@inproceedings{datta-etal-2026-early,
title = "From Early Encoding to Late Suppression: Interpreting {LLM}s on Character Counting Tasks",
author = "Datta, Ayan and
Marreddy, Mounika and
Mehler, Alexander and
Zhao, Zhixue and
Mamidi, Radhika",
editor = "Zhang, Yue and
Xu, Ruifeng and
Purwarianti, Ayu and
Rong, Tong and
Suzuki, Jun",
booktitle = "Findings of the {A}ssociation for {C}omputational {L}inguistics: {IJCNLP}-{AACL} 2026",
month = nov,
year = "2026",
address = "Hengqin, China",
publisher = "Association for Computational Linguistics",
url = "https://aclanthology.org/2026.findings-aacl.70/",
pages = "1370--1391",
ISBN = "979-8-89176-444-6",
abstract = "Large language models (LLMs) still fail at counting the characters in a word, even while they excel on far harder benchmarks. Using character counting (e.g., ``How many `p{'}s are in apple?'') as a controlled task, we find a consistent pattern across Llama, Qwen, and Gemma in both base and instruction-tuned variants. The models compute the correct answer internally but fall back on degenerate output strategies, such as always predicting ``1'' or guessing uniformly among ``1'', ``2'', and ``3''. Probing classifiers, activation patching, and logit lens analysis show that character-level information is encoded in early and mid layers, then actively attenuated by a small set of later components, mainly the penultimate and final layer MLP and sometimes the final layer attention heads. The failure is not missing information or insufficient scale but structured interference in the computation graph, which explains why these errors persist under scaling and instruction tuning."
}<?xml version="1.0" encoding="UTF-8"?>
<modsCollection xmlns="http://www.loc.gov/mods/v3">
<mods ID="datta-etal-2026-early">
<titleInfo>
<title>From Early Encoding to Late Suppression: Interpreting LLMs on Character Counting Tasks</title>
</titleInfo>
<name type="personal">
<namePart type="given">Ayan</namePart>
<namePart type="family">Datta</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Mounika</namePart>
<namePart type="family">Marreddy</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Alexander</namePart>
<namePart type="family">Mehler</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Zhixue</namePart>
<namePart type="family">Zhao</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Radhika</namePart>
<namePart type="family">Mamidi</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<originInfo>
<dateIssued>2026-11</dateIssued>
</originInfo>
<typeOfResource>text</typeOfResource>
<relatedItem type="host">
<titleInfo>
<title>Findings of the Association for Computational Linguistics: IJCNLP-AACL 2026</title>
</titleInfo>
<name type="personal">
<namePart type="given">Yue</namePart>
<namePart type="family">Zhang</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Ruifeng</namePart>
<namePart type="family">Xu</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Ayu</namePart>
<namePart type="family">Purwarianti</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Tong</namePart>
<namePart type="family">Rong</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Jun</namePart>
<namePart type="family">Suzuki</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<originInfo>
<publisher>Association for Computational Linguistics</publisher>
<place>
<placeTerm type="text">Hengqin, China</placeTerm>
</place>
</originInfo>
<genre authority="marcgt">conference publication</genre>
<identifier type="isbn">979-8-89176-444-6</identifier>
</relatedItem>
<abstract>Large language models (LLMs) still fail at counting the characters in a word, even while they excel on far harder benchmarks. Using character counting (e.g., “How many ‘p’s are in apple?”) as a controlled task, we find a consistent pattern across Llama, Qwen, and Gemma in both base and instruction-tuned variants. The models compute the correct answer internally but fall back on degenerate output strategies, such as always predicting “1” or guessing uniformly among “1”, “2”, and “3”. Probing classifiers, activation patching, and logit lens analysis show that character-level information is encoded in early and mid layers, then actively attenuated by a small set of later components, mainly the penultimate and final layer MLP and sometimes the final layer attention heads. The failure is not missing information or insufficient scale but structured interference in the computation graph, which explains why these errors persist under scaling and instruction tuning.</abstract>
<identifier type="citekey">datta-etal-2026-early</identifier>
<location>
<url>https://aclanthology.org/2026.findings-aacl.70/</url>
</location>
<part>
<date>2026-11</date>
<extent unit="page">
<start>1370</start>
<end>1391</end>
</extent>
</part>
</mods>
</modsCollection>
%0 Conference Proceedings
%T From Early Encoding to Late Suppression: Interpreting LLMs on Character Counting Tasks
%A Datta, Ayan
%A Marreddy, Mounika
%A Mehler, Alexander
%A Zhao, Zhixue
%A Mamidi, Radhika
%Y Zhang, Yue
%Y Xu, Ruifeng
%Y Purwarianti, Ayu
%Y Rong, Tong
%Y Suzuki, Jun
%S Findings of the Association for Computational Linguistics: IJCNLP-AACL 2026
%D 2026
%8 November
%I Association for Computational Linguistics
%C Hengqin, China
%@ 979-8-89176-444-6
%F datta-etal-2026-early
%X Large language models (LLMs) still fail at counting the characters in a word, even while they excel on far harder benchmarks. Using character counting (e.g., “How many ‘p’s are in apple?”) as a controlled task, we find a consistent pattern across Llama, Qwen, and Gemma in both base and instruction-tuned variants. The models compute the correct answer internally but fall back on degenerate output strategies, such as always predicting “1” or guessing uniformly among “1”, “2”, and “3”. Probing classifiers, activation patching, and logit lens analysis show that character-level information is encoded in early and mid layers, then actively attenuated by a small set of later components, mainly the penultimate and final layer MLP and sometimes the final layer attention heads. The failure is not missing information or insufficient scale but structured interference in the computation graph, which explains why these errors persist under scaling and instruction tuning.
%U https://aclanthology.org/2026.findings-aacl.70/
%P 1370-1391
Markdown (Informal)
[From Early Encoding to Late Suppression: Interpreting LLMs on Character Counting Tasks](https://aclanthology.org/2026.findings-aacl.70/) (Datta et al., Findings 2026)
ACL