Fathom-Vaidya: Advancing Medical Reasoning with Rubric-Based Rewards
本文针对医疗领域大语言模型在诊断推理和临床交互中的不足，通过合成数据和基于评分标准的强化学习方法提升了模型在这两方面的能力。
0 citationsRead paper
Selected work
Representative Papers
Mixture of Chapters: Scaling Learnt Memory in Transformers
This work addresses the limitation of Transformers in lacking an explicit knowledge storage mechanism, which hinders efficient retention and retrieval of learned information. To overcome this, the authors propose a chaptered sparse memory bank, where learnable memory tokens are queried by the Transformer via cross-attention. Inspired by Mixture-of-Experts, a dynamic chapter routing strategy selectively activates relevant subsets of memory, enabling scalable knowledge access while maintaining computational efficiency. The approach expands memory capacity to 262K tokens—introducing a new scaling dimension beyond model parameters—without incurring prohibitive computational overhead. Experiments demonstrate that, under matched FLOPs, the proposed model outperforms standard Transformers in both pretraining and instruction fine-tuning tasks, while also exhibiting substantially improved knowledge retention and robustness against catastrophic forgetting in continual learning scenarios.
0 citationsRead paper
Sustainable NARMA-10 Benchmarking for Quantum Reservoir Computing
This study addresses the nonlinear time-series prediction task of the NARMA-10 benchmark. We systematically compare quantum reservoir computing (QRC), classical echo state networks (ESNs), long short-term memory (LSTM) networks, and a hybrid quantum-classical QLSTM under a unified experimental framework. Prediction accuracy is evaluated via normalized root-mean-square error (NRMSE), while computational resource consumption and inference latency are rigorously quantified—establishing, for the first time, a sustainability-oriented evaluation paradigm for quantum time-series modeling. Results demonstrate that QRC achieves prediction accuracy comparable to LSTM and ESN, yet with substantially reduced parameter count, memory footprint, and forward-inference latency—yielding superior energy efficiency. This work not only validates QRC’s practical viability in resource-constrained edge environments but also pioneers a green, low-overhead pathway for quantum-enhanced time-series modeling.
0 citationsRead paper
Recent publications
Latest Papers
Fathom-Vaidya: Advancing Medical Reasoning with Rubric-Based Rewards
本文针对医疗领域大语言模型在诊断推理和临床交互中的不足，通过合成数据和基于评分标准的强化学习方法提升了模型在这两方面的能力。
0 citationsRead paper
Mixture of Chapters: Scaling Learnt Memory in Transformers
This work addresses the limitation of Transformers in lacking an explicit knowledge storage mechanism, which hinders efficient retention and retrieval of learned information. To overcome this, the authors propose a chaptered sparse memory bank, where learnable memory tokens are queried by the Transformer via cross-attention. Inspired by Mixture-of-Experts, a dynamic chapter routing strategy selectively activates relevant subsets of memory, enabling scalable knowledge access while maintaining computational efficiency. The approach expands memory capacity to 262K tokens—introducing a new scaling dimension beyond model parameters—without incurring prohibitive computational overhead. Experiments demonstrate that, under matched FLOPs, the proposed model outperforms standard Transformers in both pretraining and instruction fine-tuning tasks, while also exhibiting substantially improved knowledge retention and robustness against catastrophic forgetting in continual learning scenarios.
0 citationsRead paper
Sustainable NARMA-10 Benchmarking for Quantum Reservoir Computing
This study addresses the nonlinear time-series prediction task of the NARMA-10 benchmark. We systematically compare quantum reservoir computing (QRC), classical echo state networks (ESNs), long short-term memory (LSTM) networks, and a hybrid quantum-classical QLSTM under a unified experimental framework. Prediction accuracy is evaluated via normalized root-mean-square error (NRMSE), while computational resource consumption and inference latency are rigorously quantified—establishing, for the first time, a sustainability-oriented evaluation paradigm for quantum time-series modeling. Results demonstrate that QRC achieves prediction accuracy comparable to LSTM and ESN, yet with substantially reduced parameter count, memory footprint, and forward-inference latency—yielding superior energy efficiency. This work not only validates QRC’s practical viability in resource-constrained edge environments but also pioneers a green, low-overhead pathway for quantum-enhanced time-series modeling.
0 citationsRead paper