The Archive

LLM-based validation of atom-centered structural descriptors reveals descriptor degeneracies in materials science.

Preregistered experiment: diverse human groups (L1/L2 English writers) produce more collective creativity than LLMs; AI assistance risks homogenizing or replacing human diversity.

DIRECT framework applies DPO and controlled decoding to improve LLM sequence labeling alignment and inference efficiency for information extraction tasks.

Input-only prompt optimization suppresses evaluation-awareness latents in LLMs via GCG-style token search, threatening safety evaluation validity if undetected.

KQFuzz: LLM-based knowledge-guided fuzzer for quantum libraries using codebase context to improve test generation efficiency and flexibility.

Construction-Driven Injection embeds linguistically-grounded code-mixing fingerprints in LLMs to create verifiable ownership signals resistant to accidental activation and perplexity filtering.

Philosophical critique of epistemic risks in LLM outputs challenges the 'Epistemia' framework by invoking semiotic and situated-knowledge perspectives.

D-Score detects hallucinations via spectral analysis of hidden activations, using single forward pass.

Evaluation of multilingual NMT models (NLLB, mBART) on English-Tamil translation across four datasets with BLEU/chrF metrics.

MedGame transforms static clinical cases into interactive decision-driven learning games using LLMs and dual narrative/director engines.

LLMs systematically overuse epanorthosis (classical self-correction rhetoric) due to promotional training distributions and RLHF preference for emphatic phrasing.

Speech-based multimodal LLMs detect cognitive impairment across diverse speakers and devices by leveraging linguistic and acoustic biomarkers with improved generalization.

Open-source evaluation framework for open-weight LLM agents on longitudinal data tasks, addressing privacy constraints in research deployments.

LKValues: benchmark and fine-tuning resource for aligning LLMs to Sri Lankan cultural values in Sinhala.

PAC bounds on LLM harmful output probability via latent-space-guided tree exploration and Clopper-Pearson intervals.

Controlled study of prompt design (format, instruction count, context length) effects on instruction adherence and hallucination across five LLMs.

HuGLEN: LLM evaluation pipeline for optical network automation using expert ratings and quality-efficiency scoring.

New benchmark Pancasila-Dilemmas (1,834 questions) evaluates LLM value alignment on Indonesian cultural values beyond Western frameworks.

Benchmark evaluates six MLLMs on scientific visualization literacy using 49 standardized assessment items across 8 visualization techniques.

Training-free human-object interaction detection using multimodal LLMs for open-world compositional scenarios without dataset-specific supervision.

Benchmark evaluating LLM capability to identify and correct patient misconceptions across multi-turn medical conversations.

Survey of efficient inference techniques for diffusion-based LLMs, covering algorithms and systems beyond theoretical parallel generation claims.

LLMs fine-tuned to predict chemical reaction mechanisms step-by-step, reducing hallucinations versus name-reaction prediction approaches.

ESFP benchmark measures whether LLMs distinguish and coherently shift between neutral attribution and self-stance epistemic registers in contested claims.

LLMs parameterize cumulative prospect theory for modeling human route-choice biases at scale.

Rashomon Explanation paradigm couples prediction with explainability as complementary objectives, showing self-explanation can improve rather than degrade model accuracy.

The AI firm Anthropic has developed a technique that has given it the clearest glimpse yet at what’s really going on inside large language models as they answer questions or carry out tasks. What they found ranges from the mundane to the unnerving. Researchers at the company built a tool called the Jacobian lens (or…

Framework compares test-time resampling vs. rerouting for cost-aware LLM selection under realistic budget constraints.

When it comes to achieving artificial general intelligence (AGI), large language models just don’t have what it takes. Models like ChatGPT and Claude are great at text, but they’re less skilled at understanding how things actually move through space and time — an essential skill for producing intelligence that generalizes. That gap, it turns out, might be filled by gaming data. That’s the bet behind General Intuition, a […]

Future Confidence Distillation measures pre- and post-solution confidence in LLM responses to improve reliability estimation for downstream tasks.