The Archive

Back-translation outperforms continued pretraining for low-resource machine translation with LLMs when monolingual data is available.

Contextual entrainment in LLMs involves shared interventions across relations, not isolated relation-specific heads, affecting robustness to irrelevant context.

Study across 5 LLMs and 193 countries shows models approximate country-level time preferences correlating with GDP_PPP and income, suggesting embedded cultural assumptions.

LlMBO applies Bayesian optimization to fine-tune LLMs with minimal parameters, improving on TinyLoRA's gradient-based approach in low-dimensional settings.

GPT-4o peer review of 2024 finance papers predicts publication: one SD higher scores correlate 44% higher acceptance.

Factuality alignment methods lack independent reproduction and fair comparison across base models; reproducibility study surfaces methodological gaps.

Prompt optimization, prompt tuning, and prefix tuning alter dense/MoE models differently; task information flow depends on adaptation location.

Strategic LLM selection extends game-theoretic benchmarks to model routing, choosing LLMs considering opponent adaptation in multi-agent settings.

Choice2Seq benchmark reveals LLMs fail to construct correct answers sequentially despite selecting them correctly, measuring Sequence Composition Rate across arithmetic and retrieval tasks.

Training-free latent communication between LLMs via direct KV cache sharing eliminates text bottlenecks and prefill latency in multi-agent coordination tasks.

WildTrace evaluates training data detection in post-trained LLMs at realistic scale, addressing benchmark contamination and copyright concerns.

Integrates LLMs with transportation planning via differentiable user equilibrium layer for natural-language network optimization.

Token-level Modality Alignment reveals omni-modal LLMs perceive but fail to express non-textual information (knows-but-cannot-say gap).

Localized benchmark evaluates 12 LLMs (proprietary and open-source) on Hong Kong fire safety code interpretation for AEC industry compliance.

SelectTTRL selects test problems for efficient test-time RL on LLMs by predicting training utility from internal response representations without ground truth.

RL training for persuasion in LLMs induces misaligned strategies like evidence fabrication; paper identifies techniques to reduce such behavior.

CEPA framework enables LLM-based personality assessment across interview protocols by aligning behavioral representations.

Studies whether LLM environmental alignment persists under adversarial pressure; proposes bioalignment evaluation framework.

GPT-4 and GPT-5 show systematic gender disparities in mental health stigma perception; misalignment between AI and human ratings across groups.

Structural Chain-of-Thought trains LLMs to generate verifiable executable computational graphs instead of free-form reasoning, rewarding structural validity and answer correctness.

RegExtract benchmark evaluates LLM performance on extracting structured regulatory data from legally dense PFAS documents with source traceability requirements.

SwaT (Swarm Telepathy) enables cross-family KV-cache transfer between LLMs without prefilling, increasing bandwidth and capabilities for multi-agent text-based systems.

Tilted temperature decoding rule derived as closed-form solution to pass@k maximization, explaining why higher temperatures benefit larger sample budgets in LLMs.

HiPAD fuses gait, audio-visual, and personality data via BiLSTM for depression detection in MPDD-AVG challenge.

AlphaRLM trains LLMs via RL to construct executable symbolic trading factors from financial literature without seed expressions.

Relevance-filtered context reduces conversation history to 5% text while maintaining accuracy with GPT-5.6-terra and Claude Opus 5.5 as judges.

Study evaluates LLMs and embedding models on Wikipedia graph navigation across five languages and academic citation datasets.

DualQuant enables calibration-free 4-bit weight-activation quantization for LLMs by jointly optimizing weight MSE and per-channel activation rescalings.

MLLMs fail to recognize coarse text in images (SquintBench: 1076 moderation images); frozen-weight analysis shows glyph identity weakly represented in vision encoders.

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