Tuesday, July 28, 2026

1. Reason-Mediated Behavioral Models for Auditing LLM Social Simulators
   T2arXiv AI/CL/LG（论文）

   Large language models (LLMs) are increasingly deployed as social simulators, e.g., synthetic survey respondents. Current evaluations mainly check whether simulated outcomes match human final answers, but ignore the rationale behind them. Through a 94-person sunscreen concept test, this paper maps human open-ended rationales into signed reason states, showing that human-derived reasons substantially improve hold-out purchase intent prediction. LLM-simulated reasons, though sounding plausible, often echo the concept board instead of recovering the respondent's acceptance or rejection path. The work contributes an interpretable audit framework to assess whether a simulator's stated reasons align with human evidence.

2. DataOrchestra: Learning to Orchestrate Per-Example Curation of Pretraining Data
   T2arXiv AI/CL/LG（论文）

   Pretraining data processing is critical for downstream performance of large language models, but existing methods often apply fixed strategies at corpus or domain level without adapting to per-example needs. This paper proposes DataOrchestra, a framework that unifies multiple processing operations and dynamically orchestrates an example-specific pipeline: an orchestrator decides to drop, keep, or clean each chunk; for cleaning, it selects programmatic editing or LLM-based rewriting with concrete instructions. Experiments from 0.5B to 7B models show stable average gains over individual methods across 11 benchmarks, outperforming stronger baselines in math continued pretraining while reducing compute by skipping unnecessary operations.

1. Stacking the Deck: Tunable Trainability in Stacked LCUs
   T2arXiv AI/CL/LG（论文）

   Variational quantum circuits (VQCs) face a fundamental trade-off: ansätze expressive enough to resist classical simulation tend to suffer from barren plateaus, while those that provably avoid barren plateaus are often classically simulable. This paper proposes Stacked Linear Combination of Unitaries (S-LCU), a variational ansatz that uses the number of layers l as a single dial to tune the balance between barren plateaus and classical simulability. The authors prove a lower bound of Ω(1/(n k^{3l})) on the loss-landscape variance for the Free Fermion S-LCU, with a classical simulation cost of O(k^{2l} n^3) and quantum gate complexity of only O(lkn^2). This work offers a systematic method for constructing ansätze with a controlled complexity-trainability trade-off.

—

—

—