Theory of Mind (ToM)-an understanding of the mental states of others-is a key aspect of human social intelligence, yet, chatbots and LLM-based social agents do not typically integrate it. In this work, we demonstrate that LLMs that explicitly use ToM get better at dialogue, achieving goals more effectively. After showing that simply prompting models to generate mental states between dialogue turns already provides significant benefit, we further introduce ToMAgent (ToMA), a ToM-focused dialogue agent. ToMA is trained by pairing ToM with dialogue lookahead to produce mental states that are maximally useful for achieving dialogue goals. Experiments on the Sotopia interactive social evaluation benchmark demonstrate the effectiveness of our method over a range of baselines. Comprehensive analysis shows that ToMA exhibits more strategic, goal-oriented reasoning behaviors, which enable long-horizon adaptation, while maintaining better relationships with their partners. Our results suggest a step forward in integrating ToM for building socially intelligent LLM agents.

As Large Language Models (LLMs) are increasingly embedded in real-world decision-making processes, it becomes crucial to examine the extent to which they exhibit cognitive biases. Extensively studied in the field of psychology, cognitive biases appear as systematic distortions commonly observed in human judgments. This paper presents a large-scale evaluation of eight well-established cognitive biases across 45 LLMs, analyzing over 2.8 million LLM responses generated through controlled prompt variations. To achieve this, we introduce a novel evaluation framework based on multiple-choice tasks, hand-curate a dataset of 220 decision scenarios targeting fundamental cognitive biases in collaboration with psychologists, and propose a scalable approach for generating diverse prompts from human-authored scenario templates. Our analysis shows that LLMs exhibit bias-consistent behavior in 17.8-57.3% of instances across a range of judgment and decision-making contexts targeting anchoring, availability, confirmation, framing, interpretation, overattribution, prospect theory, and representativeness biases. We find that both model size and prompt specificity play a significant role on bias susceptibility as follows: larger size (>32B parameters) can reduce bias in 39.5% of cases, while higher prompt detail reduces most biases by up to 14.9%, except in one case (Overattribution), which is exacerbated by up to 8.8%.

Implicit in-context learning (ICL) has newly emerged as a promising paradigm that simulates ICL behaviors in the representation space of Large Language Models (LLMs), aiming to attain few-shot performance at zero-shot cost. However, existing approaches largely rely on injecting shift vectors into residual flows, which are typically constructed from labeled demonstrations or task-specific alignment. Such designs fall short of utilizing the structural mechanisms underlying ICL and suffer from limited generalizability. To address this, we propose In-Context Routing (ICR), a novel implicit ICL method that internalizes generalizable ICL patterns at the attention logits level. It extracts reusable structural directions that emerge during ICL and employs a learnable input-conditioned router to modulate attention logits accordingly, enabling a train-once-and-reuse framework. We evaluate ICR on 12 real-world datasets spanning diverse domains and multiple LLMs. The results show that ICR consistently outperforms prior implicit ICL methods that require task-specific retrieval or training, while demonstrating robust generalization to out-of-domain tasks where existing methods struggle. These findings position ICR to push the boundary of ICL's practical value. Large Language Models (LLMs) have been widely adopted for text understanding and generation tasks. As applications broaden, the ability to adapt these models efficiently at inference time has become increasingly important (Brown et al., 2020; Wang et al., 2020b). In-context learning (ICL) is a central mechanism for this adaptation (Dong et al., 2022; Min et al., 2021): by conditioning on a few labeled examples inserted before the query, known as in-context demonstrations (ICDs), the model can perform new tasks without any parameter updates (Wies et al., 2023; Pan, 2023). Despite its broad adoption, ICL faces two practical limitations: (i) inserting ICDs into the prompt inflates sequence length and inference cost compared to zero-shot use (Peng et al., 2024; Li et al., 2025a), and (ii) performance is brittle, varying with small changes in ICD order or format (Wu et al., 2022; Guo et al., 2024). To address these issues, recent work has explored implicit ICL, which converts ICDs into dense vectors that steer intermediate residual flows to approximate the effect of explicit prompting (Hendel et al., 2023; Todd et al., 2023; Liu et al., 2023; Li et al., 2024). While vector-based implicit ICL offers a new way to simulate ICL behaviors in LLMs, it struggles to generalize across real-world tasks.

