@inproceedings{sun-kazakov-2026-peak,
title = "Peak-Then-Collapse and the Four Interface Channels of Knowledge-Graph Tool Use",
author = "Sun, Tianda and
Kazakov, Dimitar",
editor = "Lee, Hung-Yi and
Dev, Sunipa and
Li, Jessy and
Elliott, Desmond",
booktitle = "Proceedings of the 2026 Conference on Empirical Methods in Natural Language Processing",
month = oct,
year = "2026",
address = "Budapest, Hungary",
publisher = "Association for Computational Linguistics",
url = "https://aclanthology.org/2026.emnlp-main.443/",
pages = "10214--10230",
ISBN = "979-8-89176-459-0",
abstract = "We test the standard RLVR tool-use recipe (GRPO on Qwen2.5-7B-Instruct) on a deliberately minimal knowledge-graph tool API: four Freebase navigation verbs over Complex WebQuestions. Under a self-verifiable retrieval reward, the policy{'}s correct-via-tool rate climbs from 3.8 to 9.6 over 250 steps, then collapses to 0 within a single 50-step window, a peak-then-collapse pattern replicated across four seeds. Across seven reward designs, we find four recurring failure modes: adding denser or more targeted proxy rewards shifts the failure mode rather than eliminating it. We argue that a key difference from Python interpreters, web search, and JSON APIs is interface feedback. Their failures often leak natural-language signal the model saw in pretraining. A Python traceback names the failing line; an empty Freebase result [] does not. Stripping away that surface exposes a degradation regime that same-family reward redesigns do not fix. A direct oracle ablation rules out relation selection: injecting gold relations at every retrieval call lifts exact-match accuracy by only +0.20 pp, and 95.4 of retrieval-dependent errors are retrieval-composition failures rather than answer-extraction failures. As a mitigation, one-iteration self-distillation reaches 40 EM at 7B; the best 14B model we train reaches 40.2 (a gap within sampling noise) and initialisation barely matters, so within the 7B{--}14B range tested the {\textasciitilde}40 ceiling appears interface-bound rather than capacity-bound.\par Paper Type: Long"
}<?xml version="1.0" encoding="UTF-8"?>
<modsCollection xmlns="http://www.loc.gov/mods/v3">
<mods ID="sun-kazakov-2026-peak">
<titleInfo>
<title>Peak-Then-Collapse and the Four Interface Channels of Knowledge-Graph Tool Use</title>
</titleInfo>
<name type="personal">
<namePart type="given">Tianda</namePart>
<namePart type="family">Sun</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Dimitar</namePart>
<namePart type="family">Kazakov</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</title>
</titleInfo>
<name type="personal">
<namePart type="given">Hung-Yi</namePart>
<namePart type="family">Lee</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Sunipa</namePart>
<namePart type="family">Dev</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Jessy</namePart>
<namePart type="family">Li</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Desmond</namePart>
<namePart type="family">Elliott</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">979-8-89176-459-0</identifier>
</relatedItem>
<abstract>We test the standard RLVR tool-use recipe (GRPO on Qwen2.5-7B-Instruct) on a deliberately minimal knowledge-graph tool API: four Freebase navigation verbs over Complex WebQuestions. Under a self-verifiable retrieval reward, the policy’s correct-via-tool rate climbs from 3.8 to 9.6 over 250 steps, then collapses to 0 within a single 50-step window, a peak-then-collapse pattern replicated across four seeds. Across seven reward designs, we find four recurring failure modes: adding denser or more targeted proxy rewards shifts the failure mode rather than eliminating it. We argue that a key difference from Python interpreters, web search, and JSON APIs is interface feedback. Their failures often leak natural-language signal the model saw in pretraining. A Python traceback names the failing line; an empty Freebase result [] does not. Stripping away that surface exposes a degradation regime that same-family reward redesigns do not fix. A direct oracle ablation rules out relation selection: injecting gold relations at every retrieval call lifts exact-match accuracy by only +0.20 pp, and 95.4 of retrieval-dependent errors are retrieval-composition failures rather than answer-extraction failures. As a mitigation, one-iteration self-distillation reaches 40 EM at 7B; the best 14B model we train reaches 40.2 (a gap within sampling noise) and initialisation barely matters, so within the 7B–14B range tested the ~40 ceiling appears interface-bound rather than capacity-bound.\par Paper Type: Long</abstract>
<identifier type="citekey">sun-kazakov-2026-peak</identifier>
<location>
<url>https://aclanthology.org/2026.emnlp-main.443/</url>
</location>
<part>
<date>2026-10</date>
<extent unit="page">
<start>10214</start>
<end>10230</end>
</extent>
</part>
</mods>
</modsCollection>
%0 Conference Proceedings
%T Peak-Then-Collapse and the Four Interface Channels of Knowledge-Graph Tool Use
%A Sun, Tianda
%A Kazakov, Dimitar
%Y Lee, Hung-Yi
%Y Dev, Sunipa
%Y Li, Jessy
%Y Elliott, Desmond
%S Proceedings of the 2026 Conference on Empirical Methods in Natural Language Processing
%D 2026
%8 October
%I Association for Computational Linguistics
%C Budapest, Hungary
%@ 979-8-89176-459-0
%F sun-kazakov-2026-peak
%X We test the standard RLVR tool-use recipe (GRPO on Qwen2.5-7B-Instruct) on a deliberately minimal knowledge-graph tool API: four Freebase navigation verbs over Complex WebQuestions. Under a self-verifiable retrieval reward, the policy’s correct-via-tool rate climbs from 3.8 to 9.6 over 250 steps, then collapses to 0 within a single 50-step window, a peak-then-collapse pattern replicated across four seeds. Across seven reward designs, we find four recurring failure modes: adding denser or more targeted proxy rewards shifts the failure mode rather than eliminating it. We argue that a key difference from Python interpreters, web search, and JSON APIs is interface feedback. Their failures often leak natural-language signal the model saw in pretraining. A Python traceback names the failing line; an empty Freebase result [] does not. Stripping away that surface exposes a degradation regime that same-family reward redesigns do not fix. A direct oracle ablation rules out relation selection: injecting gold relations at every retrieval call lifts exact-match accuracy by only +0.20 pp, and 95.4 of retrieval-dependent errors are retrieval-composition failures rather than answer-extraction failures. As a mitigation, one-iteration self-distillation reaches 40 EM at 7B; the best 14B model we train reaches 40.2 (a gap within sampling noise) and initialisation barely matters, so within the 7B–14B range tested the ~40 ceiling appears interface-bound rather than capacity-bound.\par Paper Type: Long
%U https://aclanthology.org/2026.emnlp-main.443/
%P 10214-10230
Markdown (Informal)
[Peak-Then-Collapse and the Four Interface Channels of Knowledge-Graph Tool Use](https://aclanthology.org/2026.emnlp-main.443/) (Sun & Kazakov, EMNLP 2026)
ACL