Khiem Le, Phuc Nguyen, Youssef Mroueh, Chi-Heng Lin, Shangqian Gao, Ting Hua, Nitesh V. Chawla
Group Relative Policy Optimization (GRPO) has become the dominant reinforcement learning method for eliciting reasoning capabilities in LLMs using verifiable rewards. Still, it suffers from two critical issues that limit models' exploration: \textbf{Gradient vanishing} and \textbf{Diversity collapse}. When a training question is too easy or too hard, all sampled responses receive identical rewards, yielding no gradient signal. Meanwhile, the model tends to collapse its responses toward a single reasoning pattern rather than exploring diverse strategies. We propose Transformation-Augmented GRPO (TA-GRPO), a simple yet effective method that addresses both issues via question transformation. For each training question, we automatically create multiple problem-equivalent rephrased questions that alter wording, format, and information order while preserving the underlying intent. As these rephrasings shift the model's perceived difficulty, responses across the original question and its rephrasings receive mixed rewards and encode more diverse reasoning strategies within it. TA-GRPO jointly computes advantages over this expanded response group and aligns all importance sampling ratios to the original question, enabling the model to learn from a richer set of attempts. Experiments with four LLMs (Qwen3-1.7B/4B-Base and Llama-3.2-1B/3B-Instruct) show that TA-GRPO consistently improves pass@$k$ on competition-level mathematics benchmarks (AMC, OlympiadBench, AIME24, AIME25) and out-of-distribution benchmarks (Minerva, GPQA-Diamond). Notably, it improves the average pass@32 of Qwen3-1.7B-Base and Qwen3-4B-Base by \textbf{4.97}\% and \textbf{4.35}\%, respectively, and matches the pass@32 of baselines trained with up to 2.5x more data.
Vu Le, Yilun Xu, Neel Vora, Gang Huang, Deepak Ganesan, Phuc Nguyen
Qubit readout, or qubit state discrimination, remains one of the slowest and most error-prone operations in superconducting quantum processors, making fast, accurate, and hardware-efficient classification essential for scalable quantum computing. Conventional discriminators integrate the I/Q record into a single point, discarding evidence a relaxing qubit leaves behind; full-resolution feed-forward classifiers recover it but are too costly to deploy under this constraint. Recurrent classifiers retain full resolution far more cheaply, but their loop-carried hidden state cannot fit a single AI Engine (AIE) tile's register capacity and spills to data memory, adding latency that grows with sequence length. We present a hardware-aware LSTM accelerator for an AMD Versal device that resolves this by partitioning at the dependency boundary - independent input projections on the AIE array, the loop-carried update in programmable logic (PL) - keeping pace with a 64 ns input cadence under a 1,000 ns budget. On the qubit affected by mid-readout state transitions, our design matches a full-resolution feed-forward network's accuracy (96.59% vs. 96.37%) using 403x fewer parameters and roughly 103x fewer DSPs while cutting relative error by 68.2% over GMM, 43.6% over QubiCML, and 16.82% over MF-NN. Using 4 out of 400 AIE tiles and 17 DSP58, it needs 2.2-8.2x fewer DSP58 blocks than QubiCML and MF-NN, and the AIE-PL partition cuts exposed tail latency by 60.3% versus a pure-AIE mapping and DSP58 usage 31.1x versus an all-PL implementation. Across eight qubits, a compact variant reaches 96.78% mean accuracy, the strongest hardware-deployed operating point. Temporal resolution and recurrent-state placement must then be treated jointly when designing low-latency quantum readout accelerators.
In virtual environments, choosing appropriate outfits requires contextual harmony with surrounding scenes, yet bridging the visual gap between 3D ambient views and catalog garment assets remains challenging. This paper investigates scene-conditioned outfit retrieval for interactive 3D environments, where a user's egocentric view serves as the retrieval signal. We present GEAR (Generative Editing for Asset Retrieval), a framework that bridges the cross-domain gap by synthesizing intermediate visual queries rather than predicting direct compatibility. GEAR uses a vision-language model to describe suitable clothing from a captured scene, employs an instruction-guided diffusion model to render the concept onto a standardized reference character, and utilizes the cropped clothing region to retrieve existing 2D outfit assets via embedding similarity. An interactive Unity prototype demonstrates real-time scene capture, asset retrieval, and in-engine 3D avatar fitting. We benchmark GEAR against direct visual similarity, aesthetic scoring, and VLM-based reranking on a controlled dataset using automated Swiss-system pairwise judgments, where GEAR achieves competitive top-ranked performance (mean Elo@5 of 1012.75 and top-10 overlap of 0.180) alongside 0.539 pairwise agreement. Furthermore, a user study across 20 virtual travel scenes demonstrates that GEAR achieves the highest overall participant preference in both visual appeal (34.5%) and contextual suitability (34.5%).
We present HyVox, a hybrid neuromorphic--von Neumann system for low-energy open-set speaker identification under tight energy budgets. HyVox splits the workload at the speaker embedding. A neuromorphic processor embeds the speech with sparse, event-driven computation, and a von Neumann microcontroller holds the speaker gallery and makes every decision: it scores queries by cosine similarity, rejects unknown speakers, and enrolls new users by writing a prototype, with no weight update. On the neuromorphic processor, a hierarchical temporal abstraction (HTA) SNN backbone learns a synaptic time constant per leaky integrate-and-fire (LIF) neuron, allocating 832 neurons across fast, medium, and slow timescales. A 32-neuron spiking embedding extractor (SEE), pretrained on episodes that each enroll new speakers and reject outsiders, turns the backbone output into an embedding for speakers unseen in training; its 32 spike counts are the only interface between the two processors. To fit the chip's synapse and integer budgets, we compress the backbone 6.5x to 29.0k synapses, train it for integer bit-shift dynamics, and fine-tune it with chip-accurate spiking training (CAST). CAST trains through a differentiable replica of the chip's integer neuron update, derived from its cycle-accurate simulator, so that gradient descent optimizes the exact weights and thresholds the chip runs; it doubles the top-1 accuracy of the mapped backbone on unseen speakers. On 20 Common Voice households of unseen speakers, HyVox reaches 99.6\% top-1 and 5.5\% rejection EER with 2 members and 94.9\% and 13.2\% with 20, and the network deployed on the neuromorphic chip matches its trained model. Each inference on a 3-s window takes 0.96 s and 1.63 mJ of dynamic energy, 42\% lower latency than the fastest embedded ANN baseline and 13.5x less dynamic and 5.8x less total energy than the most energy-efficient one.