University of Sheffield

5 related articles

Subscribe to updates

aclanthology.org Oct 31, 2026
Chain-of-thought need not fire on every query: confidence gating cuts reasoning use by 30% without losing accuracy when the signal is strong
Chain-of-thought prompting improves complex reasoning but spends extra computation even on simple queries. A Sheffield team tested "think only when unsure" gating across four training-free confidence signals, three model families and seven reasoning benchmarks, finding that effectiveness varies widely by model and task and tracing the variation to confidence-score saturation and a selection bias that sends the hardest queries to chain-of-thought. Their Pareto-optimal threshold selection reduces chain-of-thought usage by about 30% without sacrificing accuracy where the confidence signal is strong, and falls back toward full chain-of-thought where it is weak.
Chain-of-thought prompting improves complex reasoning but spends extra computation even on simple queries. A Sheffield team tested "think only when unsure" gating across four training-free confidence signals, three model families and seven reasoning benchmarks, finding that effectiveness varies widely by model and task and tracing the variation to confidence-score saturation and a selection bias that sends the hardest queries to chain-of-thought. Their Pareto-optimal threshold selection reduces chain-of-thought usage by about 30% without sacrificing accuracy where the confidence signal is strong, and falls back toward full chain-of-thought where it is weak.
University of Sheffield
Confidence-Gated Chain-of-Thought Pareto-Optimal Threshold Selection Uncertainty Estimation Efficient Reasoning
arXiv Oct 29, 2026
Multi-domain fine-tuning stops stepping on itself: 256 examples per domain lift expert routing to 90–98%
Dense Transformer layers are expanded into sparse mixture-of-experts layers, giving law and biomedicine their own full MLP experts while a router is supervised with a small number of examples to pick the right expert per domain. On Llama-3.1-8B and Qwen-3-8B the method posts the highest biomedical and legal averages and the highest general-domain average; with only 256 examples per domain for router supervision, routing accuracy on specialized domains reaches 90–98% after adaptation, against 43–70% for routers trained without it.
Dense Transformer layers are expanded into sparse mixture-of-experts layers, giving law and biomedicine their own full MLP experts while a router is supervised with a small number of examples to pick the right expert per domain. On Llama-3.1-8B and Qwen-3-8B the method posts the highest biomedical and legal averages and the highest general-domain average; with only 256 examples per domain for router supervision, routing accuracy on specialized domains reaches 90–98% after adaptation, against 43–70% for routers trained without it.
University of Sheffield
Mingzi Cao（First author）
Domain-Specialized Upcycling Mixture of Experts Domain Adaptation Catastrophic Forgetting
researchsquare.com Oct 27, 2026
Assembly order stops needing after-the-fact repair: a constraint-preserving planner returns the exact best remaining sequence in all 128 recoverable disturbance cases
In modular assembly the robot's task order must respect assembly precedence and be reordered when conditions change after work begins. Constraint-Aware Genetic Task Planning (CGTP) keeps initialization and variation inside the precedence-feasible region and, on replanning, freezes completed operations and re-optimizes only the remaining task graph. On an exhaustively verifiable eight-task benchmark it returned the exact best feasible suffix in all 128 structurally recoverable scenarios and flagged all 3 unrecoverable ones; on a 90-instance corpus, per-class recovery ran from 0.881 to 1.000 with no budget-conditional search failures.
In modular assembly the robot's task order must respect assembly precedence and be reordered when conditions change after work begins. Constraint-Aware Genetic Task Planning (CGTP) keeps initialization and variation inside the precedence-feasible region and, on replanning, freezes completed operations and re-optimizes only the remaining task graph. On an exhaustively verifiable eight-task benchmark it returned the exact best feasible suffix in all 128 structurally recoverable scenarios and flagged all 3 unrecoverable ones; on a 90-instance corpus, per-class recovery ran from 0.881 to 1.000 with no budget-conditional search failures.
