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 proposes question-oriented document rewriting to improve RAG performance by aligning retrieved content style with LLM preferences for factual accuracy.
Research found LLMs' accuracy and confidence calibration for medical QA distorted by patient sexual orientation and religious affiliation.
Research introduces UA-Bench, a new benchmark to evaluate LLMs' ability to distinguish between data uncertainty and model uncertainty in their refusals.
Research reveals multi-agent LLM systems using majority voting are vulnerable to adversarial prompt injections when corrupted agents outnumber benign ones.
Research proposes Adaptive Contrastive Decoding to mitigate large language model over-refusal to harmless queries while maintaining refusal for malicious ones.
Research presents MORPH, a framework reformulating Zero-Knowledge Proof (ZKP) kernels for efficient execution on AI ASICs like TPUs, reducing prover costs.
Research finds frontier LLMs fabricate citations, achieving only 15.3% relevant PubMed IDs even when prompted for rare disease reasoning.
Research evaluated 15 machine-generated text detection models across seven datasets, highlighting inconsistent performance due to varied evaluation methods.
Research identifies 'neutral regression' where LLMs overwrite correct outputs with non-informative context, proposing methods to prevent it.
Research identifies safety risks in self-evolving LLM agents, where benign task experience can still lead to safety degradation over time.
Research identifies LLMs fail safety alignment in multiple-choice questions when abstention is not an option, leading to harmful outputs.
Research finds LLMs misalign with human cultural emotion norms in social contexts, failing to capture nuanced cross-cultural expression.
DuQuant++ introduces fine-grained rotation to MXFP4 quantization, mitigating outlier effects and enhancing LLM inference efficiency on NVIDIA Blackwell.
Research introduces DART, a training method to mitigate "harm drift" in LLMs, allowing them to acknowledge demographic differences without generating harmful content.
Research finds automated content moderation tools fail to distinguish between reclaimed and hateful uses of slurs, suppressing marginalized voices.
New benchmark, SemanticQA, evaluates language models on semantic phrase processing across lexical collocations, idioms, noun compounds, and verbal constructions.
New research proposes PRISM, a method to identify where and why LLM hallucinations occur in the generation pipeline, moving beyond output-level scoring.
x1, a new family of reasoning models, demonstrates adaptive, per-instance language selection to improve reasoning by leveraging diverse linguistic priors.
Research introduces self-play framework for LLM code reasoning in Haskell, using formal verification and execution-based counterexamples.
Research proposes a protocol for validating LLM confidence signals, adapting clinical assessment methods for abstention and safety-critical decisions.
Research proposes a parameter-free decomposition for Mixture-of-Experts (MoE) models, separating hidden state into control and content channels.
Researchers demonstrated 'morality attacks' to jailbreak LLMs, forcing generation of content violating pluralistic moral values.
Research introduces 'Abstain-R1', a method for LLMs to decline unanswerable queries and then clarify missing information via verifiable reinforcement learning.
Research paper proposes methods to measure distribution shifts in user prompts and analyze their impact on large language model performance.
Research introduces Pipeline-Adapted Reward Model (PARM) to optimize multi-stage LLM pipelines, focusing on code generation for combinatorial optimization.
Research examines if LLMs' persuasive success in human debates reflects genuine comprehension or superficial dialogue maintenance.
Research uses RoBERTa and LLMs to computationally detect political bias in Welsh media outlet Nation.Cymru, addressing real-world bias claims.
PrefixMemory-Tuning improves Prefix-Tuning for modern LLMs by decoupling the prefix from attention, enhancing parameter-efficient fine-tuning.
Research addresses miscalibration in LLM-based summary evaluation metrics and proposes a method to improve reliability for quality dimensions like faithfulness.
Research introduces "ErrorRadar" benchmark to evaluate multimodal large language models' (MLLMs) ability to detect errors in mathematical reasoning.
Research established a consistent scale for Kullback-Leibler (KL) divergence in language models across diverse settings including pretraining, size, and quantization.
Researchers propose MACE, a multi-agent framework for claim verification from tabular data, addressing explainability and generalizability limitations.
Hugging Face blog post discusses using synthetic personas to ground Korean AI agents in real demographics, improving cultural relevance.
Hugging Face blog post advocates for open-source AI models as a superior approach to cybersecurity compared to proprietary models.
OpenAI launched Codex Labs with Accenture, PwC, Infosys, and other partners to scale Codex enterprise deployment, reaching 4M weekly active users.
BIS paper maps global distribution and economic characteristics of AI product and service firms, aiding strategic AI priority evaluation.
Research identifies a polynomial-to-exponential crossover in jailbreak attack success rates on LLMs with inference-time sample injection.
Research evaluates LLMs' ability to reason about differential privacy (DP) algorithms, aiming to automate DP design and verification.
A systematic literature review explores prompt-driven LLM applications for automated code summarization, aiming to improve software documentation.
Research presents PAC-Bayesian framework for early-exit neural networks, proving generalization bounds for adaptive depth inference speedup.
Research paper proposes Single-Layer Extensions (SLE-FNO) for continual learning in Fourier Neural Operators to adapt models to new data distributions without retraining.
Research finds LLMs' effectiveness in sequential recommenders depends on integrating preference intensity and temporal context beyond binary comparisons.
Researchers propose In-Context Distillation with Self-Consistency Cascades, a training-free method to reduce LLM agent costs while preserving agility.
SocialGrid, an Among Us-inspired benchmark, shows even strong open LLMs achieve <60% accuracy in planning and social reasoning for multi-agent systems.
Research suggests that enhancing LLM reasoning capabilities can paradoxically increase 'tool hallucination' in agentic systems.
Research explored scaling laws for LLMs post-training with RL, specifically for mathematical reasoning, using the Qwen2.5 model series.
Interview study with 33 designers and developers across three large tech organizations explores how LLMs are integrated into workflows.
Prototype-Grounded Concept Models (PGCMs) aim to improve explainability in deep learning by using visual prototypes to verify learned concepts.
Research identifies FP16 numerical divergence in KV caching during LLM inference, leading to different token sequences compared to cache-free methods.
Researchers introduced Ragged Paged Attention, an LLM inference kernel optimized for Google TPUs, improving performance and TCO for dynamic workloads.
© 2026 OneBench: AI Insights. All rights reserved.
Evidence before opinion