Elamur Research⌕
☰
home / authors / Ilia Kulikov

Ilia Kulikov

Ilia Kulikov works on neural text generation and dialogue systems, with a focus on how the training objective of a language model shapes the text it produces. Their papers here ask why models trained purely by maximum likelihood drift toward bland, repetitive and sometimes self-contradictory output. They answer by changing the loss function itself rather than adjusting the search or sampling procedure used at decoding time. The work sits within text generation and pairs a new objective with careful diagnosis of specific failure modes.

The most cited of these papers, Neural Text Generation with Unlikelihood Training (2019), contends that much of the degeneration seen in generated text comes from the likelihood objective. Decoding tricks such as top-k or nucleus sampling leave the model's probabilities untouched. The paper introduces unlikelihood training, which keeps the ordinary term rewarding correct tokens but adds a penalty that pushes down the probability of selected negative candidates, for instance tokens that have already appeared. The follow-up, Don't Say That! (2020), carries the idea into conversation. It turns the loss into a general framework in which the choice of negative candidates and their weights is tuned to particular problems in dialogue agents: copying too heavily from the context, repeating phrases within a reply, leaning on frequent words, and contradicting earlier turns.

Kulikov's papers list affiliations with New York University and Meta, and also with Carnegie Mellon University. Co-authors on the work here include Sean Welleck, Kyunghyun Cho, Jason Weston and Stephen Roller, along with Emily Dinan and Y-Lan Boureau on the dialogue study.

Ilia Kulikov has 2 papers in our library, in 2019–2020, most of them on Text generation, Dialogue systems; often with Jason Weston, Kyunghyun Cho, Sean Welleck.

2 papers in our library · 2019–2020

current affiliationPhysikalisch-Technische Bundesanstalt, Braunschweig, Germany; Federal Institute For Materials Research and Testing, Berlin, Germany

OpenAlex · ORCID · h-index 11 · 55 works · 812 citations OpenAlex, 2026-10-14

2 works in our library · 853 citations in our library · 5 more works in our library · ⭐ Paper Star rank #2456

The profile is the library's match of the name; two people with one name are not always told apart.

Elsewhere: Google Scholar · Semantic Scholar · dblp · ACL Anthology

Affiliations

From the bylines of their papers (OpenAlex, 2026-10-14), ORCID and Wikidata; institutions on papers are matched by machine. 3 more institutions on single bylines are not listed.

Topics

Text generation1Dialogue systems1

Works by year

2020s

20201 paper

2010s

20191 paper

Papers

The most cited first (OpenAlex and our library).

  1. Neural Text Generation with Unlikelihood Training (2019 · 498 citations)
    Language models trained with standard maximum likelihood tend to produce dull text that repeats itself. They overuse frequent words and loop on the same phrases, especially with greedy or beam search. Earlier fixes such as top-k sampling, nucleus sampling and n-gram beam blocking only change how text is decoded. The…
  2. Don't Say That! Making Inconsistent Dialogue Unlikely with Unlikelihood Training (2020 · 214 citations)
    Generative dialogue models trained only with maximum likelihood show several weaknesses. They copy too much from the context, repeat n-grams inside a single reply, overuse frequent tokens, and can contradict what was said earlier. The authors extend the unlikelihood loss of Welleck et al. (2019a) into a general…

Also in our library

5 more works of theirs are in our library without a page here yet; they will appear as we publish them.

Co-authors here

6 co-authors with a page, the most papers together first.

Cited by

Papers of our library cite their works 853 times (of the 140 884 papers whose bibliographies we read); 74 citing papers with a page here →.

This page in Markdown → https://md.research.elamur.ai/free/en/authors/ilia-kulikov/