New Approach to Anomaly Detection in 5G and Open RAN Telemetry
A proposed method targets deployment bottlenecks in 5G/Open RAN observability by addressing heterogeneous KPIs, class imbalance, and alert budgets.
Research Digest
Preprints and papers, rewritten in plain English.
Can AI weather forecasts stay reliable for extreme events?
A new study benchmarks four AI weather models against physics-based forecasts to see if they hold up in the tail of the distribution.
Diffusion Models Bring Probabilistic Forecasting to Limit Order Books
A new arXiv paper applies conditional diffusion models to limit order book forecasting, addressing a gap in existing snapshot models.
New Benchmark Tests RAG on Perovskite Solar Data
A proposed benchmark shifts evaluation of retrieval-augmented language models from qualitative overlap to structured, domain-specific extraction in perovskite photovoltaics.
PRISM uses prior relational info to boost solubility prediction
A new self-supervised model improves out-of-distribution generalization for solubility, a key property in drug discovery and chemical engineering.
BatchParScale Proposes Sparse Mixing for Parallel Test-Time Scaling
A new arXiv paper introduces a sparse batch-mixing method to improve parallel sampling and aggregation in sequence models.
Benign Fine-Tuning Can Still Misalign LLMs, Study Finds
New research challenges the common assumption that only harmful fine-tuning data poses safety risks, showing that benign fine-tuning can also cause misalignment.
Adversarial Post-Training Targets Compounding Errors in Trading Models
A new approach applies adversarial post-training to limit order book foundation models, addressing the compounding sampling errors left unpenalized by standard pre-training.
Factorized Tabular Foundation Models Target Architecture Trade-off
A new paper proposes a factorized tabular foundation model to overcome a trade-off in current architectures.
LLMs Aim to Capture Human Irrationality in Prisoner's Dilemma
A new arXiv paper uses large language models to model how human decisions in the Prisoner's Dilemma diverge from rational choice.
RASTA framework targets style loss in machine translation
A new open-weight framework aims to preserve linguistic style in machine translation, especially for non-Western languages.
Curriculum Learning Helps LLMs Reason Implicitly, Up to a Limit
A new arXiv study finds that curriculum learning can teach large language models to reason implicitly over long horizons, but the approach still hits performance cliffs.
Demographic Injection Aims to Improve LLM Survey Simulation
A new arXiv paper tackles the challenge of making LLM-generated survey respondents reflect real population diversity.
Divergence-Based Method Aligns LLMs with Ranked Preferences
A new approach uses divergence measures to align large language models with ranked feedback, capturing richer signal than pairwise comparisons.
MOSAIC: Per-Layer KV Cache Allocation for Long-Context LLMs
A new search-based approach allocates inference cache budgets across transformer layers instead of using a uniform budget.
Probing Whether Language Models Grasp Variable Roles in Code
A new arXiv study uses an established programming taxonomy to test which variable-role distinctions language models can recover from code representations.
Distributionally Robust Policies for Real-World Deployment
A new framework aims to keep learned policies from failing when deployment conditions differ from the observational data they were trained on.
Prefix-Tuning Offers Parameter-Efficient Control of LLM Outputs
A new arXiv paper describes prefix-tuning, a method that steers pretrained transformers by prepending task-specific prefixes rather than retraining all parameters.