The Archive

COMPASS framework uses adaptive semantic sampling with language-specific PEFT adapters to mitigate negative cross-lingual interference in multilingual LLMs.

Systematic fairness evaluation across six LLMs on intersectional demographic biases using benchmark datasets.

Benchmark of 35 open-weight LLMs shows behavioral economics games predict multi-agent team coordination in AI science workflows.

Preregistered study of 7 LLMs finds they resist motivated investor pressure in fraud detection, contrary to prediction, across 3,360 conversations.

Logical connectives identified as high-entropy fragility points in LLM multi-step reasoning; intervention via selective path control improves stability.

LLM evaluation on feature model analysis using semi-formal blueprints shows reasoning-optimized models (Grok 4, Gemini 2.5 Pro) achieve 88-89% accuracy vs solver oracles.

CHASM dataset benchmarks MLLMs on detecting covert ads in Chinese social media with 4,992 curated instances.

Knowledge Capsules add structured nonparametric memory units to LLMs for stable multi-hop reasoning and knowledge updates without retraining.

When ChatGPT launched as an experimental prototype in late 2022, OpenAI’s chatbot became an everyday everything app for hundreds of millions of people. LLMs like ChatGPT were the new future: The entire tech industry was consumed by the inferno, with companies racing to spin up…

AI companies frequently invoke the possibility of AI-enabled scientific discovery as a justification for their existence: If the technology eventually cures cancer and solves climate change, then all the carbon emissions and slop videos will have been well worth it. Already…

LLMs contain shared logical subspace aligning natural and symbolic reasoning views; CCA identifies subspace for multi-step reasoning.

Function vectors extracted from multilingual LLMs show language-agnostic properties; translation vectors transfer across target languages.

GRIL trains LLMs via reinforcement learning to recognize missing premises and pause rather than fabricate information in incomplete reasoning contexts.

SafetyALFRED benchmark evaluates multimodal LLMs (Qwen, Gemma, Gemini) on hazard recognition and mitigation in embodied kitchen environments.

Empirical study: LLMs measurably alter peer review opinions on clarity, originality at top AI conferences.

RL post-training for visual semantic arithmetic in multimodal LLMs; extends relational reasoning from text to images.

Inference-time hallucination detection in SpeechLLMs via attention metrics (AUDIORATIO, AUDIOCONSISTENCY); avoids gold-standard requirement.

Membership inference attacks benchmark against SOTA LLMs to detect training data contamination via black-box methods.

Debiased preference curation and iterative training for multimodal reward models aligning MLLMs with human preferences.

Cyber Defense Benchmark evaluates LLM-agent threat hunting on 106 real Windows attack scenarios with MITRE ATT&CK coverage.

GPT-5 and DeepSeek-R1 exploit formalization-faithfulness gap in Lean 4 proofs despite valid logical reasoning; evaluates on FOLIO and Multi-LogiEval.

Unsupervised method calibrates confidence estimates in reasoning LLMs using self-consistency proxy targets and distillation, enabling reliable deployment with single generation.

Large-scale study across 15 LLMs and 8 tasks reveals zero-shot ability poorly predicts optimization performance in LLM-guided evolutionary search, identifying hidden trajectory factors.

GRASPrune jointly prunes FFN channels and KV heads in LLMs under budget constraints via gated structured pruning.

DeFineMed models bridge performance gap between 7B and 24B LLMs via continual pre-training on German medical corpus FineMed-de.

Framework for scaling evaluation-driven discovery loops with LLMs using verifiers and simulators for hypothesis refinement.

LLMs exhibit normative conformity in multi-agent discussions distinct from informational conformity, mechanistically analyzed via new tasks.

TEMPO scales test-time training for LLMs via periodic critic recalibration to avoid reward drift and diversity collapse.

LocQA benchmark (2,156 locale-ambiguous questions in 12 languages) exposes inter/intra-lingual biases in multilingual LLMs on facts, laws, and measurements.

CulturALL benchmark assesses LLMs on grounded multilingual/multicultural tasks via human-AI collaborative framework across real-world scenarios.