ICLR 2026 Orals
← All orals

Trends across the 2026 Orals

Themes identified by Claude Sonnet 4.6 across the author-stated future-work and limitations sections of every oral paper. Each theme cites the papers that contribute to it. See how AI content is labeled.

Theme · 10 papers

Reinforcement learning as the primary post-training axis for LLMs

Across papers on reasoning, code generation, and instruction following, authors consistently identify RL-based fine-tuning—GRPO, PPO variants, trust-region methods—as the key next step beyond supervised learning. Multiple papers note current experiments are limited to small models (≤14B) or narrow task distributions and call for scaling RL to larger architectures, multi-turn settings, and agentic interaction. The convergence on RL as the dominant lever for eliciting complex behaviors marks a field-wide shift away from pure SFT pipelines.

“Apply framework to multi-turn RL, agentic interaction, and long-form reasoning”
— The Art of Scaling Reinforcement Learning Compute for LLMs
“Scale TROLL to larger models and Mixture-of-Experts architectures”
— TROLL: Trust Regions Improve Reinforcement Learning for Large Language Models
“Scaling method to larger architectures with adequate compute”
— Overthinking Reduction with Decoupled Rewards and Curriculum Data Scheduling
10 contributing papers
Theme · 10 papers

Extending models to longer contexts and sequences at inference time

A recurring limitation across language, vision, and audio models is the mismatch between training-time sequence lengths and deployment demands. Authors frequently cite the need for improved positional encodings, KV-cache compression, and memory architectures to handle longer inputs without retraining. Several papers also note the theoretical gap between results proved for short sequences and real-world long-context behaviors, and call for both architectural and theoretical work to bridge it.

“Develop more advanced memory architectures and training strategies for enhancing long-context capabilities of LLMs”
— MemAgent: Reshaping Long-Context LLM with Multi-Conv RL-based Memory Agent
“Characterize how different positional embedding schemes affect minimum training length”
— Quantitative Bounds for Length Generalization in Transformers
“Call on LLM builders to prioritize multi-turn reliability, as known remediations for simpler settings prove ineffective”
— LLMs Get Lost In Multi-Turn Conversation
10 contributing papers

LLMs Get Lost In Multi-Turn Conversation

Study showing LLMs exhibit 39% average performance drop in multi-turn conversations, failing to recover from wrong contextual assumptions.

Avg rating: 8.00 (6–10) · Philippe Laban et al.
Theme · 11 papers

Extending unimodal or text-centric models to video, audio, and additional modalities

Papers working on language, image, or audio systems uniformly name video understanding and generation as an outstanding next step, and many additionally flag missing modalities such as 3D, tabular, radar, or multilingual data. The pattern is consistent: benchmarks, reward models, and generation pipelines built for text or static images are noted as needing systematic generalization to richer temporal and sensory modalities. This reflects a convergence on omni-modal systems as the target architecture.

“Incorporate video understanding and generation tasks into evaluation system”
— Omni-Reward: Towards Generalist Omni-Modal Reward Modeling with Free-Form Preferences
“Extend ADP beyond text to images, screen recordings, and multimodal data”
— Agent Data Protocol: Unifying Datasets for Diverse, Effective Fine-tuning of LLM Agents
“Build unified audio representation for more scalable joint training”
— UALM: Unified Audio Language Model for Understanding, Generation and Reasoning
11 contributing papers

Latent Speech-Text Transformer

Aggregates speech tokens into latent patches for efficient speech-text modeling with cross-modal alignment.

Avg rating: 6.00 (2–10) · Yen-Ju Lu et al.
Theme · 10 papers

Scalable benchmarks and simulation environments for training and evaluating AI agents

Multiple papers building evaluation suites or agent training environments note that current benchmarks are limited to short, simple tasks and that simulation fidelity, coverage of realistic adversarial conditions, and scalable automated evaluation remain open problems. Authors call for environments that support RL fine-tuning loops, multi-step interaction, and adversarial injection, pointing toward a community-wide infrastructure effort to support the next generation of agent research.

“Use OpenApps for scaling agent training pipelines via RL fine-tuning”
— OpenApps: Simulating Environment Variations to Measure UI Agent Reliability
“Developing fresh benchmark problems using latest scientific knowledge, contamination-resistant and past training cutoff dates”
— AstaBench: Rigorous Benchmarking of AI Agents with a Scientific Research Suite
“Enable agents to learn through trial and error inside simulator rather than imitating recorded examples”
— SimuHome: A Temporal- and Environment-Aware Benchmark for Smart Home LLM Agents
10 contributing papers
Theme · 10 papers

Detecting and defending against adversarial manipulation of LLMs at fine-tuning and inference

Papers on backdoor attacks, steganographic triggers, reward hacking, and model fingerprinting share a common structure: a threat is demonstrated at small scale and authors call for mitigation strategies, extension to larger models, and evaluation against adaptive adversaries. The repeated limitation that experiments are confined to models ≤3B–13B reveals a gap between the studied threat surface and the deployed model scale, and points toward scalable detection and defense as a pressing research direction.

“Develop technical mitigations for finetuning-activated attacks”
— Watch your steps: Dormant Adversarial Behaviors that Activate upon LLM Finetuning
“Evaluate TRACE on more realistic, heterogeneous loopholes”
— Is it Thinking or Cheating? Detecting Implicit Reward Hacking by Measuring Reasoning Effort
“Redesign deception benchmarks using statistical methods for detecting deception rather than assuming correctness of LLM responses”
— Beyond Prompt-Induced Lies: Investigating LLM Deception on Benign Prompts
10 contributing papers
Theme · 10 papers

Mechanistic interpretability: better tools for attributing and auditing model internals

Authors working on sparse autoencoders, causal interventions, computational graphs, and topological analysis of representations consistently note that current tools are insufficiently precise, do not scale to large models, and rely on unvalidated assumptions about linear representations. Future work directions coalesce around more faithful attribution methods, semantic grounding of identified features, and automated verification pipelines, suggesting that interpretability is moving from exploratory to engineering-grade.

“Improve interpretability tools such as more faithful sparse autoencoders and more precise attribution methods”
— Verifying Chain-of-Thought Reasoning via Its Computational Graph
“Use learned features as state trackers for detecting significant changes in model behavior”
— Temporal Sparse Autoencoders: Leveraging the Sequential Nature of Language for Interpretability
“Extract concepts and features from low-rank representation space for model-agnostic interpretability”
— Sequences of Logits Reveal the Low Rank Structure of Language Models
10 contributing papers
Theme · 8 papers

World models for embodied and robotic agents: dynamics, planning, and sim-to-real transfer

Papers on latent world models, compositional planning, and robot learning share a common roadblock: models trained on narrow domains (robotics, video games) must be generalized to diverse real-world settings, and current approaches lack reliable dynamics modeling and multi-modal conditioning. Authors point toward physics-guided training, integration of language and reward signals, and evaluation on real robotic hardware as the required next steps, reflecting a convergence on interactive simulation as foundational infrastructure for embodied AI.

“Enable unified multi-modal conditioning with simultaneous action, language and image signals”
— Latent Particle World Models: Self-supervised Object-centric Stochastic Dynamics Modeling
“Physics-guided motion generation and physics-aware reinforcement post-training for precise dynamics modeling”
— World-In-World: World Models in a Closed-Loop World
“Benchmark on large-scale real robotic datasets”
— TD-JEPA: Latent-predictive Representations for Zero-Shot Reinforcement Learning
8 contributing papers