Computing professor at Georgia Tech - NLP, ML, AI
Atlanta, GA
Joined March 2009
- Tweets438
- Following1.2K
- Followers5.2K
- Likes3.6K
Come check out our tutorial on multilingual and multicultural LLMs today at 2pm #EMNLP2026
emnlp2026-multilingual-cultuโฆ
En route to Budapest ๐ญ๐บ for #EMNLP2026! Come chat at our posters in Hall H:
Monday, Oct 26:
9:00โ10:30 โ Harm amplification in LLM interactions
arxiv.org/abs/2606.02423
11:00โ12:30 โ LaRA: detecting data contamination in RL post-training
arxiv.org/abs/2605.29888
16:45โ18:15 โ Who Pays More for Safety?
arxiv.org/abs/2608.22490
Tuesday, Oct 27:
9:00โ10:30 โ Camellia: cultural biases in LLMs for Asian languages
arxiv.org/abs/2510.05291
Looking forward to seeing everyone!
Alan Ritter retweeted
๐ชฝDFlash draft models are available for Ornith-1.5 9B, 35B-A3B, and 397B.
DFlash and Ornith 1.5 make a dynamic duo: DFlash performs as the draft model, accelerating inference by drafting multiple tokens at once. Then Ornith 1.5 verifies the draft tokens in one pass instead of generating one token at a time.
๐ฎ We conducted a performance evaluation of the DFlash models. DFlash achieves up to a 2.54ร speedup without any loss in generation quality.
๐ Come and plug this super buddy into your workstation.
huggingface.co/collections/oโฆ
Alan Ritter retweeted
๐ฆOrnith 1.5 series models are now available on ollama!
๐ซกThanks to our awesome friends at @ollama
๐Come give Ornith 1.5 a try!
ollama.com/library/ornith-1.โฆ
Aloha! ๐บIntroducing Ornith-1.5, a family of open-source LLMs spanning 9B Dense, 35B MoE, and 397B MoE, trained with self-improving strategies.
It achieves state-of-the-art performance among open-source models of comparable size and delivers performance comparable to Claude Opus 4.8 across reasoning, agentic, and coding tasks:
โ
Terminal-Bench 2.1 (86.1)
โ
SWE-Bench (86 on verified, 65.1 on pro, 79.6 on Multilingual)
โ
DeepSWE (56)
โ
HLE (44.6)
โ
ClawEval (81.4)
โ
Tool Decathlon (71.2)
Ornith-1.5 takes a major step toward training foundation models through end-to-end self-improvement, extending the self-scaffolding strategies introduced in Ornith-1.0 into a more complete self-improvement loop: the model proposes new tasks, generates task-specific scaffolds, and produces solution rollouts for reinforcement learning, continuously creating new learning experiences from which it can improve.
All models, along with their quantized versions (FP8, GGUF, MLX, and NVFP4), have been released under the MIT License, enabling unrestricted commercial and research use.
๐Tech Blog: ornith.ai/ornith_1_5.html
๐คHuggingface: huggingface.co/collections/oโฆ
Alan Ritter retweeted
My team at meta is hiring a 1-year role! Come work with me!
us.meta.talentnet.community/โฆ
Qs I'm working on:
What to eval on/hill climb during pretraining to maximize post-train perf?
What evals are most correlated/predictive across pretrain -> midtrain -> post-train?
How to datamix?
Alan Ritter retweeted
Classic tech bro being tech bro. Recruits from academia, builds on academia's innovation, outbids academia for every GPU on earth, publishes nothing peer-reviewed, gives back pennies, then declares academia "mainly fraud." That's pulling up the ladder and spitting down it.
Alan Ritter retweeted
There's an OpenReview profile of someone who:
- Was a postdoc at the IAS
- Has many papers with respected coauthors, approx. 15k citations
- Is now an independent researcher
- Submitted 54 ICLR papers (!!!)
