RL has emerged as a powerful paradigm for enhancing the instruction following capabilities of LLMs. While existing training recipes achieve substantial gains, we find that they suffer from **exploration bias towards easy instructions** when the training data has multiple instructions in a prompt. This bias is caused by two main reasons: 1) **the policy model’s initial ability to satisfy hard instructions is too low** to trigger successful exploration during RL training, so the optimization is biased to easy instructions and 2) canonical RL training recipes typically employ a cumulative reward (the number of instructions fulfilled), **treating all the instructions equally**; the policy model is biased to fulfill the easy instructions to get the same amount of rewards. To address these, we first propose two metrics to measure the exploration bias of instruction following and then introduce a two-stage framework to alleviate it: 1) Behavioral Bootstrapping: a lightweight rejection sampling fine-tuning before RL to activate hard instructions and 2) Scarcity-Aware Rewards: a new RL reward function that assigns rewards to instructions based on their empirical scarcity. Experiments show that the proposed metrics are highly correlated with model performance and our methods unleash the potential of RL training: our best models outperform the baselines by a significant margin across three verifiable instruction following benchmarks.

Rubric-based reward shaping provides interpretable and editable reward signals for fine-tuning LLMs via reinforcement learning (RL), but existing adaptive rubric methods typically update criteria from local evidence such as the current batch or instance-level comparisons. This local view discards diagnostic information produced during training, making it difficult to track recurring failures, evaluate previous rubric edits, or raise standards once earlier criteria become saturated. We introduce AMARIS, a Memory-Augmented Rubric Improvement System that grounds rubric updates in longitudinal training evidence. AMARIS stores rollout analyses, step-level summaries, and rubric update records in a persistent evaluation memory, then retrieves recent and semantically relevant history to revise rubrics. We evaluate AMARIS across science, medicine, instruction following, and creative writing under both global and instance-specific rubric settings, with policy models from the Qwen and Llama families trained by GRPO and PPO. AMARIS improves over static, local-adaptive, and memory-ablated baselines, such as +2.8 points on GPQA-Diamond and +2.2 points on IFBench over the strongest baselines, while analysis shows that memory reduces oscillatory rubric edits and supports a progression from early failure correction to later curriculum advancement. The final rubric set that AMARIS discovers is itself reusable, and the online trajectory contributes measurably beyond it. AMARIS runs asynchronously alongside the normal RL loop, reducing blocking latency relative to synchronous rubric updates. The source code is available at https://github.com/qualidea1217/AMARIS.

While Large Language Models (LLMs) excel at factual retrieval, they often struggle with the “curse of two-hop reasoning” in compositional tasks. Recent research suggests that parameter-sharing transformers can bridge this gap by forming a “Generalization Circuit” during a prolonged “grokking” phase. A fundamental question arises: Is a grokked model truly superior to its non-grokked counterparts? Furthermore, is the extensive computational cost of waiting for the grokking phase worthwhile? In this work, we conduct a mechanistic study to evaluate the Generalization Circuit’s role in knowledge assimilation and transfer. We demonstrate that: (i) The inference paths established by non-grokked and grokked models for in-distribution compositional queries are identical. This suggests that the “Generalization Circuit” does not represent the sudden acquisition of a new reasoning paradigm. Instead, we argue that grokking is the process of integrating memorized atomic facts into an easy-acquire, naturally established reasoning path. (ii) Achieving high accuracy on unseen cases after prolonged training and the formation of a certain reasoning path are not bound; they can occur independently under specific data regimes. (iii) Even a mature circuit exhibits limited transferability when integrating new knowledge, suggesting that “grokked” Transformers do not achieve a full mastery of compositional logic.

While large language models (LLMs) have been thoroughly evaluated for deductive and inductive reasoning, their proficiency in holistic rule learning in interactive environments remains less explored. We introduce RULEARN, a novel benchmark to assess the rule-learning abilities of LLM agents in interactive settings. In RULEARN, agents strategically interact with simulated environments to gather observations, discern patterns, and solve complex problems. To enhance the rule-learning capabilities for LLM agents, we propose IDEA, a novel reasoning framework that integrates the process of Induction, Deduction, and Abduction. The IDEA agent generates initial hypotheses from limited observations through abduction, devises plans to validate these hypotheses or leverages them to solve problems via deduction, and refines previous hypotheses through induction, dynamically establishing and applying rules that mimic human rule-learning behaviors. Our evaluation of the IDEA framework, which involves five representative LLMs, demonstrates significant improvements over the baseline. Furthermore, our study with human participants reveals notable discrepancies in rule-learning behaviors between humans and LLMs. We believe our benchmark will serve as a valuable and challenging resource, and IDEA will provide crucial insights for the development of LLM agents capable of human-like rule learning in real-world scenarios. Our code and data have been released on GitHub: https://github.com/KaiyuHe998/RULEARN_IDEA.

Agentic Retrieval-Augmented Generation (RAG) systems enhance Large Language Models (LLMs) by enabling dynamic, multi-step reasoning and information retrieval. However, these systems often exhibit sub-optimal search behaviors like over-search (retrieving redundant information) and under-search (failing to retrieve necessary information), which hinder efficiency and reliability. This work formally defines and quantifies these behaviors, revealing their prevalence across multiple QA datasets and agentic RAG systems (e.g., one model could have avoided searching in 27.7% of its search steps). Furthermore, we demonstrate a crucial link between these inefficiencies and the models’ uncertainty regarding their own knowledge boundaries, where response accuracy correlates with model’s uncertainty in its search decisions. To address this, we propose β-GRPO, a reinforcement learning-based training method that incorporates confidence threshold to reward high-certainty search decisions. Experiments on seven QA benchmarks show that β-GRPO enable a 3B model with better agentic RAG ability, outperforming other strong baselines with a 4% higher average exact match score.

Large language model (LLM) agents have demonstrated remarkable potential in advancing scientific discovery. However, their capability in the fundamental yet crucial task of reproducing code from research papers, especially in the NLP domain, remains underexplored. This task includes unique complex reasoning challenges in the intellectual synthesis of abstract concepts and the comprehension of code repositories with interdependent files. Motivated by this gap, we present LMR-BENCH, a benchmark designed to systematically evaluate the capability of LLM agents on code reproduction from Language Modeling Research. It consists of 28 code reproduction tasks derived from 23 research papers published in top-tier NLP venues over the past five years, spanning nine fundamental categories. Models are provided with a research paper, a code repository containing one or more masked functions, and instructions for implementing these functions. We conduct extensive experiments in standard prompting and LLM agent settings with state-of-the-art LLMs, evaluating the accuracy of unit tests and performing LLM-based evaluation of code correctness. Experimental results reveal that even the most advanced models still exhibit persistent limitations in scientific reasoning and code synthesis, highlighting critical gaps in LLM agents’ ability to autonomously reproduce scientific research.