University of Hertfordshire Loughborough University University of Sheffield
Yiping Meng（First author, Corresponding author）
Constraint-Aware Genetic Task Planning (CGTP)Assembly Sequence Planning Disturbance-Triggered Partial Replanning Evolutionary Search
International Journal of Computer Assisted Radiology and Surgery Oct 17, 2026
Alzheimer's classifiers stop using age as a shortcut: residualizing age in feature space nearly zeroes the age-prediction link in cognitively normal and mild cognitive impairment groups
On structural MRI from three independent cohorts, Alzheimer's disease classifiers use age as a shortcut: older patients receive higher predicted risk. Residualizing age linearly in feature space (PMDN) brings the age-prediction correlation near zero in cognitively normal and mild cognitive impairment groups, with balanced accuracy 0.71 and Brier score about 0.22, and far less fold-to-fold spread than the adversarial CF-Net. Post hoc calibration further lowers probability error, with nonparametric isotonic regression and Bayesian binning performing best.
On structural MRI from three independent cohorts, Alzheimer's disease classifiers use age as a shortcut: older patients receive higher predicted risk. Residualizing age linearly in feature space (PMDN) brings the age-prediction correlation near zero in cognitively normal and mild cognitive impairment groups, with balanced accuracy 0.71 and Brier score about 0.22, and far less fold-to-fold spread than the adversarial CF-Net. Post hoc calibration further lowers probability error, with nonparametric isotonic regression and Bayesian binning performing best.
Charité - Universitätsmedizin Berlin Otto-von-Guericke-Universität Magdeburg University of Aveiro University of Cologne University of Edinburgh University Hospital Bonn University of Göttingen Universitätsmedizin Göttingen
Majid Ramedani（First author）Martin Dyrba（Corresponding author）
PMDN CF-Net Probability Calibration adversarial training Convolutional Neural Networks
arXiv Oct 1, 2026
Scaling synthetic pre-pretraining from 1B to 7B and 100B tokens keeps the token savings, but the authors trace them to long-range retrieval rather than a grammatical prior
The study systematically evaluates synthetic pre-pretraining (PPT) across four parameter scales from 500M to 7B, four pre-training data mixtures, five PPT tasks, and pre-training budgets up to 100B tokens, finding that downstream performance and token-efficiency gains persist at scale (saving at least 21B pre-training tokens at 3B) but that there is no consistent evidence the gains come from a grammatical prior; instead they arise from PPT tasks that improve long-range retrieval, are insensitive to code and math share, and diminish only when web text is absent.
The study systematically evaluates synthetic pre-pretraining (PPT) across four parameter scales from 500M to 7B, four pre-training data mixtures, five PPT tasks, and pre-training budgets up to 100B tokens, finding that downstream performance and token-efficiency gains persist at scale (saving at least 21B pre-training tokens at 3B) but that there is no consistent evidence the gains come from a grammatical prior; instead they arise from PPT tasks that improve long-range retrieval, are insensitive to code and math share, and diminish only when web text is absent.
Mohamed bin Zayed University of Artificial Intelligence Hugging Face University of Sheffield
Atsuki Yamaguchi（First author）
Synthetic Pre-pretraining Long-range Retrieval Language model pretraining

You’re all caught up

Chain-of-thought need not fire on every query: confidence gating cuts reasoning use by 30% without losing accuracy when the signal is strong

Multi-domain fine-tuning stops stepping on itself: 256 examples per domain lift expert routing to 90–98%

Assembly order stops needing after-the-fact repair: a constraint-preserving planner returns the exact best remaining sequence in all 128 recoverable disturbance cases

Alzheimer's classifiers stop using age as a shortcut: residualizing age in feature space nearly zeroes the age-prediction link in cognitively normal and mild cognitive impairment groups

Scaling synthetic pre-pretraining from 1B to 7B and 100B tokens keeps the token savings, but the authors trace them to long-range retrieval rather than a grammatical prior