arXiv subjects

Changyi Yang

Publications and source records attributed to Changyi Yang.

6 recordsLinked to original sources

arXiv2026

Efficient and Trainable Language Model Test-Time Scaling via Local Branch Routing

Test-time scaling improves language-model reasoning, but existing approaches often face a difficult trade-off: long chain-of-thought sampling remains single-threaded, while sentence- or solution-level search can be computationally expensive and hard to train end-to-end. We introduce Local Branch Routing (LBR), a token-level test-time scaling framework that expands a small local lookahead tree, forwards all sampled branches through the language model, and uses a lightweight router to select the depth-1 subtree to commit. By routing over the hidden states of candidate local futures, LBR allows each token decision to use evidence beyond the root next-token distribution while avoiding full solution-level search. The resulting prune-shift-grow decoding process preserves discrete branch identities and defines a tractable tree-trajectory likelihood: newly grown nodes are counted when first sampled, and router decisions are assigned explicit probabilities. This enables end-to-end reinforcement learning with verifiable rewards, jointly optimizing the base model and router under the same likelihood-ratio principle as discrete-token RLVR. On synthetic hierarchical-planning tasks, LBR shows that post-candidate hidden states provide useful routing evidence. On mathematical reasoning benchmarks, LBR improves both Pass@1 and Pass@32 over discrete chain-of-thought, vanilla discrete-token RLVR, and RL-compatible soft-token branching baselines. These results suggest that lightweight local branching offers an efficient, trainable, and discrete form of language-model test-time scaling.

cs.CL↗

arXiv2026

Length Value Model: Scalable Value Pretraining for Token-Level Length Modeling

Tokens are the fundamental units of computation in modern autoregressive models, and generation length directly influences both inference cost and reasoning performance. Despite its importance, existing approaches model length primarily at the coarse sequence level. We introduce the Length Value Model (LenVM), a token-level framework that estimates the remaining generation length at every decoding step. By formulating length modeling as a value estimation problem and assigning a constant negative reward to each generated token, LenVM predicts a bounded, discounted return that is a monotone proxy for the remaining generation horizon. This value formulation provides annotation-free, dense, unbiased, and scalable supervision. Experiments on LLMs and VLMs show that LenVM supports exact control, continuous performance--efficiency steering, length prediction, and interpretation. On LIFEBench-token, it raises the exact-length score of Qwen2.5-7B-Instruct from $30.9$ to $64.8$ in one pass and reaches $83.6$ when combined with LCG. Near $200$ tokens on GSM8K, LenVM retains about $63\%$ Pass@1 versus $6\%$ under a hard token budget. Across other settings, LenVM improves matched-length quality over prompting and EOS calibration, while its length-prediction accuracy improves with scale and exceeds specialized predictors. Its token-level values also identify markers of shifts between longer and shorter trajectories. These results establish generation length as an effective scalable token-level value signal for control, prediction, and future value-based training.

cs.CL↗

arXiv2014

Modeling of ion beam induced charge sharing experiments for the design of high resolution position sensitive detectors

In a multi-electrode device, the motion of free charge carriers generated by ionizing radiation induces currents on all the electrodes surrounding the active region [1]. The amount of charge induced in each sensitive electrode is a function of the device geometry, the transport parameters and the generation profile. Hence this charge sharing effect allows the signal from each sensitive electrode to provide information about the electrical characteristics of the device, as well as information on the location and the profile of each ionization track. The effectiveness of such approach was recently demonstrated in Ion Beam Induced Charge (IBIC) experiments carried out using a 2 MeV He microbeam scanning over a sub-100 lm scale silicon device, where the ion strike location point was evaluated through a comparative analysis of the charge induced in two independent surface electrodes coupled to independent data acquisition systems [2]. In this report, we show that the Monte Carlo method [3] can be efficiently exploited to simulate this IBIC experiment and to model the experimental data, shedding light on the role played by carrier diffusion, electronic noise and ion beam spot size on the induction of charge in the sensitive electrodes. Moreover, the Monte Carlo method shows that information on the ion strike position can be obtained from the charge signals from the sensitive electrodes.

cond-mat.mtrl-sci↗

arXiv2010

Single-shot readout of an electron spin in silicon

The size of silicon transistors used in microelectronic devices is shrinking to the level where quantum effects become important. While this presents a significant challenge for the further scaling of microprocessors, it provides the potential for radical innovations in the form of spin-based quantum computers and spintronic devices. An electron spin in Si can represent a well-isolated quantum bit with long coherence times because of the weak spin-orbit coupling and the possibility to eliminate nuclear spins from the bulk crystal. However, the control of single electrons in Si has proved challenging, and has so far hindered the observation and manipulation of a single spin. Here we report the first demonstration of single-shot, time-resolved readout of an electron spin in Si. This has been performed in a device consisting of implanted phosphorus donors coupled to a metal-oxide-semiconductor single-electron transistor - compatible with current microelectronic technology. We observed a spin lifetime approaching 1 second at magnetic fields below 2 T, and achieved spin readout fidelity better than 90%. High-fidelity single-shot spin readout in Si opens the path to the development of a new generation of quantum computing and spintronic devices, built using the most important material in the semiconductor industry.

cond-mat.mes-hall↗

arXiv2010

Transport Spectroscopy of Single Phosphorus Donors in a Silicon Nanoscale Transistor

We have developed nano-scale double-gated field-effect-transistors for the study of electron states and transport properties of single deliberately-implanted phosphorus donors. The devices provide a high-level of control of key parameters required for potential applications in nanoelectronics. For the donors, we resolve transitions corresponding to two charge states successively occupied by spin down and spin up electrons. The charging energies and the Lande g-factors are consistent with expectations for donors in gated nanostructures.

cond-mat.mes-hall↗

arXiv2009

Electron tunnel rates in a donor-silicon single electron transistor hybrid

We investigate a hybrid structure consisting of $20\pm4$ implanted $^{31}$P atoms close to a gate-induced silicon single electron transistor (SiSET). In this configuration, the SiSET is extremely sensitive to the charge state of the nearby centers, turning from the off state to the conducting state when the charge configuration is changed. We present a method to measure fast electron tunnel rates between donors and the SiSET island, using a pulsed voltage scheme and low-bandwidth current detection. The experimental findings are quantitatively discussed using a rate equation model, enabling the extraction of the capture and emission rates.

cond-mat.mes-hall↗