The Archive

Theoretical learning complexity of shuffle ideals via membership and contrastive queries in formal language theory.

High update-to-data ratio in SAC causes rank collapse and value divergence; early warning signals detected via magnitude tracking.

First-order response models predict per-fact forgetting margins in LoRA fine-tuning with 0.97+ correlation but fail at true forgetting prediction.

Clinical LLMs exhibit faithfulness gaps: high accuracy without correct reasoning. Introduces ESI, CFS, PSS metrics for explanation fidelity.

Associative algebra layers replace standard matrix multiplication in Transformers with sparser operations, achieving quadratic arithmetic complexity for fixed block sizes.

USAI-Quant benchmark evaluates vision-language models on quantitative reasoning over remote sensing imagery for urban/spatial AI applications.

OpenTumorBoard: 611-case benchmark from 12,534 minutes of real tumor board recordings evaluating LLMs on specialist turn prediction and board simulation.

Large-scale study reveals substantial LLM accuracy degradation as multiple-choice candidate sets scale, challenging conclusions from small-candidate benchmarks.

Analysis of massive activation spikes in vision-language models: text spikes appear consistently in early layers; visual spikes vary across model architectures.

Clinical LLM evaluation reveals six models fail counterfactual validity and demographic robustness tests: 22% of CFV attempts and significant demographic bias detected.

Framework quantifies information loss when LLM agents compress context into written state summaries, measuring write-time regret across decision sequences.

Temporal Affective Grounding reformulates audio emotion recognition from single labels to temporally bounded speech spans with vocal tone descriptions.

Nutri-ATLAS embodied agent integrates graph-based lookup and LLM reasoning for nutrition assistance in IoT healthcare, addressing ambiguous food descriptions.

Re-derivability metric predicts cost of upstream faults in staged LLM agent pipelines; inspector agents grounded in original problem show +0.6 recovery gain.

PlanGuard detects physical safety risks in multi-step embodied agent plans before execution using MSP-Safe dataset of compositional task interactions.

Sketched tangent consistency loss enables on-the-fly derivative-informed training for neural operators without offline label generation.

HABIT taxonomy characterizes 23 behavioral axes of LLM agents across 17 domains, distinguishing model preferences better than existing value taxonomies.

Analysis of anisotropy in LLM post-training reveals 5% of residual channels carry essential language modeling signal, with implications for optimization.

T⁵ twin-critic method improves token-level credit assignment in reinforcement mid-training by calibrating advantage estimates and reducing update drift.

FedHV applies closed-form hypervolume weighting to federated multi-objective optimization with heterogeneous data and communication constraints.

Motif prompt injection adds control handles to NotaGen-style music generation models for stylistically coherent composition with user-specified constraints.

CLAIRE hybrid workflow automates healthcare form completion via schema-grounded field mapping, validation, and audit tracing across EHR integration points.

Token-adaptive recursion mitigates catastrophic forgetting and class imbalance in ophthalmic domain incremental learning via selective model retraining.

CollisionGAT graph-attention network screens multi-agent motion plans for collisions, enabling controller-agnostic safety checks before robot step execution.

Proposes response-aware patient dynamics model for respiratory support prediction in critical care using explicit physiological change representation.

CT-OPD method for post-training diffusion vision-language models via counterfactual on-policy distillation with trajectory-aware feedback.

OmniMoE-VL sparse vision-language model with coupled visual-depth routing for dynamic visual representation selection across mixture-of-experts.

Copper-Policy learns compact world representations for robot manipulation via temporal joint-embedding prediction without pixel reconstruction.

Framework for monitoring agentic AI network traffic using hypersparse matrices, DBOS, PostgreSQL, and GraphBLAS for alignment auditing.

A-B-D framework infers LLM agent personality traits from behavioral data rather than self-reports or expensive LLM judge ratings.