From Observation to Intervention: A Causal Audit of Expert Importance in Mixture-of-Experts Models

Paper 2 likely has higher impact due to a more broadly applicable, timely contribution to scaling LLM pretraining under real distributed-systems constraints. GASLoC proposes a novel decentralized algorithm compatible with adaptive optimizers, local steps, and sparse randomized communication, and shows empirical gains over strong baselines (including heterogeneous bandwidth settings). This targets a major bottleneck for frontier training and can influence both ML systems and optimization research. Paper 1 is rigorous and valuable as a cautionary interpretability result, but its impact is narrower (MoE pruning/interpretability methodology) and primarily negative/diagnostic rather than enabling new capabilities.

Paper 2 introduces a novel and practical methodology for protein property prediction that addresses a significant challenge in protein design. Its contributions—evolutionary substitution matrix-based kernels, structure-aware learning, and multi-task capabilities—have broad applicability in biotechnology, drug design, and protein engineering. Paper 1, while methodologically rigorous and offering important cautionary insights about interpretability practices in MoE models, is primarily a negative result with narrower scope, showing that observational metrics don't predict causal importance. Paper 2's positive, constructive contribution with demonstrated practical utility gives it higher potential for real-world impact and broader adoption.

Paper 2 is more novel and broadly impactful: it challenges a common inferential leap in interpretability (associational metrics → interventional claims) with a concrete, multi-model causal audit, yielding a general negative result and an evidential standard that can reshape evaluation practices beyond MoE pruning. Its methodological rigor is high (token-level interventions, multiple-comparison correction, power control), and the implications span interpretability, causal evaluation, pruning, and safety auditing. Paper 1 is strong and timely for PTQ efficiency, but its impact is likely more incremental and narrower to quantization workflows.

Paper 2 (K-Forcing) addresses a critical bottleneck in LLM deployment—inference efficiency—with a novel paradigm (push-forward language modeling) that offers concrete 2.4-3.5x speedups while maintaining compatibility with existing infrastructure. This has immediate, broad real-world applications in industrial-scale LLM serving. Paper 1, while methodologically rigorous and raising important points about the observation-intervention gap in MoE interpretability, is more narrowly focused on a negative/cautionary result about existing pruning metrics. Paper 2's potential to influence both research (new decoding paradigms) and practice (deployment costs) gives it broader impact.

Paper 2 has higher potential impact: it challenges a widely used inferential leap in interpretability/pruning (observational routing stats → causal expert importance) with a systematic interventional audit across multiple popular MoE families. The negative result is methodologically rigorous (token-level interventions, multiple-comparison correction, power control) and broadly relevant to interpretability, causal evaluation, and model compression practices. Its implications generalize beyond MoE pruning to many observational interpretability claims, making it timely and cross-cutting. Paper 1 is useful engineering for speech+LLM efficiency, but is narrower and more incremental.

Paper 2 addresses a fundamental methodological gap in ML interpretability—the conflation of observational and interventional evidence (Pearl's causal hierarchy)—with rigorous experimental methodology including multiple-comparison correction and effect size reporting. Its findings challenge widely-used assumptions in MoE pruning and have broad implications for interpretability research standards across the field. Paper 1, while technically sound, offers an incremental improvement to PPO-style trust regions for LLM RL, a rapidly evolving area where methods are frequently superseded. Paper 2's contribution is more foundational and likely to influence research practices across multiple subfields.

Paper 1 addresses a timely and broadly impactful issue in AI interpretability and MoE model pruning—a rapidly growing area given the deployment of large MoE models. It introduces a rigorous causal framework (connecting Pearl's causal hierarchy to interpretability practices) and provides concrete empirical evidence challenging widely-used assumptions, which could reshape pruning methodologies and interpretability standards across the field. Paper 2, while technically strong, extends known impossibility results from ReLU to tanh networks in a narrower theoretical niche, with less immediate practical impact and a smaller affected community.

Flow-DPPO addresses a practical and timely problem in RL-based fine-tuning of flow matching models for image/video generation, proposing a principled replacement for ratio clipping with exact KL divergence constraints. It has immediate real-world applications in generative AI, demonstrates concrete improvements across multiple metrics, and provides open-source code. Paper 2 makes a valuable methodological point about the gap between observational and interventional evidence in MoE pruning, but its scope is narrower (a negative/cautionary result on existing metrics) with less direct applicability to improving systems, limiting its broader impact.

Paper 1 (QGF) addresses a fundamental challenge in scaling RL with expressive policies by proposing test-time policy improvement that avoids actor-critic training instability. It has broad practical applications in robotics and control, demonstrates strong empirical results across multiple benchmarks, and offers favorable scaling properties. Paper 2 makes a valuable methodological point about the gap between observational and interventional evidence in MoE interpretability, but its scope is narrower—primarily a cautionary finding about existing pruning heuristics rather than enabling new capabilities. Paper 1's contribution is more actionable and broadly impactful across robotics and RL communities.

Paper 2 likely has higher impact: it introduces a generally applicable, plug-and-play gradient estimation framework for optimizing order-statistic (distributional/risk-sensitive) objectives in RL, covering many important criteria (VaR/CVaR, trimmed means, top-m) with unbiased estimators and broad real-world relevance (robustness, tail risk, best-of-K), plus code and demonstrations (incl. LLM post-training). Paper 1 is methodologically rigorous and timely as a cautionary interventional audit, but its impact is narrower (MoE pruning/interpretability) and primarily corrective rather than enabling a new widely reusable method.