@inproceedings{ta-etal-2026-proof,
title = "{PROOF}-Gen: From Optimized Data to Better Distillation",
author = "Ta, Anh and
Zhu, Junjie and
Shayandeh, Shahin",
editor = "Atanasova, Pepa and
Haffari, Gholamreza and
Wang, Minghan",
booktitle = "Proceedings of the 2026 Conference on Empirical Methods in Natural Language Processing: Industry Track",
month = oct,
year = "2026",
address = "Budapest, Hungary",
publisher = "Association for Computational Linguistics",
url = "https://aclanthology.org/2026.emnlp-industry.168/",
pages = "2844--2864",
ISBN = "979-8-89176-460-6",
abstract = "Supervised fine-tuning on teacher-generated trajectories is the standard first stage for distilling tool-calling capabilities into deployable models. Post-training pipelines that drive shipped tool-calling agents re-run this stage on a daily or weekly cadence, paying the frontier teacher cost each cycle, yet the mechanism is generate-and-filter (keep the teacher{'}s passing trajectories, discard the rest) and each cycle leaves behind the same hard scenarios because failures supply no signal. On {\ensuremath{\tau}}2-bench, 57{\%} of teacher trials fail, two-thirds of them near-misses (most tool calls correct, undone by one decisive error).\par We introduce PROOF-Gen (Per-scenario Reflective Optimization to Overcome Failed Generation), which recovers golden trajectories from these failures via per-scenario prompt optimization. For each failed task, a reflector analyzes the execution trace and evaluation feedback, then writes corrective guidance that steers the teacher to a passing trajectory. The guidance is stripped before training, so the student learns from clean demonstrations with no task-specific scaffold.\par On {\ensuremath{\tau}}2-bench, per-scenario optimization recovers 93{\%} of failed scenarios. Fine-tuned on the combined data, Qwen3-4B-Instruct-2507 improves from Pass{\textasciicircum}1=0.132 to 0.529 and Gemma 4 E4B-it gains +7.2pp on BFCLv4 multi-turn. In a deployed pipeline, the method lifts trajectory quality by +6.3pp goal completion and transfers to a deployed on-device model (+1.5pp goal completion; +1.7 to +5.0pp across response-quality metrics), with positive transfer in every locale (non-English average +1.48pp)."
}<?xml version="1.0" encoding="UTF-8"?>
<modsCollection xmlns="http://www.loc.gov/mods/v3">
<mods ID="ta-etal-2026-proof">
<titleInfo>
<title>PROOF-Gen: From Optimized Data to Better Distillation</title>
</titleInfo>
<name type="personal">
<namePart type="given">Anh</namePart>
<namePart type="family">Ta</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Junjie</namePart>
<namePart type="family">Zhu</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Shahin</namePart>
<namePart type="family">Shayandeh</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<originInfo>
<dateIssued>2026-10</dateIssued>
</originInfo>
<typeOfResource>text</typeOfResource>
<relatedItem type="host">
<titleInfo>
<title>Proceedings of the 2026 Conference on Empirical Methods in Natural Language Processing: Industry Track</title>
</titleInfo>
<name type="personal">
<namePart type="given">Pepa</namePart>
<namePart type="family">Atanasova</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Gholamreza</namePart>
<namePart type="family">Haffari</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Minghan</namePart>
<namePart type="family">Wang</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<originInfo>
<publisher>Association for Computational Linguistics</publisher>
<place>
<placeTerm type="text">Budapest, Hungary</placeTerm>
</place>
</originInfo>
<genre authority="marcgt">conference publication</genre>
<identifier type="isbn">979-8-89176-460-6</identifier>
</relatedItem>
