Towards Certified Unlearning for Deep Neural Networks
Research proposes techniques to extend certified unlearning methods to deep neural networks, addressing challenges in highly nonconvex models.
arXiv cs.LG — Machine Learning·24 Apr
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.
Research proposes techniques to extend certified unlearning methods to deep neural networks, addressing challenges in highly nonconvex models.
Researchers propose F²LP-AP, a fast and flexible label propagation method for semi-supervised node classification, addressing GNN computational overhead and homophily assumptions.
Research identifies 'precision-induced output disagreements' in LLMs due to varying numerical precision (e.g., bfloat16, int8) during deployment.
Research proposes adaptive conformal anomaly detection for time series, leveraging pre-trained foundation models without fine-tuning, yielding interpretable p-value anomaly scores.
Research identifies training-time interventions to improve neural network calibration, addressing overconfidence in predictions without post-hoc adjustments.
Research indicates that co-locating tests with code improves foundation model code generation quality across multiple models and providers.
Research explores using generative models to create synthetic flight diversion records, addressing data imbalance for predictive model training.
MIRROR benchmark evaluates 16 LLMs across 8 labs on metacognitive calibration, assessing self-knowledge for decision-making.
Research presents a unified theory for sparse dictionary learning in mechanistic interpretability, addressing piecewise biconvexity and spurious minima.
Researchers propose ridge spectral sparsification to improve large-scale graph learning in distributed streaming settings.
Research paper details performance analysis and optimization of a BentoML-based AI inference system for scalable model serving, in collaboration with graphworks.ai.
Research proposes Differentiable Conformal Training (DCT) to provide statistically valid confidence guarantees for LLM factuality, reducing hallucinations.
PayPal empirically evaluated speculative decoding with EAGLE3 on a fine-tuned Llama 3.1-Nemotron model for its Commerce Agent, showing inference speedups.
LiteParse, an open-source tool for PDF text extraction, now runs entirely in the browser using standard PDF parsing and OCR, without AI models.
Google's Threat Intelligence notes indirect prompt injection (IPI) is a top security priority, but actual exploitation in the wild remains limited.
The 'One Useful Thing' newsletter speculates on a hypothetical GPT-5.5 model, suggesting incremental advancements in capability.
OpenAI's GPT-5.5 model is rolling out via ChatGPT and a semi-official Codex backdoor API, with the primary API release delayed for safeguards.
The UK National Cyber Security Centre (NCSC) released guidance on the careful oversight and new capabilities needed for AI adoption in cyber defence.
International cyber agencies, including the NCSC, issued new guidance on defending against China-linked covert network tactics in cyber activity.
The UK National Cyber Security Centre (NCSC) details a widespread shift to covert, China-nexus compromised device networks for cyberattacks and defense strategies.
OpenAI announced GPT-5.5, claiming it is their smartest, fastest model, designed for complex tasks including coding, research, and data analysis.
OpenAI published a 'System Card' for GPT-5.5, a speculative future model, detailing anticipated safety and alignment considerations.
Expert commentary deep dives into advanced AI research covering agent reasoning, multimodal generation, and humanoid learning techniques.
LoRA-FA proposes an improved parameter-efficient fine-tuning method, enhancing LoRA by addressing its performance limitations on certain tasks.
Research proposes neural bandit for optimal LLM selection across subtasks in an agentic pipeline, aiming for cost-efficient success.
Researchers introduced AVISE, a modular open-source framework for identifying vulnerabilities and evaluating the security of AI systems.
Research introduces Task-Stratified Knowledge Scaling Laws to analyze how Post-Training Quantization (PTQ) differentially impacts LLM memorization, application, and reasoning capabilities.
BatchLLM is a research paper optimizing large-batched LLM inference by exploiting global prefix sharing and throughput-oriented token batching.
ActuBench proposes a multi-agent LLM pipeline to generate and evaluate actuarial reasoning tasks aligned with IAA syllabus, using distinct LLM roles.
Research proposes a continuous semantic caching framework for LLM serving to reduce inference costs and latency by reusing responses to semantically similar queries.
Research identifies distinct internal model features influencing LLM confidence versus actual correctness via sparse autoencoders.
Research explored using small language models' self-reported numerical confidence for routing in cascade systems, escalating uncertain tasks to larger models.
Research proposes a joint stochastic approximation method to improve end-to-end training and optimization for Retrieval-Augmented Generation (RAG) models.
Research investigates quantization robustness of diffusion-based language models (d-LLMs) for coding tasks, focusing on memory and inference cost reduction.
Research re-evaluates Human Label Variation (HLV) in NLP, suggesting it's a signal for model robustness, especially with LLM post-training.
SkillGraph uses a directed weighted execution-transition graph from 49,831 tool sequences to improve LLM agent tool selection and ordering, addressing data dependencies.
Research finds Transformers can infer transitive relations in some graph structures but fail in others, impacting causal reasoning. arXiv paper.
Research claims retrofitting smaller, 300M parameter multilingual models can achieve 7B model performance in retrieval tasks.
Research indicates LLMs develop universal sinusoidal representations for numbers, largely interchangeable across different model architectures.
Research proposes LoID, a method to extract informative prior distributions from LLMs for Bayesian logistic regression, improving generalization on small datasets.
LLMs prioritize surface cues over implicit constraints, showing systematic failure in reasoning tasks like the 'car wash problem' due to sigmoid heuristics.
Research paper introduces CodedLang dataset of 7,744 Chinese Google Maps reviews to improve LLM handling of coded language.
Research analyzes LLM 'over-refusal' by mapping internal refusal mechanisms to specific representation subspaces to mitigate unwarranted safety denials.
Research found LLM-generated resume summaries exhibit race-gender bias based on candidate names, even when grounded in identical synthetic resumes.
Research indicates AI-generated text detectors often fail beyond benchmarks, exploiting dataset biases rather than true machine authorship signals.
Research proposes Plackett-Luce (PLR) model to reorder in-context learning examples, improving LLM performance by optimizing example sequence.
SpeechParaling-Bench introduces a new benchmark for evaluating paralinguistic cues in Large Audio-Language Models, covering over 100 features.
Research paper systematically evaluates intersectional fairness across six LLMs using ambiguous and disambiguated contexts from two benchmark datasets.
Research explores prompt optimization and judge selection for LLM-as-a-Judge evaluations in legal QA, assessing transferability across judges.
AstaBench proposes a new benchmark suite for evaluating AI agents across scientific research tasks, including literature review and data analysis.
© 2026 OneBench: AI Insights. All rights reserved.
Evidence before opinion