Searcharxiv⌕ Search
Home/Archive/arXiv

arXiv · 2610.10460

Composing What Each Teacher Learned: Multi-Teacher On-Policy Distillation through Teacher-Relative Shifts

Hejian Sang·Zhengze Zhou·Shayan Mohajer Hamidi·Xiaomin Li·Rohit Jain·Alborz Geramifard
View original source ↗Read source document ↗Export metadata ↓

Abstract

Multi-teacher on-policy distillation (MOPD) is used in two settings. In common-domain composition, several teachers score each student rollout from one prompt domain and their signals form a single target; in routed-domain distillation, prompts from different domains are assigned to the corresponding specialist. Both settings usually transfer each teacher's endpoint policy, which mixes what post-training changed with preferences inherited from the teacher's base. We introduce $Δ$-MOPD, which transfers each teacher's teacher-minus-base logit shift re-anchored at the student's frozen initialization, and compare it with endpoint supervision in both settings while holding teacher selection fixed. We first expose the mechanism that impedes endpoint transfer: inherited base pull can exceed the post-training shift. Removing it reduces the teacher-term norm ratio and target--student KL. Across our experiments, the results suggest that shift targets are particularly useful when teacher signals are combined at a state. With three composed teachers, $Δ$-MOPD exceeds endpoint composition by $4.11$ Math and $1.95$ five-benchmark points; with two, it matches endpoint accuracy. Under phased routing, it achieves higher mean performance in both phase orders and reduces the observed order gap from $10.50$ to $6.42$ points. Under interleaved routing, where each update involves one teacher, the two targets perform comparably. The phased results provide supporting evidence that the benefit may extend to signals accumulated across training phases. Target construction is thus an independent design axis in MOPD, complementary to teacher selection.

Explore related subjects

cs.LGcs.AI

Keep this discovery

Explore connections, maps & timelines

BibTeXRIS

Hejian Sang, Zhengze Zhou, Shayan Mohajer Hamidi, Xiaomin Li, Rohit Jain, Alborz Geramifard. 2026-10-07. Composing What Each Teacher Learned: Multi-Teacher On-Policy Distillation through Teacher-Relative Shifts. https://arxiv.org/abs/2610.10460

Cite the original work for its findings. Save a collection to share your selection of sources.

KEEP EXPLORING

Related papers

arXiv2026

Learning in the Recurrent State: Gradient Descent with Linear Recurrent Networks

In-context learning lets a sequence model adapt to a new task from examples in its input. A prominent line of work shows how self-attention can be constructed to implement gradient descent on a linear predictor fit to the in-context examples during the forward pass. State-space models (SSMs) and other linear recurrent networks (LRNNs) model sequences at linear time cost, but it is unclear how their recurrent update could carry out the same in-context gradient descent. We introduce Gradient-based Recurrent In-context Learner (GRIL), a diagonal LRNN that factorizes a supervised gradient step into a short-window cross-product write and a multiplicative readout of the next query. For linear regression, this construction accumulates the context gradient in a matrix state and applies it in a single forward pass, with $O(f^2)$ learned degrees of freedom. The same design extends to multi-step updates and cross-entropy classification, with a limited MLP-based extension to non-linear regression. We show empirically that trained GRILs recover the behavior and parameters analytically predicted by the construction on synthetic ICL tasks. Furthermore, the same architecture can be extended and trained on general-purpose benchmarks, including Long Range Arena, language modeling and associative recall. Together, these results establish windowed cross-product self-attention as a concrete inductive bias that lets LRNNs learn in context through gradient-descent-like updates, while remaining trainable on general-purpose tasks.

cs.LG↗
arXiv2026

LLaTA: Unlocking Graph Structure Learning with Tree-Guided Large Language Models

The emergence of large language models (LLMs) has popularized text-attributed graphs (TAGs), creating an urgent need for graph structure learning (GSL) methods that effectively leverage textual information. However, existing GSL approaches are designed for traditional graphs without text, and adapting them to LLMs faces two challenges: defining a suitable optimization objective given LLMs' massive parameters, and designing an efficient architecture without costly fine-tuning. To address these, we propose LLaTA (Large Language and Tree Assistant), which reformulates GSL as a tree optimization framework---shifting from training edge predictors to designing a language-aware tree sampler. LLaTA constructs structural encoding trees via entropy minimization to capture topology, then leverages tree-guided LLM in-context learning to integrate textual semantics without fine-tuning. Extensive experiments on 11 datasets demonstrate LLaTA's flexibility with any backbone, superior scalability over LLM-based GSL methods, and state-of-the-art effectiveness across diverse domains.

cs.LG↗
arXiv2026

Causal Posterior Estimation

We present Causal Posterior Estimation (CPE), a novel method for Bayesian inference in simulator models, where evaluating the likelihood function is intractable or computationally expensive, but generating outputs given parameter values is straightforward. CPE approximates the posterior distribution using flow matching while directly incorporating the conditional dependence structure induced by the model's graphical representation into the neural network architecture. Across extensive experiments, we demonstrate that hard-coding these conditional dependencies into the network, rather than requiring them to be learned from data, enables CPE to achieve highly accurate posterior inference that matches or outperforms state-of-the-art baselines.

cs.LG↗