arXiv subjects

Andy Zeyi Liu

Publications and source records attributed to Andy Zeyi Liu.

7 recordsLinked to original sources

arXiv2026

How Good Are Frontier Models at Physics? Expert Re-Grading Reveals Broken Evaluations and Near-Saturation of Leading Benchmarks

Low reported scores on leading physics benchmarks, including those featured in the Artificial Analysis Intelligence Index (2026), suggest that frontier language models still struggle with advanced physics, a demanding test of their scientific reasoning and quantitative problem-solving abilities. Yet this impression does not always align with domain experts' experiences using these models in their work. We revisit these reported findings by evaluating frontier models on six widely used physics benchmarks and auditing them with experts, focusing on text-only problems with verifiable final answers. For each subfield of physics, faculty and graduate researchers with relevant expertise carefully review problem statements, reference solutions, and model responses to distinguish genuine model errors from grader errors, incorrect reference solutions, and ambiguous or underspecified questions. Most audited cases initially evaluated as incorrect reflect these benchmarking issues rather than errors in the models' physics reasoning. We then ask experts to address these benchmarking issues by correcting erroneous reference solutions and repairing or excluding flawed questions. We find that GPT-5.6-Sol's measured mean@4 rises from 47.3% to 78.7% on HLE-Physics and from 61.0% to 87.2% on CMT-Benchmark, while its corrected pass@4 reaches 94.4% on the 54 retained CritPt challenges. Corrected scores are computed on the retained evaluation subsets following expert review. Scores on the audited subsets of UGPhysics, PRISM-Physics, and PHYBench also rise substantially after correction. These findings suggest that current benchmarks substantially understate frontier models' ability to solve well-posed physics problems. Near-saturation on these closed-ended tasks highlights the need for more demanding, expert-validated evaluations.

cs.AI↗

arXiv2026

Semigroup-JEPA: Latent Dynamics Consistency for Zero-Shot Physics Generalization

Joint-Embedding Predictive Architecture (JEPA) world models learn a compact latent representation of the world that supports prediction and planning, but their capability to learn physics and generate physically realistic dynamics remains hitherto untested. In this work, we introduce SemiGroup-JEPA (SG-JEPA), which extends the LeWorldModel framework by supplying the parameter governing the physics to the temporal model via action-conditioning and jointly training an encoder and predictor through an autoregressive latent rollout. To evaluate the model's ability to generalize out of distribution, we design dynamical tasks under different gravitational fields that, despite obeying the same physical law, exhibit qualitatively different dynamics, ranging from floating motion in weak gravitational fields to rapid bouncing in strong ones. In contrast to DINO-WM, SG-JEPA reduces open-loop prediction error by up to 2 times on two-dimensional datasets, and increases control success rate up to 2.5 times for three-dimensional robotic datasets, for which we train independent diffusion policies. To explain this advantage, we develop a linear feature model that separates local law-conditioned error from its recursive amplification under rollout. Guided by this model, we find that back-propagating the multi-step rollout loss into the representation trains the encoder to keep the features that the predictor can carry forward, and that those are the features the dynamics depend on, so most of the gain comes from the encoder learning better features rather than from the predictor learning better dynamics. See project page at https://sg-jepa.github.io.

cs.LG↗

arXiv2026

Memory Inception: Latent-Space KV Cache Manipulation for Steering LLMs

Steering large language models (LLMs) is usually done by either instruction prompting or activation steering. Prompting often gives strong control, but caches guidance tokens at every layer and can clutter long interactions; activation steering is compact but typically weaker and does not support large structured reminders. We introduce memory inception (MI), a training-free method that steers in latent attention space by inserting text-derived key-value (KV) banks only at selected layers. Rather than materializing reminder content throughout the prompt cache, MI treats steering as selective KV allocation, injecting latent slots only where the model routes to them. On matched personality-steering tasks, MI gives the best overall control--drift trade-off, remaining competitive with prompting while consistently outperforming CAA. On updateable guidance, MI supports mid-conversation behavior shifts without rewriting the visible transcript, achieving the highest post-shift alignment on Qwen3. On structured reasoning, MI outperforms visible prompting on HARDMath and PHYSICS (10/12 subject$\times$mode cells), serving as proxies for structured reasoning in verifiable domains, while cutting content-matched KV storage by up to 118$\times$. These results position MI as a powerful steering method when guidance is persistent, structured, or expensive to keep in the visible transcript.

cs.LG↗

arXiv2026

Spectral Lens: Activation and Gradient Spectra as Diagnostics of LLM Optimization

