cs.AINov 14, 2026

Staged Semantic Distillation: Preserving Free-Form Semantics in a Low-Compute Non-LLM Surrogate of a Staged LLM Agent

Authors: Zhiyang Niu, Huayuan Shen, Chuanjie Li, Mengmeng Zhao, Jiajia Zhang

Organizations: Harbin Institute of Technology, Shenzhen, China. · Nanyang Technological University, Singapore. · Pingan Group, China.

Abstract

Staged LLM agents make market and social simulators realistic, but one 102-day run of a 200-agent, ten-index market simulator issues 203,100 language-model calls and returns one realization. Conventional low-compute distillation drops their free-form intermediate state. We present staged semantic distillation, a low-compute method that preserves non-structured semantics in a non-LLM student trained from one recorded trace. A frozen encoder maps each teacher output to a 1024-d vector; the student sees typed stage-tagged vectors with stage-specific visibility, is jointly supervised, and trains with scheduled self-history. One TwinMarket trace yields a 13.2M-parameter deterministic student that trains in 1,658 seconds on one 8-GB GPU and runs with zero teacher or decoder calls and no text generation (0.14 ms per decision). Its analysis stays close to the teacher (cosine 0.943), and swapping semantic inputs between agents moves 64.7% of decisions against 12.8% for a matched donor, so the channel is preserved and live. We use the surrogate to study what the transfer carries. It reaches 0.8347 balanced accuracy on the teacher-forced replay and 0.7783 [0.766,0.791] on the deployment decoder; the five-seed gain over a no-semantic retrain is +2.2 [1.21,3.14]. Against the teacher's own re-sampling level, round-one R is 97.1% [95.8,98.4] and 92.3% [89.0,95.4]; R is a diagnostic, not an accuracy. A five-seed content-by-schedule study isolates the gain: scheduled self-history carries it (+2.2 [1.21,3.14]); shuffled and constant content controls reproduce it, while content matters teacher-forced. The design rule: invest in the deployment schedule, let content earn its place at the teacher-forced operating point. A 20-seed, 200,000-decision sweep takes 12 minutes with zero LLM calls, versus three hours for the teacher. The paper is a method-and-case-study contribution on one simulator and trace.

Figures & tables

Appendix figures & tables14 assets

Supplementary material from the paper’s appendix.

Appendix

Explore similar work

Nov 10, 2026

SteerDistill: Quantizing LLM Agents via Closed-Loop Teacher Steering

Low-bit quantization reduces the inference cost of LLM agents but can compromise task success. Examining failed rollouts, we find that quantized agents overlook task requirements or environmental conditions, producing environment transitions that diverge from successful full-precision references. Correcting these behavioral divergences remains difficult: offline data may miss the student's failure contexts, while its own rollouts may provide few correct responses in those contexts. We introduce SteerDistill, a quantization-aware distillation method in which a full-precision teacher guides the student to revise failed responses and iteratively refines its guidance using feedback from student execution. This closed-loop teacher steering produces locally verified behavioral corrections. We internalize these corrections by training the student without guidance to match the steered student's predictions at corrected turns and the full-precision teacher's at other turns, enabling unassisted inference. Across five agentic benchmarks and three models, SteerDistill exceeds the strongest quantized baseline by 1.67--2.18 points in average score under W4A16 and stays within one point of BF16. On τ2τ^2-Bench, it outperforms QeRL by 1.84 percentage points in success rate using 67% fewer training GPU hours and produces shorter successful trajectories than the quantized baselines.
Oct 22, 2026

Beyond On-Policy vs. Off-Policy: At What Level Should LLM Agent Distillation Follow the Student?

Distilling interactive LLM agents requires choosing between teacher-generated demonstrations and student-generated experience, corresponding broadly to off-policy and on-policy distillation. While both have proven effective, we argue that this distinction is inherently hierarchical: the policy that induces an interaction history need not be the same as the policy that generates the token prefix of the current action. We introduce a hierarchical agent distillation framework that formalizes these two levels of on-policiness and enables explicit control over their pairing, mixture proportions, and training schedule. We study this problem through self-distillation, where a frozen teacher has access to task guidance unavailable to the student, and conduct a three-stage controlled study across three models and four interactive environments. We find that student-induced histories paired with teacher-generated action prefixes generally provide the strongest standalone training signal, whereas following the student at both levels generally performs worst. Nevertheless, fully on-policy data becomes beneficial when incorporated into a mixture with other configurations. Moreover, mixtures with identical teacher/student proportions at each level can yield substantially different performance, showing that how the two levels are paired matters beyond the overall amount of on-policy data. Finally, front-loading fully teacher-generated contexts and delaying fully on-policy contexts improves over a static mixture. These findings provide practical principles for deciding where, how much, and when agent distillation should follow the student.
Sep 20, 2026

PSD: Pseudo Self-Distillation of Memory Representation Capabilities for LLM Agents

Memory systems are becoming a core component of LLM agents, but constructing and maintaining memory remains expensive because it relies on repeated calls to large proprietary language models. This cost creates a major barrier to deploying memory-enhanced agents at scale. In this paper, we present Pseudo Self-Distillation (PSD), a framework that enables small language models (SLMs) to construct hierarchical memory representations by distilling behavior from a strong black-box oracle through a multi-stage training pipeline. Standard distillation methods require access to teacher logits or hidden states, which closed models do not expose. Unlike conventional self-distillation settings, where supervision is derived from a model's own predictions, sampled rollouts, or aggregated outputs, PSD enables a single-model distillation setup while channeling external oracle knowledge through the prompt. PSD uses a single small model in two roles: a teacher that sees a privileged prompt containing the oracle's answer as reference context, and a student that sees only the task prompt. The student learns to reproduce the teacher's output distribution, absorbing oracle-guided behavior into its own weights without accessing the oracle's internals. On LoCoMo, PSD-trained Qwen3-0.6B, 1.7B, and 4B match or exceed GPT-4.1-mini on downstream retrieval at a fraction of the deployment cost, with off-policy PSD achieving the strongest results across most conditions. We further show that this memory-construction capability transfers out-of-distribution to LongMemEval, despite the students being trained exclusively on LoCoMo with no exposure to LongMemEval data.