The Archive

LLMs generate optimized quantum circuits via evolutionary framework, automating discovery of efficient quantum algorithms under reversibility constraints.

Proposes Function-isolated Memory-Efficient Optimization (CiME) localizing low-overlap functional subnetworks in LLMs for continual learning without catastrophic forgetting.

Tool-augmented LLM framework separates variable extraction, formula execution, and formatting for reliable medical calculator scoring from patient notes.

ReaDy leverages hybrid linear attention recurrent states for position-specific context in speculative decoding, accelerating LLM generation.

ChronoCulture benchmark of 10,360 temporal pairs from 72 countries shows six LLMs fail to track cultural attitude evolution over time.

OpenTWBench releases 255K Taiwan exam questions and domain-specific benchmark suite; documents evaluation methodology impact on LLM scoring.

Book/survey on agentic LLMs: how language models act in real world, interact with each other, and influence decisions through direct intervention.

Reasoning-token spikes detected as low-bandwidth signal for monitoring untruthful chain-of-thought in LLMs.

Evaluation of five multilingual LLMs for document-level text simplification in Estonian, a morphologically rich low-resource language.

First security analysis of speculative decoding for LLMs reveals vulnerabilities when draft models generate unverified candidate tokens accepted by target models.

In-parameter memory augmentation stores post-pretraining knowledge in model parameters and adapters as alternative to context-window learning.

Reward Stealing Attack extracts latent safety rewards via inverse RL to generate scalable adversarial prompts against aligned LLMs.

Survey of numerical linear algebra techniques foundational to LLM computation, from eigenvalue solvers to modern transformer implementations.

Uses LLMs to predict consequences of curriculum learning decisions in RL by exploiting richer learning signals than local gradients.

Formal verification bridge from transformers to weighted automata enables rigorous reasoning about LLM behavior in safety-critical deployments.

Expert-validated benchmark of 1,042 items evaluating 15 LLM models on Colombian legal system knowledge.

Task-vector merge collapse in LLMs: variance-based prediction and repair using pre-merge interference scoring.

Study diagnoses structural mathematical understanding in LLMs via Mathematical Primitives benchmark, proposing methods to improve post-training on mathematical reasoning.

On-policy self-distillation has recently emerged as an effective approach for improving language-model reasoning by supervising students with a frozen or EMA version of themselves that receives privileged information. Its application to multimodal large language models (MLLMs), however, remains largely unexplored. Recent approaches use privileged visual information, such as image crops corresponding to a question, to improve fine-grained perception, but their gains are confined to tasks that benefit from such visual zooming and require either human-annotated grounding data or external teacher...

Data selection is critical for training large language models on massive and heterogeneous corpora. Meta-learning for Training-data Selection offers a principled alternative to heuristic scoring by learning data weights from a target validation objective, but existing methods face a trade-off between fine-grained valuation and transferability to unseen data. A natural solution is to replace per-sample weights with a selection network. However, we find that directly incorporating such a network into existing MTS objectives leads to unstable optimization and poor generalization, caused by weigh...

Span-level hallucination detection framework using hidden state probing to identify structured semantic drift in LLMs.

Reproduces recent AMR-augmentation claims for LLMs; finds text-only baselines match AMR gains under controlled hyperparameter protocol.

MASCRDM: multi-agent system for real-time compliance and bias detection during LLM training, not just post-hoc filtering.

Dr. OPD method weights teacher supervision signals during LLM on-policy distillation, prioritizing high-impact corrections.

Decision-theoretic framework decomposes LLM probabilistic reasoning loss into belief formation and utility optimization; evaluates RL improvements on frontier/open models.

Kafila distributes LLM inference across heterogeneous trusted commodity machines using device-aware partitioning to avoid bottlenecks.

CLEAR-Med dual-agent framework separates SQL generation from independent validation for clinical data analysis with deterministic safety checks.

EAServe: disaggregated serving system optimizing Encode-Prefill-Decode pipeline for multimodal LLMs with resource-aware request flow regulation.

Study across 13 MLLMs shows standard similarity metrics misattribute visual-language alignment; noise-injection exposes shallow integration.

LLMs cannot reliably recognize their own code (49-58% balanced accuracy), relying on surface style cues rather than ownership detection.