Recent reinforcement learning (RL) methods have substantially enhanced the planning capabilities of Large Language Models (LLMs), yet the theoretical basis for their effectiveness remains elusive. In this work, we investigate RL's benefits and limitations through a tractable graph-based abstraction, focusing on policy gradient (PG) and Q-learning methods. Our theoretical analyses reveal that supervised fine-tuning (SFT) may introduce co-occurrence-based spurious solutions, whereas RL achieves correct planning primarily through exploration, underscoring exploration's role in enabling better generalization. However, we also show that PG suffers from diversity collapse, where output diversity decreases during training and persists even after perfect accuracy is attained. By contrast, Q-learning provides two key advantages: off-policy learning and diversity preservation at convergence. We further demonstrate that careful reward design is necessary to prevent reward hacking in Q-learning. Finally, applying our framework to the real-world planning benchmark Blocksworld, we confirm that these behaviors manifest in practice. Planning is a fundamental cognitive construct that underpins human intelligence, shaping our ability to organize tasks, coordinate activities, and formulate complex solutions such as mathematical proofs. It enables humans to decompose complex goals into manageable steps, anticipate potential challenges, and maintain coherence during problem solving. Similarly, planning plays a pivotal role in state-of-the-art Large Language Models (LLMs), enhancing their ability to address structured and long-horizon tasks with greater accuracy and reliability. Early generations of LLMs primarily relied on next-token prediction and passive statistical learning, which limited their planning capabilities to short-horizon, reactive responses.

Traditional video-based tasks like soccer action spotting rely heavily on visual inputs, often requiring complex and computationally expensive models to process dense video data. We propose a shift from this video-centric approach to a text-based task, making it lightweight and scalable by utilizing Large Language Models (LLMs) instead of Vision-Language Models (VLMs). We posit that expert commentary, which provides rich descriptions and contextual cues contains sufficient information to reliably spot key actions in a match. To demonstrate this, we employ a system of three LLMs acting as judges specializing in outcome, excitement, and tactics for spotting actions in soccer matches. Our experiments show that this language-centric approach performs effectively in detecting critical match events coming close to state-of-the-art video-based spotters while using zero video processing compute and similar amount of time to process the entire match.

Transformer-based architectures traditionally employ softmax to compute attention weights, which produces dense distributions over all tokens in a sequence. While effective in many settings, this density has been shown to be detrimental for tasks that demand precise focus on fixed-size patterns: as sequence length increases, non-informative tokens accumulate attention probability mass, leading to dispersion and representational collapse. We show in this paper that dynamically sparse attention mechanisms using $α$-entmax can avoid these issues, due to their ability to assign exact zeros to irrelevant tokens. Furthermore, we introduce Adaptive-Scalable Entmax (ASEntmax), which endows $α$-entmax with a learnable temperature parameter, allowing the attention distribution to interpolate between sparse (pattern-focused) and dense (softmax-like) regimes. Our empirical evaluation on synthetic tasks and language modeling demonstrates that ASEntmax substantially outperforms softmax, scalable softmax, and fixed-temperature $α$-entmax baselines, achieving up to 1000$\times$ length extrapolation on synthetic benchmarks and superior long-context generalization on language modeling while preserving short-context performance, including better perplexity trends and higher retrieval accuracies at 8$\times$ training length.

Transformer architectures, and their attention mechanisms in particular, form the foundation of modern large language models. While transformer models are widely believed to operate in high-dimensional hidden spaces, we show that attention outputs are confined to a surprisingly low-dimensional subspace, where about 60\% of the directions account for 99\% of the variance--a phenomenon that is consistently observed across diverse model families and datasets, and is induced by the attention output projection matrix. Critically, we find this low-rank structure as a key factor of the prevalent dead feature problem in sparse dictionary learning, where it creates a mismatch between randomly initialized features and the intrinsic geometry of the activation space. Building on this insight, we propose a subspace-constrained training method for sparse autoencoders (SAEs), initializing feature directions into the active subspace of activations. Our approach reduces dead features from 87\% to below 1\% in Attention Output SAEs with 1M features, and can further extend to other sparse dictionary learning methods. Our findings provide both new insights into the geometry of attention and practical tools for improving sparse dictionary learning in large language models.

Large language models (LLMs) are increasingly developed as autonomous agents using reinforcement learning (agentic RL) that reason and act in interactive environments. However, sparse and sometimes unverifiable rewards make it extremely challenging to assign credit when training LLM agents that serve as a policy. Recent work attempts to integrate process supervision into RL but suffers from biased annotation, reward hacking, high-variance from overly fine-grained rewards or failtures when state overlap is rare. We therefore introduce implicit step rewards for agentic RL (iStar), a general credit-assignment strategy that integrates seamlessly with standard RL algorithms without relying on additional rollouts or explicit step labels. Particularly, we alternatively optimize an implicit process reward model (PRM) with the policy model to generate implicit step rewards via a trajectory-based DPO objective. Theoretical analysis shows that this learning objective produces a step-wise reward function. Then the implicit step rewards are used to compute step-level advantages, which are combined with trajectory (or episode)-level advantages for policy updates, creating a self-reinforcing training loop. We evaluate our method on three challenging agent benchmarks, including WebShop and VisualSokoban, as well as open-ended social interactions with unverifiable rewards in SOTOPIA. Crucially, iStar shows superior performance over frontier LLMs and strong RL baselines across domains, achieving state-of-the-art results with higher sample-efficiency and training stability. Further analysis also demonstrates efficient exploration by iStar with increased rewards in both step- and episode-level while maintaining fewer steps to achieve task success. Code will be available soon.

