Persistent Teacher Anchoring for Tool-Using Agents
Hyun Bin Park (Sogang University)Kyungho Song (University of MichiganAnn Arbor)Sangmin Lee (Sogang University)Du-Seong Chang (Sogang University)
Sep 2026
Artificial IntelligenceMachine LearningNatural Language Processing
Abstract
Distillation is common in LLM post-training, where on-policy knowledge distillation (OPKD) uses student-generated trajectories to prepare the student for downstream RL. At each state, the student matches a next-token distribution supplied by the teacher. As the rollout enters states the teacher would not visit, the teacher-student distribution gap can accumulate. In tool use, this gap becomes consequential because student-written calls execute before supervision and their observations shape later prefixes. Proposer-verifier generation addresses this drift by letting the teacher decide which student-proposed text is retained during generation. Existing formulations govern text but leave tool execution outside their scope. We propose Persistent Teacher Anchoring (PTA), a student-induced but teacher-committed rollout construction. PTA retains chunk-level verification and adds turn-level commitment, allowing a call to reach the environment only after the teacher has verified the entire turn. Treating verified chunks as atomic generation units, we introduce persistent lookahead, which fills idle rollout capacity by advancing future samples and carrying unfinished ones across student updates under the fixed verifier. Across Search-R1-style retrieval and DeepEyes-style perception RL, applying PTA before downstream RL improves macro best@4 by 2.5 and 2.8 points over OPKD under the same downstream RL budget, while lookahead improves throughput by 24%.
This publication proposes a definition and a classification of agile software development approaches and analyses ten software development methods that can be characterized as being "agile" against the defined criterion.
P. Abrahamsson, O. Salo, Jussi Ronkainen et al.· arXiv.org· 727 citations· ⚡54
The results indicate that software engineering work practices are chosen opportunistically, adapted and configured to provide value under the constrains imposed by the startup context.
Nicolò Paternoster, Carmine Giardino, M. Unterkalmsteiner et al.· Information and Software Tec...· 394 citations· ⚡54
The possibility of inferring high-dimensional data inference in a model that consists of a prior and an auxiliary differentiable constraint given some additional information is considered, thereby allowing a range of potential applications in adapting models to new domains and tasks.
Alexandros Graikos, Esmeralda S. Whitammer, N. Jojic et al.· Neural Information Processin...· 316 citations· ⚡15
It is proved that any global minimizer of the trajectory balance objective can define a policy that samples exactly from the target distribution, and empirically demonstrate the benefits of the trajectories balance objective for GFlowNet convergence, diversity of generated samples, and robustness to long action sequences and large action spaces.
Esmeralda S. Whitammer, Moksh Jain, Emmanuel Bengio et al.· Neural Information Processin...· 302 citations· ⚡60
The growing literature on affect among software developers mostly reports on the linkage between happiness, software quality, and developer productivity. Understanding happiness and unhappiness in all its components -- positive and negative emotions and moods -- is an attractive and important endeavor. Scholars in industrial and organizational psychology have suggested that understanding happiness and unhappiness could lead to cost-effective ways of enhancing working conditions, job performance, and to limiting the occurrence of psychological disorders. Our comprehension of the consequences of (un)happiness among developers is still too shallow, being mainly expressed in terms of development productivity and software quality. In this paper, we study what happens when developers are happy and unhappy while developing software. Qualitative data analysis of responses given by 317 questionnaire participants identified 42 consequences of unhappiness and 32 of happiness. We found consequences of happiness and unhappiness that are beneficial and detrimental for developers' mental well-being, the software development process, and the produced artifacts. Our classification scheme, available as open data enables new happiness research opportunities of cause-effect type, and it can act as a guideline for practitioners for identifying damaging effects of unhappiness and for fostering happiness on the job.
D. Graziotin, Fabian Fagerholm, Xiaofeng Wang et al.· Journal of Systems and Softw...· 236 citations· ⚡13
Mobile phones have been closed environments until recent years. The change brought by open platform technologies such as the Symbian operating system and Java technologies has opened up a significant business opportunity for anyone to develop application software such as games for mobile terminals. However, developing mobile applications is currently a challenging task due to the specific demands and technical constraints of mobile development. Furthermore, at the moment very little is known about the suitability of the different development processes for mobile application development. Due to these issues, we have developed an agile development approach called Mobile-D. The Mobile-D approach is briefly outlined here and the experiences gained from four case studies are discussed.
P. Abrahamsson, Antti Hanhineva, H. Hulkko et al.· Conference on Object-Oriente...· 225 citations· ⚡18
A new machine-learning framework aims to improve the success rate of computational protein design while moving away from results that reproduce sequences found in nature.
