The Steering Budget: Examples beat Knobs
New research suggests generative model steering is limited by training data, not model capacity, with 'examples' providing more control than 'knobs'.
arXiv cs.LG — Machine Learning·17 Jul
Search signals, briefings, company results, benchmarks and glossary terms.
Search signals, briefings, company results, benchmarks and glossary terms.
Use this view to inspect the underlying evidence corpus. For ranked developments, decision posture and interpretation, use Signals.
Raw feed or Signals?
Raw feed is chronological evidence. Signals ranks and interprets material change.
New research suggests generative model steering is limited by training data, not model capacity, with 'examples' providing more control than 'knobs'.
Research finds adaptive Runge-Kutta optimizers, when compute-matched, yield training loss but not generalization improvement over plain Adam.
Research details SAGA, a schema-aware grounding agent for converting natural language questions into SPARQL queries over knowledge bases.
Research proposes zero-shot quantization for object detection models using off-the-shelf generative models, addressing data access limitations.
Research explores optimal linear combinations of binary classifiers using truth tables to partition datasets and analyze convexified empirical risk.
New research proposes RENEW, a method to repair model exploitation in world models used in offline reinforcement learning by leveraging human preferences.
New research proposes an operational framework for understanding why closed-loop AI systems saturate and how external information can overcome this.
Research explores data augmentation methods for robust imitation learning in streamed video games, addressing network artifacts and data scarcity.
Research proposes a noise-robust framework integrating Inverse Reinforcement Learning (IRL) and Reinforcement Learning (RL) for eliciting risk preferences.
Research introduces CatalogAgent, a supervisor-mediated self-learning system that uses LLM-based generator-evaluator frameworks for product catalog data completion.
Research introduces Probabilistic Physics-Informed Neural Networks (PPINNs) to estimate heterogeneous elastic properties from low-resolution, noisy data.
Research introduces GeoDetect, a new method for detecting adversarial attacks on Vision-Language Pre-trained models (VLPs) by analyzing their embedding space geometry.
Research presents a novel system for image generation, integrating LLM-optimized negative prompts and latent-space classifier guidance to improve Stable Diffusion output quality.
Research details parameter-efficient prompt tuning of vision foundation models using adaptive focal loss for interpretable MCI screening from drawing tests.
DriftWorld introduces a new generative model for faster world modeling, enabling quicker rollouts for action planning in robotic control.
Research addresses domain adaptation challenges for plug-and-play image reconstruction denoisers used outside their training domains, impacting convergence.
New research proposes a novel 'diffusion distance' measure to quantify discrepancies between probability distributions on a graph, using Metropolis-Hastings.
Research explores optimal self-distillation for rectified flow generative models, proving a student can improve a suboptimal teacher.
Research introduces MeanFlowNFT, a forward-process reinforcement learning framework for efficient average-velocity generative models, enhancing alignment.
ConFlow is a research paper introducing a constraints-guided learning method for flow matching models used in robot motion generation.
COAT, a new framework combining counterfactual outcome estimation with mixed-integer optimization, learns interpretable prescriptive policies from observational data.
Research finds LLMs exhibit 'covert value leakage,' influencing answers based on internal values without disclosure, impacting sensitive queries.
Research explores human-in-the-loop ML for safe autonomous vehicles, addressing challenges in complex scenarios and data labeling bottlenecks.
New research extends policy learning methods for treatment allocation under budget constraints to address missing treatment data.
SafeOR-Gym introduces a new benchmark suite for safe reinforcement learning algorithms focused on operations research problems, unlike prior robotics benchmarks.
Research introduces a Dynamic, Automatic, and Systematic (DAS) red-teaming framework to continuously stress-test LLMs for health, addressing static benchmark obsolescence.
Research introduces a Weak Penalty Neural ODE method for improved forecasting of chaotic dynamical systems from noisy time series data.
A new research paper proposes a common mathematical framework and reporting checklist for local additive feature attribution methods in explainable AI.
LyaGuide proposes a Lyapunov-guided framework for stabilizing generative flow models, offering explicit stability guarantees for adapting pretrained flows.
Research proposes Generalized Fisher-Weighted SVD to efficiently compress large language models by approximating full Fisher information.
Research paper MIDiff proposes a multivariate-imaging diffusion model to generate synthetic mobile usage data, addressing sparsity and variable type heterogeneity.
AlphaWiSE, a new weight interpolation method, addresses catastrophic forgetting in multimodal models like CLIP by adapting to new data without losing prior alignments.
Research introduces one-shot generative design for disordered metamaterials using self-organizing neural cellular automata to overcome design complexity.
Researchers introduced SOReL, a fully offline Bayesian model-based Reinforcement Learning method for hyperparameter tuning and performance estimation.
Research details a statistical and economic framework for determining when to switch from an incumbent production model to a challenger model leveraging new data sources.
Research finds interleaved noise injection in stochastic optimization improves model performance across clean, corrupted, and out-of-distribution data.
New research establishes the first non-vacuous generalization bounds for parameter-efficient reinforcement learning with verifiable rewards (RLVR) fine-tuning.
New research proposes a continuous-time reinforcement learning framework for fine-tuning discrete diffusion models, deriving continuous-time PPO variants.
Research identifies data-parameter interference as a source of instability in federated fine-tuning of LLMs using LoRA, proposing dynamic subspace boosting.
Research questions the reliability of temporal graph learning benchmarks, finding simple heuristics often competitive with complex models.
Research on an Android app store shows increasing sponsored search ad load by up to 43% revenue gain can reduce total search conversions by 5%.
Research proposes a method for accelerating A/B tests through counterfactual estimation and policy overlap, reducing variance and cost.
Research introduces flow-based generative models with native extrapolation awareness to prevent silent failures in safety-critical prediction tasks.
Research demonstrates stable pretraining of Large Language Models using exclusively low-rank weights, matching dense model performance while reducing compute.
Research introduces Trajectory-Aware Flow Matching for generative topology optimization, aiming to reduce computational costs in design exploration.
Research proposes an adaptive layer-freezing strategy for energy-efficient federated learning in medical image conversion to reduce computational load.
Research proves 'grokking'—generalization long after overfitting—in ridge regression using gradient descent with weight decay.
ARTIST, a new method for adaptive time series reasoning, selects relevant segments rather than encoding entire sequences, improving efficiency.
Research explores how fairness-enhancing algorithms impact privacy leakage, a reverse analysis from prior work on privacy's effect on fairness.
New research proposes Angular Gaussian Supervised Contrastive Learning (AG-SCL) to improve deep learning reliability for long-tailed ECG arrhythmia diagnosis.
© 2026 OneBench: AI Insights. All rights reserved.
Evidence before opinion