- Had two of them accepted
What is this person actually trying to do?
Alan Ritter retweeted
We're organizing the Workshop on Evaluation of Interactive Agents at NeurIPS 2026!
Consider submitting your work (non-archival) and joining us in Atlanta.
Topics include (not limited to): evaluation benchmarks, trajectory-level metrics, user/environment simulation, long-horizon reliability, safety, and ethics.
โฐ Deadline: Aug. 29 (AoE)
๐ Website: eval-interactive-agents-workโฆ
#NeurIPS
Alan Ritter retweeted
๐ซกAt DeepReinforce, we believe that open weights matter.
๐งEveryone should have the right to control their own intelligence, protect their data, and shape how AI works for them.
๐งโ๐ณ๐ฉโ๐ณThe team is cooking. More to come!!
๐ฆThe Ornith-1.0 family has surpassed 12M downloads on Hugging Face in 30 days.
โค๏ธThanks to everyone who has tried Ornith and shared feedback with us. The team is working to push open-weight LLM forward.
๐ซกWe believe open weights matter. Everyone should have the freedom to control their own intelligence, protect their data, and shape how AI works for them.
๐ซถMore to come!
Alan Ritter retweeted
Excited to share my first work after joining @MSFTResearch!
LLMs have entered the agentic era, and we now collaborate with agents on complex tasks over many turns of interaction.
But does your agent actually follow what you intended?
We show where agents get lost: LLMs Get Lost in Evolving User Intent.
๐ arxiv.org/abs/2607.20734 ๐งต 1/N
Itโs great to see so much excitement around self-distillation!
However, some recent work highlights an important failure mode: when the teacher is given the solution, it can suppress verification, backtracking, and exploration, which make reasoning models effective.
We encountered a closely related issue. A teacher conditioned on an expert solution may skip crucial steps, or rationalize its way to the provided answer.
We just got back from presenting this work at ICML ๐ฐ๐ท. Our method combines mixed-policy generation with a contrastive objective to discourage shortcuts while still learning effectively from expert solutions on the most challenging instances.
Excited to be heading to ICML! I will be presenting our paper on rethinking how to train models with expert solutions using self-distillation.
๐๏ธ July 8, 2026, 5:00 PM โ 6:45, Hall A #2502
Alan Ritter retweeted
๐๐ถ๐ณ๐ฒ ๐จ๐ฝ๐ฑ๐ฎ๐๐ฒ! ๐
I'm excited to share that I've ๐ฆ๐๐ฐ๐ฐ๐ฒ๐๐๐ณ๐๐น๐น๐ ๐ฑ๐ฒ๐ณ๐ฒ๐ป๐ฑ๐ฒ๐ฑ my M.S. thesis at Yonsei University on Revisiting the Uniform Information Density Hypothesis in LLM Reasoning (lnkd.in/g-SvEwDT) and will be joining ๐๐ฒ๐ผ๐ฟ๐ด๐ถ๐ฎ ๐ง๐ฒ๐ฐ๐ต this fall as an M.S. student in Computer Science! Details are in my homepage (lnkd.in/gXFk-gV6)
Before then, I'll be presenting two recent works:
๐ ๐๐๐ ๐ฎ๐ฌ๐ฎ๐ฒ (San Diego)
Revisiting the Uniform Information Density Hypothesis in LLM Reasoning
๐ ๐ฃ๐ผ๐๐๐ฒ๐ฟ ๐ฆ๐ฒ๐๐๐ถ๐ผ๐ป: July 6, 11:00โ12:30
๐ ๐๐ฟ๐ฎ๐ป๐ฑ ๐๐ฎ๐น๐น (๐ฆ๐ฒ๐๐๐ถ๐ผ๐ป ๐ต: ๐ข๐ฟ๐ฎ๐น/๐ฃ๐ผ๐๐๐ฒ๐ฟ๐/๐๐ฒ๐บ๐ผ๐ ๐)
๐ ๐๐๐ ๐ ๐ฎ๐ฌ๐ฎ๐ฒ ๐ง๐ฟ๐๐๐๐๐ผ๐ฟ๐๐ต๐ ๐๐ ๐ณ๐ผ๐ฟ ๐๐ผ๐ผ๐ฑ ๐ช๐ผ๐ฟ๐ธ๐๐ต๐ผ๐ฝ (Seoul)
LaRA: Layer-wise Representation Analysis for Detecting Data Contamination in RL Post-Training
๐ July 10 (KST)
๐ ๐๐ฎ๐น๐น๐ฟ๐ผ๐ผ๐บ ๐ญ๐ฌ๐ฏ
If you're attending either conference, I'd love to connectโfeel free to stop by the poster or reach out for a coffee chat!
