arXiv AI

Coarse-Graining Recursive Expert Computation: Continuation-Relative Interfaces and Failure Modes

arXiv AI
1d ago

From Local Prediction Error to Final Outputs: A Theory of Expert Sharing in Multi-Step Reasoning

The paper investigates how sharing experts in mixture‑of‑experts (MoE) models across layers or computational steps can reduce local prediction error but may not always improve final outputs. By analyzing linear regression models, the authors derive criteria that balance variance reduction against bias introduced by sharing across roles with different target behaviors, and they identify conditions under which local gains persist in the final output. The study also provides error bounds for general routed computations, demonstrates that correct predictions alone do not confirm expert attribution, and presents controlled experiments showing cases where local improvements reverse into higher final‑output error.

By Tiantong Wang, Zhiren Gong, Yixin Wang, Tiantong Wu, Wei Yang Bryan Lim
arXiv Machine Learning
Oct 5

The Conflict Between Logic and Memory: Training Conditions for Optimizer-Dependent Rule Acquisition

The paper investigates how different optimizers acquire distinct generalizing relations in single‑hidden‑layer ReLU networks. By manipulating nuisance‑connected weights and performing a strict‑parity sweep, the authors show that training conditions reshape both shared failures and relative optimizer advantages, with specific interventions improving accuracy on hard stages. The study also compares optimizer behavior across tasks and an external architecture (CIFAR‑10).

By Gongyue Zhang, Honghai Liu
arXiv Machine Learning
Oct 9

Conditional Transfer from Controlled Pretraining Mixtures to Code

The paper investigates how synthetic tasks influence language‑model pretraining by distinguishing diagnostic, teachable, and transferable signals. In a controlled setting where 70% of the data is general Python and 30% is a mixture of curated synthetic tasks, literature‑derived probes, and OpenCodeInstruct, 14 of 27 tasks show teachability, with curated tasks outperforming literature probes. Curated synthetic data provides conditional value: it boosts downstream performance when mixed with OpenCodeInstruct but is less effective when used alone, and a loss‑based scheduler can misalign short‑term loss reduction with long‑term transfer.

By Ohad Rubin
arXiv Machine Learning
Nov 5

How Far Can You Steer? Behavior-Capability Frontiers in Contrastive Activation Addition

The paper investigates Contrastive Activation Addition (CAA), a method that steers language‑model behavior by adding activation vectors derived from contrasting examples. By framing CAA as a behavior‑capability frontier, the authors select interventions that maximize behavioral gain while limiting capability degradation to specified budgets (1%, 3%, 5%). Across four instruction‑tuned models and seven behaviors, they find that validated behavioral effects generalize well, and that joint behavior‑capability survival improves significantly with higher budgets, though additional gains plateau. Control experiments confirm that the effects rely on learned behavior directions rather than generic perturbations, and representational geometry only modestly predicts cross‑behavior transfer. Secondary tests reveal that even interventions meeting protected benchmarks can still degrade untested capabilities or transfer unevenly to free‑form generation, underscoring that activation steering is a constrained, model‑dependent intervention rather than a universal control knob.

By Deepanshu Raj
arXiv AI
Oct 26

LeafSkill: Efficient Leaf-Level Action Grounding for Recursive LLM Agents

LeafSkill introduces an external reuse layer for recursive LLM agents that retrieves semantically related action grounding experiences, binds action templates to current objects, and uses a lightweight 1.7B selector to choose or defer actions. This approach preserves high coverage of candidate actions under paraphrases (96–97%) and achieves a 97.34% adoption rate of unseen skill types. In experiments on Robotouille tasks, LeafSkill outperforms ReCAP, completing more tasks and reducing token usage and calls by over 14%.

By Shen Liu, Ze Zhou, Zhenguo Xu, Shaopu Wang, Yike Gao, Chunlei Wang
arXiv AI
Jun 15

Learning What to Predict: Downstream-Guided Task Design for Continued Pretraining

arXiv:2601. 22108v2 Announce Type: replace-cross Abstract: Continued pretraining is optimized with fixed self-supervised tasks but selected by downstream performance, creating a coarse feedback loop in which practitioners evaluate checkpoints, change data mixtures or objectives, and restart runs, while individual updates remain blind to target capabilities.

By Shuqi Ke, Giulia Fanti