Reinforcement Learning (RL) is increasingly utilized to enhance the reasoning capabilities of Large Language Models (LLMs). However, effectively scaling these RL methods presents significant challenges, primarily due to the difficulty in maintaining high AI accelerator utilization without generating stale, off-policy data that harms common RL algorithms. This paper introduces PipelineRL, an approach designed to achieve a superior trade-off between hardware efficiency and data on-policyness for LLM training. PipelineRL employs concurrent asynchronous data generation and model training, distinguished by the novel in-flight weight updates. This mechanism allows the LLM generation engine to receive updated model weights with minimal interruption during the generation of token sequences, thereby maximizing both the accelerator utilization and the freshness of training data. Experiments conducted on long-form reasoning tasks using 128 H100 GPUs demonstrate that PipelineRL achieves approximately $\sim 2x$ faster learning compared to conventional RL baselines while maintaining highly on-policy training data. A scalable and modular open-source implementation of PipelineRL is also released as a key contribution.

Detecting AI-generated text is an increasing necessity to combat misuse of LLMs in education, business compliance, journalism, and social media, where synthetic fluency can mask misinformation or deception. While prior detectors often rely on token-level likelihoods or opaque black-box classifiers, these approaches struggle against high-quality generations and offer little interpretability. In this work, we propose DivEye, a novel detection framework that captures how unpredictability fluctuates across a text using surprisal-based features. Motivated by the observation that human-authored text exhibits richer variability in lexical and structural unpredictability than LLM outputs, DivEye captures this signal through a set of interpretable statistical features. Our method outperforms existing zero-shot detectors by up to 33.2% and achieves competitive performance with fine-tuned baselines across multiple benchmarks. DivEye is robust to paraphrasing and adversarial attacks, generalizes well across domains and models, and improves the performance of existing detectors by up to 18.7% when used as an auxiliary signal. Beyond detection, DivEye provides interpretable insights into why a text is flagged, pointing to rhythmic unpredictability as a powerful and underexplored signal for LLM detection.

Training Large Language Models(LLMs) is one of the most compute-intensive tasks in high-performance computing. Predicting end-to-end training time for multi-billion parameter models distributed across hundreds of GPUs remains challenging due to complex interactions between transformer components, parallelism strategies(data, model, pipeline, tensor), and multi-tier communication. Learned models require costly sampling, while analytical models often struggle with real-world network and hardware complexities. We address this by decomposing LLMs into core computational primitives and modeling them with: (1) operator-level decomposition for fine-grained analysis; (2) lightweight sampling based hardware-aware prediction models for key operations; (3) an end-to-end prediction system integrating these components across complex parallelization strategies. Crucially, our methodology has been validated on two large-scale HPC systems. Our framework achieves low average prediction errors-4.98\% on Perlmutter(A100) and 9.38\% on Vista(GH200)-for models up to 20B parameters across 128 GPUs. Importantly, it runs entirely on CPUs, enabling rapid iteration over hardware configurations and training strategies without costly on-cluster experimentation.

Research on Large Language Models (LLMs) increasingly focuses on identifying mechanistic explanations for their behaviors, yet the field lacks clear principles for determining when (and how) findings from one model instance generalize to another. This paper addresses a fundamental epistemological challenge: given a mechanistic claim about a particular model, what justifies extrapolating this finding to other LLMs -- and along which dimensions might such generalizations hold? I propose five potential axes of correspondence along which mechanistic claims might generalize, including: functional (whether they satisfy the same functional criteria), developmental (whether they develop at similar points during pretraining), positional (whether they occupy similar absolute or relative positions), relational (whether they interact with other model components in similar ways), and configurational (whether they correspond to particular regions or structures in weight-space). To empirically validate this framework, I analyze "1-back attention heads" (components attending to previous tokens) across pretraining in random seeds of the Pythia models (14M, 70M, 160M, 410M). The results reveal striking consistency in the developmental trajectories of 1-back attention across models, while positional consistency is more limited. Moreover, seeds of larger models systematically show earlier onsets, steeper slopes, and higher peaks of 1-back attention. I also address possible objections to the arguments and proposals outlined here. Finally, I conclude by arguing that progress on the generalizability of mechanistic interpretability research will consist in mapping constitutive design properties of LLMs to their emergent behaviors and mechanisms.

The growing deployment of large language model (LLM) based agents that interact with external environments has created new attack surfaces for adversarial manipulation. One major threat is indirect prompt injection, where attackers embed malicious instructions in external environment output, causing agents to interpret and execute them as if they were legitimate prompts. While previous research has focused primarily on plain-text injection attacks, we find a significant yet underexplored vulnerability: LLMs' dependence on structured chat templates and their susceptibility to contextual manipulation through persuasive multi-turn dialogues. To this end, we introduce ChatInject, an attack that formats malicious payloads to mimic native chat templates, thereby exploiting the model's inherent instruction-following tendencies. Building on this foundation, we develop a persuasion-driven Multi-turn variant that primes the agent across conversational turns to accept and execute otherwise suspicious actions. Through comprehensive experiments across frontier LLMs, we demonstrate three critical findings: (1) ChatInject achieves significantly higher average attack success rates than traditional prompt injection methods, improving from 5.18% to 32.05% on AgentDojo and from 15.13% to 45.90% on InjecAgent, with multi-turn dialogues showing particularly strong performance at average 52.33% success rate on InjecAgent, (2) chat-template-based payloads demonstrate strong transferability across models and remain effective even against closed-source LLMs, despite their unknown template structures, and (3) existing prompt-based defenses are largely ineffective against this attack approach, especially against Multi-turn variants. These findings highlight vulnerabilities in current agent systems.