The Archive

Controlled study finds task decomposition in LLM-as-judge frameworks does not improve NLG evaluation; prior gains stem from confounded baselines.

SVD-MBR applies singular value decomposition to minimum Bayes risk decoding, reducing metric overfitting in text generation via low-rank approximation.

Analysis of Transformer attention heads reveals scaled idempotence pattern in OV operators across 2.8B–235B parameter models via principal-coordinate factorization.

Leakage-free evaluation protocol for online adaptation in edge time-series forecasting reveals warmup bias significantly affects baseline comparisons.

ChatGPT gains EHR integration for healthcare; clinicians can securely access patient context and medical research via enterprise connections.

Dual Neural-Calibrated IMM method improves lane-change intention recognition for autonomous driving by incorporating driving context via neural calibration.

Latent Recurrent Thoughts framework enables chain-of-thought reasoning in continuous representation space with frozen LLM and auxiliary network refinement.

EDRAC benchmark covers dialectal Arabic machine reading comprehension across five dialects (Egyptian, Moroccan, Emirati, Syrian, Saudi) with 499 passages.

ClinTraceBench evaluates 8 clinical EHR history representations via 385 MIMIC-IV dialogues with provenance, finding compact summaries degrade longitudinal reasoning signal.

Replication study of TRACE causal discovery method confirms 0.90+ F1 on synthetic data; reveals optimal threshold tied to error margin, not constant.

Empirical study on Qwen2.5-3B and Phi-3.5-mini shows hints in code generation provide steering, not missing information; unrelated hints rescue 40-50% as many failures.

Analyzes CLIP zero-shot classification via decision margins and class-wise hubness; smaller modality gap alone insufficient for accuracy gains without improved margin structure.

Neural Symbolic Regression framework combines neural nets as preconditioners with sparse regression, addressing scalability and noise sensitivity of genetic programming methods.

Contrastive Routing Mechanism (CoRM) improves MoE expert specialization by scoring against EMA reference instead of absolute magnitude, concentrating routing to separable subspace.

Trait-direction drift mechanism explains subliminal learning in model distillation; biased teacher generates clean data masking hidden preference transfer to student.

Modelpedia: automated LLM-assisted framework extracting and aggregating AI model findings from papers into searchable catalog, linking findings to models, datasets, methods.

StateSwap protocol probes LLM internal representations via untrained tokens to explain inconsistent MCQ answers under support vs. elimination framing.

TFMat applies text-conditioned flow matching to crystal structure generation; structured materials language semantic prior improves one-candidate accuracy on Perov-5, Carbon-24, MP-20.

Study on aligning LLM-as-judge outputs to human judgment distributions rather than aggregated labels, improving evaluation robustness.

CORD method for post-hoc calibration that preserves model predictions while correcting confidence scores in multiclass settings.

OUTLETS predicts LLM output lengths using speculative decoding backbones to optimize resource provisioning and cluster scheduling.

Artificial Rosetta Stone framework uses Markov models to reconstruct damaged symbolic raga music sequences via constrained inference.

Framework for deploying diffusion-model image generation on solar-powered satellites with energy constraints and compressed downlinking.

Method combining RL and gradient-based policy learning to optimize context-dependent MPC weights via implicit solver differentiation.

ARISE-RL framework for self-evolving agents via rubric-mediated co-evolution between task generator and solver, addressing sparse rewards.

CM-PTM pre-training method for cross-source multi-granular user behavior representation in mobile gaming personalization tasks.

WorldBench: multilingual agent benchmark with 1,600 persona-grounded tasks across 7 languages testing state preservation and cultural grounding.

SAGE uses subpopulation-aware generative enhancement to reduce spurious correlations in minority groups without explicit group labels.

Linear probes read internal variables from Pythia models at step 1,000 but steering along those directions remains ineffective; readability precedes causal efficacy independent of scale.

QXymb framework and QILP-0 system construct logic programs from quantum circuit behavior; niche specialized application with limited relevance to AI systems.