SubspacePath Pruner introduces an inference-time pruning mechanism that couples representation probes with parameter pathways, enabling scenario-adaptive routing that retains robustness while reducing unnecessary computation.

This work combines active inference with explicit rule constraints to model habit dynamics in sequential decision settings, improving action quality and latency in realistic environments.

SoT studies endogenous state-conditioned reasoning to improve quality-efficiency trade-offs across backbones and benchmark families.

LoTR enhances existing external reasoning paradigms by routing internal computation pathways according to evolving logic states at each reasoning step.

CoAx (conditional co-ablation) recovers dormant self-repair backup components that first-order ablation scores miss, improving backup-head discovery, repair-aware attribution, capability knockout, and structured pruning across transformer circuits.

Almost every structured-pruning method scores each unit in isolation, assuming set damage is additive — the wrong granularity for a Transformer's shared residual stream. CoCurve prunes attention and FFN units jointly: a single Fisher matrix whose diagonal is node saliency and whose off-diagonal entries are cross-module co-pruning curvature edges, provably a Gram product recoverable from just M single-unit ablations. It leads every capability class on Llama-3.1-8B at 20% and best preserves fragile code — prune the edges, not just the nodes.

AnchorPath shows the high-sparsity collapse of training-free structured pruning is a property of the one-shot paradigm — a single quadratic extrapolated across the whole budget — not the importance score. Recasting compression as a calibration-only anchored continuation (a fixed dense anchor, a moving expansion point, K re-anchored steps) cuts perplexity by up to two orders of magnitude versus one-shot at 50% sparsity, while changing only the mask. One-shot second-order pruning is exactly the K=1 case.

CrossAgentMem improves multi-agent reasoning quality while controlling context overhead, achieving stronger accuracy under a bounded token budget through memory-aware collaboration.

This paper provides a theory for why sparse MoE architectures can reduce sample complexity barriers compared with dense models via geometry-aware routing and exponent-collapse behavior.