Hakki C. Karaimer

Staff Research Scientist at Samsung Research America (SRA)

[Armen Jeddi @armen_jeddi]

🚀 New paper: AVIS: Adaptive Visual Inference Scaling for Vision-Language Models
Most test-time scaling methods for vision-language models optimize a single axis of compute: either scale visual context or perform more reasoning rollouts.
AVIS jointly adapts both. It allocates compute across visual context scaling and visual reasoning scaling on a per-query basis, combining efficient visual token pruning with adaptive reasoning search.
📊 Across diverse image and video reasoning benchmarks, AVIS achieves a better accuracy–compute trade-off than methods that scale only visual context or only reasoning, while remaining deployment-friendly through shared-prefill inference and KV-cache reuse.
Vision-language inference is not just about seeing more or thinking longer. AVIS learns when to do each.
📄 Paper: https://t.co/KIJ2E7EMCw
🌐 Website: https://t.co/Y7x7qtD0a6

Amazing teamwork with @amirkz99 @karaimer @CSProfKGD @Babak_Taati

[Armen Jeddi @armen_jeddi]

🚀 New paper: GEAR: Genetic AutoResearch for Agentic Code Evolution
Most autonomous research agents search like greedy hill climbers: edit one program, run one experiment, and keep only the best result.
GEAR replaces this with a population-based frontier of research states, using mutation, crossover, and an evolvable search policy to preserve and recombine promising ideas.
📊 Under the same AutoResearch setup and 100-experiment budget, all GEAR variants outperform the baseline.
Research is not one path. GEAR gives agents a frontier.

📄 Paper: https://t.co/HgAN6LC9vg
🌐 Website: https://t.co/2jhAhtrU2E

#autoresearch #llm #agentic_ai #agentic_coding

@karaimer @CSProfKGD @Babak_Taati

[Armen Jeddi @armen_jeddi]

I’ll be in Rio 🇧🇷 presenting LoopFormer at #ICLR2026!
Stop by our poster in Pavilion 3 this Thursday, 10:30 AM–1 PM — or DM me if you’d like to chat.

[Armen Jeddi @armen_jeddi]

Loops, loops, loops… LoopFormer.
Looped Language Models are becoming one of the defining themes of 2026, and next week I’ll be presenting our paper, LoopFormer, at #ICLR2026.
Lmk if you want to chat about looped LMs in Rio.

[Armen Jeddi @armen_jeddi]

Website: https://t.co/8OAjpnheTp
Github: https://t.co/lNJBJH8V5g

[Armen Jeddi @armen_jeddi]

LoopFormer is the next step in Recursive Language Modeling (RLMs): it treats iterative refinement in looped Transformers as a trajectory in representation space, conditioning each loop on time and step size so the same model is both parameter-efficient and compute-adaptive.

[Armen Jeddi @armen_jeddi]

New Paper: (ICLR 2026): LoopFormer: Elastic-Depth Looped Transformers for Latent Reasoning via Shortcut Modulation
https://t.co/zC4r3WysU3

[Armen Jeddi @armen_jeddi]

🙏🚀

[DailyPapers @HuggingPapers]

Dive deeper into PC-GRPO, an innovative approach for VLM self-improvement using puzzle-based RL.

Explore the paper: https://t.co/svd7dAWwCa

Project details: https://t.co/DMRX9tMhkM

[DailyPapers @HuggingPapers]

Samsung introduces PC-GRPO for Vision-Centric Reasoning

This novel framework uses self-supervised visual puzzles & a difficulty-aware curriculum to train VLMs. It achieves state-of-the-art visual reasoning without costly annotations, improving stability and accuracy.

[AK @_akhaliq]

Puzzle Curriculum GRPO for Vision-Centric Reasoning
https://t.co/9TMg5hnYEg

[Armen Jeddi @armen_jeddi]

Thanks @_akhaliq for highlighting our work! https://t.co/Aws8BHlXiQ

[Armen Jeddi @armen_jeddi]

Thanks to my teammates at Samsung, UofT, and Vector Institute.
@karaimer @Babak_Taati

[Armen Jeddi @armen_jeddi]

https://t.co/zTRFWZxaFI

[Armen Jeddi @armen_jeddi]

We achieve the best performance on Qwen-VL-2.5 (3B/7B). We also propose a remedy for cleaning existing benchmarks and evaluate both our model and baselines on the cleaned sets. Our model consistently improves. See the project page for details.

[Armen Jeddi @armen_jeddi]

We identify a critical challenge in VLM GRPO post-training: the chain of reasoning and the final answer drift apart. We introduce the Reasoning/Answer Consistency (RAC) metric to quantify this, and we integrate consistency-enhancing reward schemes into our setup.

[Armen Jeddi @armen_jeddi]

Instead of noisy labels, we use verifiable vision pretext puzzles — Jigsaw, Rotation, PatchFit — for RLVR. Rewards are exact (Jigsaw gives partial credit). With a difficulty-aware curriculum, we overcome GRPO’s flat/sparse rewards.

[Armen Jeddi @armen_jeddi]

This realization forced us to change our entire approach. If human annotations are this noisy (and expensive), we shouldn’t rely on them for RLVR. So we asked: can we teach a VLM to reason without a single labeled image? Enter Puzzle Curriculum GRPO (PC-GRPO).

[Armen Jeddi @armen_jeddi]

Not just typos. We audited MMStar, SEEDBench, MME, and ColorBench with human raters + strong model committees. Found subjective answers, invisible details, and flat-out wrong facts. If 1/5 questions are wrong, how can we trust leaderboards or measure progress?

[Armen Jeddi @armen_jeddi]

🚨 New paper: I spent weeks digging into our VLM’s “failures” and found something unsettling: the model was right; the benchmark was wrong. 😑
Our audits show ~10–20% of “ground truth” in top vision benchmarks is incorrect or ambiguous. 🧵👇