Training loss and throughput can hide distinct internal representation in language-model training. To examine these hidden mechanics, we use spectral measurements as practical and operational diagnostics. Using a controlled family of decoder-only models adapted from the modded NanoGPT codebase, we introduce an empirical protocol based on activation covariance and per-sample gradient SVD spectra. This dual-view reveals three empirical findings and one mechanistic explanation. First, batch size acts as a latent determinant of representation geometry: runs that reach equal loss settle into systematically distinct activation spectra. Second, the activation covariance tail measured early in training reliably forecasts downstream token efficiency. Third, movement of the activation spectrum head (leading modes), together with gradient spectra, characterizes underlying learning-dynamics changes, separating learning-side architectural improvements from primarily execution-side gains. These predictive and diagnostic signals persist across the 12-, 36-, and 48-layer model tiers. Finally, a mechanistic model proves the main observations and explains how activation covariance spectra correlate with task-aligned feature learning.

stat.ML↗

arXiv2026

QGPU: Parallel logic in quantum LDPC codes

Quantum error correction is critical to the design and manufacture of scalable quantum computing systems. Recently, there has been growing interest in quantum low-density parity-check codes as a resource-efficient alternative to surface codes. Their adoption is hindered by the difficulty of compiling fault-tolerant logical operations. A key challenge is that logical qubits do not necessarily map to disjoint sets of physical qubits, which limits parallelism. We introduce clustered-cyclic codes, a quantum low-density parity-check code family with finite-size instances such as [[136,8,14]] and [[198,18,10]] that are competitive with state-of-the-art constructions. These codes admit a directly addressable logical basis, enabling highly parallel logical measurement layers. To leverage this structure, we propose parallel product surgery for quantum product codes. Using an auxiliary copy of the data patch and an engineered product-connection structure, the protocol performs many logical Pauli-product measurements in a single surgery round with small, fixed overhead. For clustered-cyclic codes, this yields surface-code-style maximal parallelism: up to k/2 disjoint Pauli-product measurements per round under explicit algebraic conditions. We prove that parallel product surgery preserves the code distance for hypergraph product codes and numerically verify distance preservation for the listed clustered-cyclic instances with k = 8. Finally, for the [[24,8,3]] clustered-cyclic code, treating half of the logical qubits as auxiliaries enables arbitrary parallel CNOTs on disjoint pairs; combined with symmetry-derived operations, these gates generate the full Clifford group fault-tolerantly.

quant-ph↗

arXiv2026

Check-weight-constrained quantum codes: Bounds and examples

Quantum low-density parity-check (qLDPC) codes can be implemented by measuring only low-weight checks, making them compatible with noisy quantum hardware and central to the quest to build noise-resilient quantum computers. A fundamental open question is how constraints on check weight limit the achievable parameters of qLDPC codes. Here, we study stabilizer and subsystem codes with constrained check weight, combining analytical arguments with numerical optimization to establish strong upper bounds on their parameters. We show that stabilizer codes with checks of weight at most three cannot have nontrivial distance. We also prove tight tradeoffs between rate and distance for broad families of CSS stabilizer and subsystem codes with checks of weight at most four and two, respectively. Notably, our bounds are applicable to general qLDPC codes, as they rely only on check-weight constraints without assuming geometric locality or special graph connectivity. In the finite-size regime, we derive numerical upper bounds using linear programming techniques and identify explicit code constructions that approach these limits, delineating the landscape of practically relevant qLDPC codes with tens or hundreds of physical qubits.

quant-ph↗

arXiv2025

Fault-tolerant Preparation of Distant Logical Bell Pair -- with application in the magic square game

Measures of quantum nonlocality traditionally assume perfect local computation. In real experiments, however, each computational primitive is imperfect. Fault-tolerant techniques enable arbitrarily accurate quantum computation but do not necessarily preserve optimized measures of nonlocality. We examine the impact of low noise on quantum nonlocality in nonlocal games, where even small imperfections can disproportionately increase entanglement consumption. Focusing on the fault-tolerant magic square game, we optimize the tradeoff between noisy entanglement consumption and deficit in the game value. We introduce an interface circuit and logical entanglement purification protocol (EPP) to efficiently translate states between physical and logical qubits and purify noisy logical Bell pair, reducing Bell pair consumption. Our analytical and numerical results, particularly for the $[[7^k,1,3^k]]$ concatenated Steane code, demonstrate exponential Bell pair savings and a higher noise threshold. We establish theoretical lower bounds for local noise threshold of $4.70\times10^{-4}$ and an initial Bell pair infidelity threshold of $18.3\%$. Our framework is adaptable to various quantum error-correcting codes (QECCs) and experimental platforms. This work not only advances fault-tolerant nonlocal games but also inspires further research on interfacing different QECCs, fostering modular quantum architectures and the quantum internet.

quant-ph↗