workflow utility.
Research shows ASR models encode both verbatim and intended transcription styles, but uncontrolled activation causes decoding instability and unreliable word timing.
Research explores 'vectorizing the Trie' for efficient constrained decoding in LLMs, addressing limitations for generative retrieval in recommendation systems.
Research explores rationale-guided knowledge distillation for cross-lingual stance detection, improving model performance in low-resource languages.
A research paper from arXiv introduces a multi-role red teaming framework to systematically uncover vulnerabilities and evaluate faithfulness in LLM outputs.
CLT-Forge is a new library for mechanistic interpretability research, using cross-layer transcoders to create more manageable attribution graphs for LLMs.
Research paper discusses agentic LLM systems moving from prototypes to production across software, science, and finance, highlighting deployment challenges.
Research explores challenges in style-personalized text generation from LLMs, noting high specificity and context dependency make evaluation difficult.
Research explores small language models (SLMs) for creative plot generation, focusing on global coherence, character, pacing, and emotional progression.
A new survey on arXiv details the evolution of AI for mathematical reasoning, from rule-based solvers to contemporary LLM-driven approaches.
Replicated prior research on human label variation in Natural Language Inference (NLI), confirming lower agreement for non-upward monotonicity operators.
Research from arXiv investigates how instruction-tuned Transformer models, LLaMA and Mistral, encode causation and antithesis.
Research indicates Diffusion Language Models (DLMs) encode an internal, latent representation of denoising progress, similar to explicit timesteps.
PRISP proposes a privacy-safe, few-shot personalization method for LLMs, addressing data scarcity, limited compute, and strict privacy.
Research explores using Vector Symbolic Architectures (VSAs) to decode internal representations and improve interpretability of Large Language Models.
Research experimentally evaluates prompt design factors—format, instruction count, and context length—on LLM instruction adherence and hallucination.
CircuitKIT is a new toolkit for mechanistic interpretability, unifying methods for discovering, evaluating, and applying model circuits.
Researchers propose HindsightBench, a black-box audit protocol to detect parametric hindsight (leakage of future knowledge) in LLMs used for time-indexed decision tasks.
Research explores using depthwise convolutions within Transformer blocks to improve local inductive bias in LLMs, comparing placements across Qwen3.
Research proposes Hierarchical Parallel Document Parsing (HPD-Parsing) to address sequential bottlenecks in unified VLM-based document parsers.
Research evaluates four language models for adaptability, effectiveness, and limitations in the low-resource Aminoacian language.
Doctorina MedBench-ICD10 introduces a new dialogue-based evaluation framework for agent-based medical AI, simulating physician-patient interactions.
AdaFlash introduces adaptive speculative decoding using diffusion drafters to accelerate large language model inference by generating drafts in a single pass.
Research explores enhancing legal machine translation using reasoning-capable language models to improve precision and address linguistic complexity.
Research identifies and quantifies "post-hoc rationalization" in Reverse Chain-of-Thought (RCG) LLM generation, where models justify a pre-committed answer.
AFIR, a Romanian government agency, deployed RAGAL, a fully local RAG-based assistant for technical support, adhering to zero data egress.
Researchers introduced MedDDC-Eval, a diagnosis-decoupled evaluation framework for multi-turn medical consultation agents to isolate policy elicitation from diagnosis generation.
Research introduces MaLoRA, a dynamic low-rank adaptation method for LLMs, enabling token and instance-level state adaptation during inference.
DAIS (Dependency-Aware Intermediate QA Supervision) is a new training framework converting teacher rationales into stage-level QA records to improve complex reasoning.
MUX proposes distilling discrete LLM reasoning steps into continuous multiplexed tokens for more efficient, high-bandwidth computation.
Research explores self-evolution of LLM dialogue skills using future-feedback prediction to address unstable validation signals in open-ended conversations.
Research evaluates multimodal LLMs' theory-of-mind (ToM) reasoning in multi-party meetings, identifying current limitations beyond overt signals.
New benchmark, GAMUT, focuses on evaluating the factual completeness of open-ended LLM generations, addressing the gap beyond factual precision.
© 2026 OneBench: AI Insights. All rights reserved.
Evidence before opinion