Alan Ritter retweeted
Excited to be heading to ICML! I will be presenting our paper on rethinking how to train models with expert solutions using self-distillation.
๐๏ธ July 8, 2026, 5:00 PM โ 6:45, Hall A #2502
Alan Ritter retweeted
๐ New research from @scale_AI
Frontier SWE benchmarks are usually single-turn, one-shot tasks: the agent gets a detailed spec upfront, then implements autonomously.
That is not how most real coding-agent workflows feel.
Introducing SWE-Interact.
๐งต
Excited to attend #ICML2026 in Seoul ๐ฐ๐ท together with my Ph.D. students: @EthanMendes3, @hyungjoochae, @CherylolGuo!
Weโll be presenting work on:
- Multilingual reasoning
- Self-distillation from expert reasoning
- Cloning synthetic web environments
Poster info + paper links below. Looking forward to catching up!
Here are the papers we'll be presenting:
Didactic to Constructive: Turning Expert Solutions into Learnable Reasoning
@EthanMendes3 , @jungsoo___park, @alan_ritter
arXiv: arxiv.org/abs/2602.02405
Poster: Wed, Jul 8 โข 5:00โ6:45 PM KST, Hall A (#2502)
Safe and Scalable Web Agent Learning via Recreated Websites
@hyungjoochae , @jungsoo___park, @alan_ritter
arXiv: arxiv.org/abs/2603.10505
Poster: Tue, Jul 7 โข 10:30 AMโ12:15 PM KST, Hall A (#1714)
Learning to Route Languages for Multilingual Preference Optimization
@CherylolGuo, Hiromi Wakaki, @mittu1204, @alan_ritter, @cocoweixu
arXiv: arxiv.org/abs/2605.25360
Poster: Tue, Jul 7 โข 10:30 AMโ12:15 PM KST, Hall A (#402)
Alan Ritter retweeted
Knowledge is not evenly distributed across languages. For a given question, what if better information comes from another language?
๐ฅย We propose Language-Routed Policy Optimization (LRPO), an online reinforcement learning framework that lets LLMs explore and learn from cross-lingual generations during training.
#ICML2026
Alan Ritter retweeted
.@gtcomputing's @alan_ritter and his Ph.D. student Ethan Mendes have been named to the 2026 list of @MSFTResearch Fellows and Fellowship Advisors. Their research develops artificial intelligence agents that function as teammates. b.gatech.edu/4uILpL3
ALT Alan Ritter
Alan Ritter retweeted
Real-world LLM interactions are conversational, yet LLMs are far more vulnerable to multi-turn than single-turn attacks.
Introducing DialTree: RL + tree search for automated multi-turn attack discovery.
๐ 81.5% avg ASR@1 across 12 LLMs
๐ +44.2% over prior SOTA
#ICLR2026 [1/6]
Alan Ritter retweeted
Checkout our new benchmark, SWE-Atlas Test Writing from @scale_AI . Its the second benchmark in the SWE-Atlas Evaluation suite.
High quality test writing tasks - Agentic, Underspecified, Professional grade evaluation.
Frontier models score <45% and GPT 5.4 leads the pack.