<abstract>Supervised fine-tuning on teacher-generated trajectories is the standard first stage for distilling tool-calling capabilities into deployable models. Post-training pipelines that drive shipped tool-calling agents re-run this stage on a daily or weekly cadence, paying the frontier teacher cost each cycle, yet the mechanism is generate-and-filter (keep the teacher’s passing trajectories, discard the rest) and each cycle leaves behind the same hard scenarios because failures supply no signal. On \ensuremathτ2-bench, 57% of teacher trials fail, two-thirds of them near-misses (most tool calls correct, undone by one decisive error).\par We introduce PROOF-Gen (Per-scenario Reflective Optimization to Overcome Failed Generation), which recovers golden trajectories from these failures via per-scenario prompt optimization. For each failed task, a reflector analyzes the execution trace and evaluation feedback, then writes corrective guidance that steers the teacher to a passing trajectory. The guidance is stripped before training, so the student learns from clean demonstrations with no task-specific scaffold.\par On \ensuremathτ2-bench, per-scenario optimization recovers 93% of failed scenarios. Fine-tuned on the combined data, Qwen3-4B-Instruct-2507 improves from Pass⌃1=0.132 to 0.529 and Gemma 4 E4B-it gains +7.2pp on BFCLv4 multi-turn. In a deployed pipeline, the method lifts trajectory quality by +6.3pp goal completion and transfers to a deployed on-device model (+1.5pp goal completion; +1.7 to +5.0pp across response-quality metrics), with positive transfer in every locale (non-English average +1.48pp).</abstract>
<identifier type="citekey">ta-etal-2026-proof</identifier>
<location>
<url>https://aclanthology.org/2026.emnlp-industry.168/</url>
</location>
<part>
<date>2026-10</date>
<extent unit="page">
<start>2844</start>
<end>2864</end>
</extent>
</part>
</mods>
</modsCollection>
%0 Conference Proceedings
%T PROOF-Gen: From Optimized Data to Better Distillation
%A Ta, Anh
%A Zhu, Junjie
%A Shayandeh, Shahin
%Y Atanasova, Pepa
%Y Haffari, Gholamreza
%Y Wang, Minghan
%S Proceedings of the 2026 Conference on Empirical Methods in Natural Language Processing: Industry Track
%D 2026
%8 October
%I Association for Computational Linguistics
%C Budapest, Hungary
%@ 979-8-89176-460-6
%F ta-etal-2026-proof
%X Supervised fine-tuning on teacher-generated trajectories is the standard first stage for distilling tool-calling capabilities into deployable models. Post-training pipelines that drive shipped tool-calling agents re-run this stage on a daily or weekly cadence, paying the frontier teacher cost each cycle, yet the mechanism is generate-and-filter (keep the teacher’s passing trajectories, discard the rest) and each cycle leaves behind the same hard scenarios because failures supply no signal. On \ensuremathτ2-bench, 57% of teacher trials fail, two-thirds of them near-misses (most tool calls correct, undone by one decisive error).\par We introduce PROOF-Gen (Per-scenario Reflective Optimization to Overcome Failed Generation), which recovers golden trajectories from these failures via per-scenario prompt optimization. For each failed task, a reflector analyzes the execution trace and evaluation feedback, then writes corrective guidance that steers the teacher to a passing trajectory. The guidance is stripped before training, so the student learns from clean demonstrations with no task-specific scaffold.\par On \ensuremathτ2-bench, per-scenario optimization recovers 93% of failed scenarios. Fine-tuned on the combined data, Qwen3-4B-Instruct-2507 improves from Pass⌃1=0.132 to 0.529 and Gemma 4 E4B-it gains +7.2pp on BFCLv4 multi-turn. In a deployed pipeline, the method lifts trajectory quality by +6.3pp goal completion and transfers to a deployed on-device model (+1.5pp goal completion; +1.7 to +5.0pp across response-quality metrics), with positive transfer in every locale (non-English average +1.48pp).
%U https://aclanthology.org/2026.emnlp-industry.168/
%P 2844-2864
Markdown (Informal)
[PROOF-Gen: From Optimized Data to Better Distillation](https://aclanthology.org/2026.emnlp-industry.168/) (Ta et al., EMNLP 2026)
ACL
- Anh Ta, Junjie Zhu, and Shahin Shayandeh. 2026. PROOF-Gen: From Optimized Data to Better Distillation. In Proceedings of the 2026 Conference on Empirical Methods in Natural Language Processing: Industry Track, pages 2844–2864, Budapest, Hungary. Association for Computational Linguistics.