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 claims spectral analysis of LoRA adapters identifies fine-tuning objectives and predicts downstream harmful compliance behavior in LLMs.
Research claims decomposing LLM classification tasks into smaller sequential problems significantly reduces prediction error, scaling with a power law.
Researchers augmented a deep anomaly detection dataset for batch distillation with simulation data to improve model training for industrial processes.
Research identifies a unified mechanism for harmful content generation in LLMs, indicating current alignment training is brittle and jailbreaks exploit a common vulnerability.
NOMAD is a new research paper proposing a method to generate embeddings for massive distributed graphs, addressing scalability limitations of existing techniques.
Research introduces Automated Instruction Revision (AIR), a rule-induction method for LLM adaptation with limited examples, comparing it to prompt optimization and fine-tuning.
Researchers propose ESPRESSO, a deep learning method, for detecting stepping-stone intrusions in networks by correlating traffic flows.
Research proposes Dictionary-Aligned Concept Control for MLLMs, dynamically steering activations during inference to mitigate unsafe responses without fine-tuning.
Research proposes a 'Two-Stage Decision-Sampling Hypothesis' explaining how RL post-training fosters self-reflection in LLMs, improving multi-turn performance.
nextAI fine-tuned LLaMa2 70B on a single A100 40GB GPU for the NeurIPS LLM Efficiency Challenge, optimizing for resource usage.
Research models LLM decision-making for automation: act vs. escalate. Applies to forecasting, content, loan approval, and autonomous driving.
Research paper proposes FP8 low-precision stack for stable reinforcement learning with LLMs to accelerate rollout/generation and reduce memory bottlenecks.
Research proposes CLIP-Inspector, a method to detect backdoors in prompt-tuned Vision-Language Models (VLMs) like CLIP, when training is outsourced.
Research introduces a new tensor decomposition method to quantify uncertainty in Large Language Model-based Multi-Agent Systems, addressing limitations of single-agent UQ methods.
Research proposes Reinforcement-aware Knowledge Distillation (RaKD) to compress large, RL-trained LLMs for reasoning while maintaining performance.
Research introduces HiFloat4, a 4-bit floating-point format for LLM pre-training on Ascend NPUs, claiming efficiency gains over existing FP4 formats.
Research identifies extrinsic gender bias in Bangla pretrained language models for sentiment, toxicity, hate speech, and sarcasm detection.
Research proposes PACED, a distillation method weighting training problems by student pass rate (p(1-p)) to improve efficiency.
Research finds LLM neurons consistently exhibit polysemantic behavior, challenging discrete neuron-concept attribution for model interpretation.
Research proposes ImageProtector, a visual prompt injection method to prevent multi-modal LLMs from analyzing images for sensitive information.
Research finds low-data supervised fine-tuning outperforms prompting for adapting vision-language models to remote sensing imagery with domain shift.
Research proposes HaloProbe, a Bayesian method to detect and mitigate object hallucinations in Vision-Language Models, improving reliability beyond attention weights.
Research proposes Uncertainty-Aware Transformers using conformal prediction to quantify prediction uncertainty in LLMs for high-stakes applications.
Research analyzed how bias mitigation reshapes embedding spaces in BERT and Llama2, reducing gender-occupation associations.
Research compared LLMs and fine-tuned BERT models for Arabic sentiment analysis on Gaza War news headlines using a 10,990 headline dataset.
Layer pruning for LLMs effective for classification, but significantly degrades generative reasoning tasks (e.g., GSM8K, HumanEval+).
Research identifies decision boundary proximity as a common cause for miscalibrated confidence and paraphrase sensitivity in medical Vision-Language Models.
Research introduces Gen-n-Val, an agentic framework for generating and validating synthetic image data to address scarcity, noise, and class imbalance in computer vision datasets.
Research demonstrates dynamic sparsity in tree-structured feed-forward layers reduces transformer compute, a drop-in MLP replacement.
Research paper benchmarks ten deep learning uncertainty quantification (UQ) methods, finding auxiliary losses often ineffective for calibration.
Research evaluated 8 SOTA LLMs on a new benchmark with 14 perturbation techniques against the AIME 2024 dataset, finding reasoning robustness varies.
StaRPO, a new RL policy optimization framework, improves LLM logical consistency and structural coherence in complex reasoning tasks by capturing internal logic.
Research evaluates temperature and prompting strategies (CoT, zero-shot) for extended reasoning in LLMs, specifically Grok-4.1.
Research introduces a kill-chain canary methodology to track prompt injection attacks through multi-stage LLM systems, moving beyond binary success/failure metrics.
Research introduces BRIXEL, a method to achieve dense feature maps with lower compute and memory, addressing the high-resolution demands of models like DINOv3.
Research proposes a proxy model framework to reduce computational cost for post-hoc interpretability of large language models.
The Weekend Windup #28 from Joe Reis discusses if fundamental data engineering principles still matter amidst AI advancements.
BenchBrowser, a research tool, retrieves evidence to evaluate if language model benchmarks accurately measure practitioner-intended capabilities.
Research proposes a rank-based uniformity test to audit black-box LLM APIs for performance degradation or model substitutions by providers.
FinTruthQA is a new benchmark for assessing financial disclosure quality using AI on Chinese stock exchange investor platforms, addressing non-substantive firm responses.
Research quantifies the contribution of individual information signals (e.g., reproduction test, edit location) to LLM agent performance in automated software engineering.
Research identifies 'problem drift' in multi-agent LLM debates where models deviate from initial tasks over longer reasoning chains, reducing performance.
Research paper unifies various LLM post-training methods (SFT, RL, preference optimization) into off-policy and on-policy learning frameworks.
Research introduces a new benchmark for evaluating the robustness of machine-generated text detectors against personalized LLM outputs, highlighting detection challenges.
Research comparing human vs. LLM judgment on plausible/implausible events, finding LLMs struggle with nuance in non-literal contexts.
OrgForge is an open-source multi-agent simulation framework for generating verifiable, internally consistent, and temporally structured synthetic corporate data.
Researchers introduced SEM-CTRL, a method integrating Monte Carlo Tree Search with LLM decoders to enforce context-sensitive semantic constraints on outputs.
LLMs struggled to detect (64% accuracy) and correct bias based on Wikipedia's Neutral Point of View policy, indicating difficulty with specialized norms.
Research indicates emotional framing in prompts degrades LLM quantitative reasoning, even when numerical content is identical.
Research paper evaluates LLMs for demographic-targeted social bias detection in large text corpora, addressing a key regulatory concern for data auditing.
© 2026 OneBench: AI Insights. All rights reserved.
Evidence before opinion