arXiv subjects

Muzhi Yang

Publications and source records attributed to Muzhi Yang.

5 recordsLinked to original sources

arXiv2026

Testing Processor Contracts in Speculative Decoding Engines

Speculative sampling can accelerate decoding without changing the target distribution. Its standard guarantee assumes that logits processors, which enforce constraints or modify token scores, receive the appropriate token prefix and state at each verification position. We formalize this processor-prefix contract and present BADC (branch-aware differential conformance testing). BADC combines enumerated references on tiny random models, checks of processor inputs and state, and simultaneous confidence bounds on output-event probabilities. Preregistered, source-informed experiments on selected Transformers and llama-cpp-python releases confirm two causes: processors read unverified drafts or advance state without rollback. For one pilot-selected prompt on a trained checkpoint, fresh samples give a 60.26-percentage-point lower bound on the difference in the probability of a predefined first-token event between ordinary and prompt-lookup decoding, with nominal simultaneous 95% coverage under within-arm i.i.d. sampling. Controlled comparisons establish no cause-recovery advantage of directed selection over random selection, while permutation testing detects each tested family at least as often as BADC. These findings connect processor-interface failures to observable decoding differences without establishing diagnostic superiority or workload prevalence.

cs.SE↗

arXiv2026

Re-executing Coding-Agent Evaluations: Unchanged Decisions and Unresolved Policy Risk

Executable tests grade large language model (LLM) coding-agent patches for leaderboards, rewards and self-improvement. Two preregistered re-execution studies returned negative results. SWE-bench Pro v1's offline option rejected none of 59 accepted Go patches whose grading configuration changed. Their exploratory, unweighted upper confidence limit is 4.95% under candidate independence, leaving a 3% exposed-patch risk unresolved; the preregistered estimator uses inverse sampling-probability weights. An extension was inconclusive. In a selected stratum, a reconstructed module cache built from the reference solution's dependencies rejected accepted alternatives from four models that passed with their own dependencies. Re-executing 1,929 selected agent-version/task evaluations from the Darwin Gödel Machine's archive preserved best-agent selection and system-variant ranking, with search history and scores of evaluations not rerun fixed. Exploratory host, image and adapted-grader comparisons also preserved them. Seven of 23 disagreements with archived verdicts did not recur on a second host; rebuilt images changed 16 three-run majority verdicts (0.83%) under four-CPU execution. We provide an evidence-to-claim summary, lessons for grader audits and an anonymous replication package. These results concern accepted patches and fixed-history decisions; general policy safety and adaptive-learning robustness require broader evaluations.

cs.SE↗

arXiv2016

Functional Hypergraph Uncovers Novel Covariant Structures over Neurodevelopment

Brain development during adolescence is marked by substantial changes in brain structure and function, leading to a stable network topology in adulthood. However, most prior work has examined the data through the lens of brain areas connected to one another in large-scale functional networks. Here, we apply a recently-developed hypergraph approach that treats network connections (edges) rather than brain regions as the unit of interest, allowing us to describe functional network topology from a fundamentally different perspective. Capitalizing on a sample of 780 youth imaged as part of the Philadelphia Neurodevelopmental Cohort, this hypergraph representation of resting-state functional MRI data reveals three distinct classes of sub-networks (hyperedges): clusters, bridges, and stars, which represent spatially distributed, bipartite, and focal architectures, respectively. Cluster hyperedges show a strong resemblance to the functional modules of the brain including somatomotor, visual, default mode, and salience systems. In contrast, star hyperedges represent highly localized subnetworks centered on a small set of regions, and are distributed across the entire cortex. Finally, bridge hyperedges link clusters and stars in a core-periphery organization. Notably, developmental changes within hyperedges are ordered in a similar core-periphery fashion, with the greatest developmental effects occurring in networked hyperedges within the functional core. Taken together, these results emphasize that the network organization of human brain emerges across multiple scales and evolves substantially through the adolescent period.

q-bio.QM↗

arXiv2016

Structure, Function, and Control of the Musculoskeletal Network

The human body is a complex organism whose gross mechanical properties are enabled by an interconnected musculoskeletal network controlled by the nervous system. The nature of musculoskeletal interconnection facilitates stability, voluntary movement, and robustness to injury. However, a fundamental understanding of this network and its control by neural systems has remained elusive. Here we utilize medical databases and mathematical modeling to reveal the organizational structure, predicted function, and neural control of the musculoskeletal system. We construct a whole-body musculoskeletal network in which single muscles connect to multiple bones via both origin and insertion points. We demonstrate that a muscle's role in this network predicts susceptibility of surrounding components to secondary injury. Finally, we illustrate that sets of muscles cluster into network communities that mimic the organization of motor cortex control modules. This novel formalism for describing interactions between the muscular and skeletal systems serves as a foundation to develop and test therapeutic responses to injury, inspiring future advances in clinical treatments.

q-bio.TO↗

arXiv2014

Learning-Induced Autonomy of Sensorimotor Systems

Distributed networks of brain areas interact with one another in a time-varying fashion to enable complex cognitive and sensorimotor functions. Here we use novel network analysis algorithms to test the recruitment and integration of large-scale functional neural circuitry during learning. Using functional magnetic resonance imaging (fMRI) data acquired from healthy human participants, from initial training through mastery of a simple motor skill, we investigate changes in the architecture of functional connectivity patterns that promote learning. Our results reveal that learning induces an autonomy of sensorimotor systems and that the release of cognitive control hubs in frontal and cingulate cortices predicts individual differences in the rate of learning on other days of practice. Our general statistical approach is applicable across other cognitive domains and provides a key to understanding time-resolved interactions between distributed neural circuits that enable task performance.

q-bio.NC↗