What I Learned at ICML 2025 Will Surprise You!

The ICML 2025 conference brought a whirlwind of innovation across every corner of machine learning — and we’re here to unpack the juiciest bits. Whether you're into reinforcement learning, NLP, vision models, or diffusion-based generation, this year’s papers had something wow-worthy for everyone.

🔹 AdaSplash: Adaptive Curriculum Learning via Soft Sampling

• Proposes AdaSplash, which dynamically adjusts training sample importance using gradient norm and loss.
• Avoids hard filtering by soft-sampling informative and stable examples.
• Shows theoretical convergence and outperforms baselines in various benchmarks.

---

🔹 TESS: Training with Exploration-Sensitive Sampling

• Tackles exploration in long-horizon decision-making tasks.
• Prioritizes rare transitions using an exploration-sensitive score.
• Integrates seamlessly into model-based RL training pipelines.

---

🔹 InstructZero: Multi-Task Zero-Shot Agent with Latent Program Inference

• Introduces a zero-shot agent trained via instruction and latent program alignment.
• Capable of multi-task generalization without fine-tuning.
• Uses inferred latent programs to align with natural language instructions.

---

🔹 B2C: Back-to-Context for Scalable In-Context RL

• Proposes an in-context RL method that uses demonstrations by resetting to past states.
• Improves sample efficiency and stability in offline and online RL settings.
• Works in both discrete and continuous control tasks.

---

🔹 LM Data Valuation via Shapley Scores

• Applies Shapley value theory to quantify the contribution of pretraining data.
• Identifies harmful and redundant subsets of training data for language models.
• Helps reduce dataset size while maintaining or improving performance.

---

🔹 PIGLeT: Pretraining with Grounded Language and Temporal Goals

• Introduces a pretraining method grounded in both language and temporal goals.
• Improves generalization in temporal reasoning and goal-directed tasks.
• Evaluated on simulation environments with strong performance.

---

🔹 ELICIT: Uncertainty-Aware Curriculum for Offline RL

• Builds a curriculum based on epistemic uncertainty estimation.
• Improves exploration and policy learning in static offline datasets.
• Outperforms several baselines on D4RL benchmarks.

---

🔹 MAGNify: Memory-Augmented Graph Networks for Long-Horizon Reasoning

• Combines graph neural networks with memory modules to handle long-term dependencies.
• Useful for physical reasoning and decision-making tasks.
• Shows superior performance in multi-step inference problems.

---

🔹 DataComp: A Data Competition for Pretraining Robust Vision Models

• Introduces a new benchmark for evaluating vision model pretraining data quality.
• Encourages better dataset design over sheer size.
• Shows that smaller, curated datasets can outperform massive noisy datasets.

---

🔹 D3PM-GEN: Diffusion Models for Structured Sequence Generation

• Extends diffusion models to structured outputs like text or code.
• Avoids exposure bias and improves generation quality.
• Achieves strong results on sequence modeling benchmarks.

---

🔹 Fleet of Agents: Coordinated Problem Solving with Large Language Models

• Focuses on coordinating multiple LLMs working as a “fleet” to collaboratively solve problems.
• Likely explores agent communication, task decomposition, and orchestration strategies.

---

🔹 Consistently Preserving Safety when Fine-Tuning LLMs on Diverse Datasets

• Investigates how to maintain safety guarantees when adapting LLMs with varied or noisy datasets.
• Important for deploying LLMs in real-world, high-stakes environments.

---

🔹 OTTER: A Vision-Language-Action Model

• Proposes a model that integrates visual understanding, language comprehension, and actionable decision-making.
• Likely contributes to embodied AI or robotics tasks.

---

🔹 CMoS: Rethinking Time Series Prediction Through the Lens of Chunk-wise Spatial Correlations

• Introduces a new perspective on time series forecasting using spatial (chunk-wise) correlations.
• Enhances interpretability and performance in sequential data modeling.

---

🔹 Context is Key: A Benchmark for Forecasting with Essential Textual Information

• Provides a benchmark that fuses time series with textual data for improved forecasting.
• Encourages models to interpret external textual cues alongside numerical trends.

---

🔹 Adaptive Multi-prompt Contrastive Network for Few-shot Out-of-distribution Detection

• Uses multiple adaptive prompts and contrastive learning to detect OOD inputs with few samples.
• Addresses generalization and robustness in low-data settings.

---

🔹 Hierarchical Planning for Complex Tasks with Knowledge Graph-RAG and Symbolic Verification

• Combines retrieval-augmented generation (RAG) with symbolic verification for complex hierarchical planning.
• Likely useful for multi-step reasoning in AI agents and automation tasks.

---

🔹 WOMD-Reasoning: A Large-Scale Dataset for Interaction Reasoning in Driving

• Introduces a dataset centered on reasoning about agent interactions in autonomous driving.
• Targets better modeling of human-like decision-making in self-driving systems.

---

🔹 SeedLoRA: A Fusion Approach to Efficient LM Fine-Tuning

• A method that fuses LoRA (Low-Rank Adaptation) with other efficient fine-tuning techniques.
• Promises resource-efficient updates to large models.

---

🔹 Metastable Dynamics of Chain-of-Thought Reasoning: Provable Benefits of Search, RL and Distillation

• Theoretically explores how reasoning chains evolve during training and inference.
• Combines search methods, reinforcement learning, and distillation for better CoT performance.

---

🔹 Leveraging Skills from Unlabeled Prior Data for Efficient Online Exploration

• Focuses on skill reuse from unlabeled data to bootstrap exploration in online RL.
• Helps reduce data requirements and improve learning speed.

---

🔹 The Limits of Predicting Agents from Behaviour

• Discusses challenges in predicting future actions of agents based on observed behavior.
• Questions generalization in behavior modeling.

---

🔹 General Agents Need World Models

• Argues for the necessity of explicit world modeling in developing general-purpose agents.
• Reinforces the value of structured environments and state representations.

---

🔹 ActiveDPO: Active Direct Preference Optimization for Sample-Efficient Alignment

• Enhances DPO (Direct Preference Optimization) with active sampling for alignment training.
• Reduces required human feedback while improving model preference alignment.

---

🔹 Aligned Textual Scoring Rule

• Applies scoring rules (used in review or grading contexts) to align LLM-generated responses.
• Could support applications like automated feedback, peer review, or interviews.

---

🔹 Few-shot Steerable Alignment: Adapting Reward and LLM Policies with Neural Process

• Uses neural processes to steer model outputs based on few-shot alignment signals.
• Combines policy learning with meta-learning techniques.

---

🔹 Personalization and Pluralistic Alignment via RL Fine-tuning

• Explores ways to align models not just to one truth, but to pluralistic and user-specific values.
• Uses reinforcement learning for more adaptive and personalized models.