Emre Can Acikgoz

I am a PhD fellow in Computer Science at UIUC advised by Dilek Hakkani-Tur and Gokhan Tur. I work on Self-Evolving Agents. I am currently an Applied Scientist Intern at Amazon Alexa AI (Bellevue).

I completed my MSc in Computer Science at Koc University, where I focused on Large Language Models and Post-Training under the supervision of Deniz Yuret. Previously, I earned my BSc in Electrical and Electronics Engineering (AI focus) from the same institution, during which I also worked with Deniz Yuret on supervised and unsupervised morphological analysis.

If you want to discuss your research (non-commercial) or seek for an advice/feedback (non-industrial), feel free to schedule a short chat-meeting.

Email  /  GitHub  /  HuggingFace  /  Google Scholar  /  LinkedIn

Research

My research focuses on Self-Evolving Agents. To achieve this, I work on Agent Learning and Test-Time Training. My high-level goal is to build agents that can continuously learn from few-samples, with minimal updates; regardles of the task, domain, and environment.

In parallel to that, I like to explore the capabilities of LLMs, identify their limitations, and enhance them to new tasks and domains. My previous work spans a wide range of topics around language modeling and post-training, including multi-modality.

2025

We propose TD-EVAL, a two-step evaluation framework for TOD systems that combines fine-grained turn-level analysis—focusing on conversation cohesion, knowledge consistency, and policy compliance—with dialogue-level comparisons via a pairwise TOD Agent Arena. This unified approach captures both local and global errors missed by traditional metrics.

In ToolRL, we explored different reward strategies for these issues and present a new approach for tool utilization tasks that achieves 17% improvement over base models and 15% over SFT versions.

Our proposed taxonomy systematically analyzes Conversational Agents around three essential dimensions: (i) Reasoning—logical and structured thinking for decision-making, (ii) Monitoring—self-awareness and continuous user intention tracking, (iii) Control—effective tool utilization and policy adherence, all together with the representative list of works.

Inspired by human metacognition, SMART enhances LLM's self-awareness to reduce tool overuse while boosting performance. Our experiments show that SMARTAgent reduces tool use by 24% while improving performance by 37%.

CoALM unifies multi-turn dialogue management and complex API usage in a single model. Trained on the CoALM-IT multi-task dataset, CoALM (8B, 70B, 405B) outperforms domain-specific models like GPT-4o on MultiWOZ 2.4, BFCL V3, and API-Bank benchmarks.

Large language models (LLMs) have driven a paradigm shift in task-oriented dialogue by enabling AI agents with stronger reasoning, tool use, and instruction-following abilities. These developments give rise to conversational AI agents—systems that merge advanced language understanding with agentic decision-making to achieve dynamic, context-aware, and task-oriented interactions. This work covers ongoing challenges include multi-turn context management, controllability, personalization, and user alignment.

PIPA is a unified evaluation protocol for task planning agents that models their behavior within a POMDP framework, enabling fine-grained assessment across the entire agentic process. Unlike traditional task completion metrics, PIPA uses atomic evaluation criteria to diagnose strengths and weaknesses in context understanding, tool use, and decision-making, aligning evaluation more closely with user satisfaction.

2024

ReSpAct is a framework that enables LLM agents to engage in interactive, user-aligned task-solving. It enhances agents' ability to clarify, adapt, and act on feedback.

This study evaluates the effectiveness of training strategies for large language models in low-resource languages like Turkish, focusing on model adaptation, development, and fine-tuning to enhance reasoning skills and address challenges such as data scarcity and catastrophic forgetting.

We present Hippocrates, an open-source LLM framework specifically developed for the medical domain. Also, we introduce Hippo, a family of 7B models tailored for the medical domain, fine-tuned from Mistral and LLaMA2 through continual pre-training, instruction tuning, and reinforcement learning from human and AI feedback.

ViLMA (Video Language Model Assessment) presents a comprehensive benchmark for Video-Language Models, starting with a fundamental comprehension test and followed by a more advanced evaluation for temporal reasoning skills.

2022

This paper describes the winning approaches in MRL: The 1st Shared Task on Multilingual Clause-level Morphology. Our submission, which excelled in all three parts of the shared task — inflection, reinflection, and analysis — won the first prize in each category.

Tutor

Comp547: Deep Unsupervised Learning (Spring'24)
Comp541: Deep Learning (Fall'23)
Comp542: Natural Language Processing (Spring'23)
Comp541: Deep Learning (Fall'22)
Comp547: Deep Unsupervised Learning (Spring'22)

Blogs

The Rise of Conversational AI Agents with Large Language Models (November 8, 2024)

Academic Service

Reviewer: AACL 2025, EMNLP 2025, NeurIPS 2025, ACL 2025, NeurIPS 2024

Two Truths and a Lie

Two truths and one fun fact about me: (i) I won national math olympiads twice, (ii) I made my own custom Harley Davidson when I was 16, (iii) I used to play college football.

(website template credits)