Proteomics data is essential to pathogenic understanding of a disease phenotype. In cancer, analysis of molecular signatures enables precision medicine through the identification of biological processes that drive individualized tumor progression, therapeutic resistance, and clinical heterogeneity. Recent advances in multimodal large language models (LLMs) have shown remarkable capacity to integrate and reason across heterogeneous data modalities. However, performing multi-modal language modeling for molecular understanding of patient-specific proteomics remains a significant challenge due to two barriers: (1) the lack of instruction-tuning datasets that enable clinical interpretation from proteomics data, and (2) the absence of language modeling architectures designed to capture the rich heterogeneity of molecular data. In this work, we introduce CPTAC-PROTSTRUCT, the first instruction tuning dataset for molecular understanding of oncology, comprising over 400k open-ended examples derived from individualized proteomic profiles curated from the largest national proteomics cancer study (CPTAC). Additionally, we propose KRONOS (Knowledge Representation of patient Omics Networks in Oncology via Structured tuning), a novel graph-LLM framework that leverages molecular interaction topology with proteomics to learn patient-specific graph representations for enhanced clinical reasoning. We show that KRONOS achieves competitive performance across benchmark clinical tasks, including molecular classification, temporal trajectory modeling, and tumor stage prediction from proteomics data. Ultimately, this approach empowers LLMs to understand patient-level pathogenesis, advancing precision medicine through more accurate diagnosis, prognosis, and treatment stratification.

Abstract--Intent-Based Networking (IBN) aims to simplify network management by enabling users to specify high-level goals that drive automated network design and configuration. However, translating informal natural-language intents into formally correct optical network topologies remains challenging due to inherent ambiguity and lack of rigor in Large Language Models (LLMs). T o address this, we propose a novel hybrid pipeline that integrates LLM-based intent parsing, formal methods, and Optical Retrieval-Augmented Generation (RAG). By enriching design decisions with domain-specific optical standards and systematically incorporating symbolic reasoning and verification techniques, our pipeline generates explainable, verifiable, and trustworthy optical network designs. Intent-Based Networking (IBN) simplifies network management by allowing users to express high-level objectives--such as connectivity, performance, or security--without specifying implementation details [1], [2]. Standardization bodies like TM Forum and the Internet Engineering Task Force define intent as a declarative statement of desired outcomes, delegating the detailed configuration and implementation tasks to automated systems. By abstracting away low-level complexities, IBN significantly reduces operational overhead, human error, and management complexity [2]. Existing research predominantly explores intent translation into configurations or incremental topology adjustments [3], [4], but largely overlooks the initial phase of comprehensive network design, particularly for optical networks. Poor initial design decisions can lead to significant performance degradation or expensive reconfigurations throughout the operational lifecycle [5], [6].

Reinforcement Learning (RL) has emerged as a popular training paradigm, particularly when paired with reasoning models. While effective, it primarily focuses on generating responses and lacks mechanisms to explicitly foster critique or reflection. Several recent studies, like Critique-Fine-Tuning (CFT) and Critique-Guided-Distillation (CGD) have shown the benefits of explicitly teaching LLMs how to critique. Motivated by them, we propose Critique Reinforcement Learning (CRL), where the model is tasked with generating a critique for a given (question, solution) pair. The reward is determined solely by whether the final judgment label $c \in \{\texttt{True}, \texttt{False}\}$ of the generated critique aligns with the ground-truth judgment $c^*$. Building on this point, we introduce \textsc{Critique-Coder}, which is trained on a hybrid of RL and CRL by substituting 20\% of the standard RL data with CRL data. We fine-tune multiple models (\textsc{Critique-Coder}) and evaluate them on different benchmarks to show their advantages over RL-only models. We show that \textsc{Critique-Coder} consistently outperforms RL-only baselines on all the evaluated benchmarks. Notably, our \textsc{Critique-Coder-8B} can reach over 60\% on LiveCodeBench (v5), outperforming other reasoning models like DeepCoder-14B and GPT-o1. Beyond code generation, \textsc{Critique-Coder} also demonstrates enhanced general reasoning abilities, as evidenced by its better performance on logic reasoning tasks from the BBEH dataset. This indicates that the application of CRL on coding datasets enhances general reasoning and critique abilities, which are transferable across a broad range of tasks. Hence, we believe that CRL works as a great complement to standard RL for LLM reasoning.

Large Language Models (LLMs) demonstrate impressive mathematical reasoning abilities, but their solutions frequently contain errors that cannot be automatically verified. Formal theorem proving systems such as Lean 4 offer automated verification with complete accuracy, motivating recent efforts to build specialized prover LLMs that generate verifiable proofs in formal languages. However, a significant gap remains: current prover LLMs solve substantially fewer problems than general-purpose LLMs operating in natural language. We introduce Hilbert, an agentic framework that bridges this gap by combining the complementary strengths of informal reasoning and formal verification. Our system orchestrates four components: an informal LLM that excels at mathematical reasoning, a specialized prover LLM optimized for Lean 4 tactics, a formal verifier, and a semantic theorem retriever. Given a problem that the prover is unable to solve, Hilbert employs recursive decomposition to split the problem into subgoals that it solves with the prover or reasoner LLM. It leverages verifier feedback to refine incorrect proofs as necessary. Experimental results demonstrate that Hilbert substantially outperforms existing approaches on key benchmarks, achieving 99.2% on miniF2F, 6.6% points above the best publicly available method. Hilbert achieves the best known result on PutnamBench. It solves 462/660 problems (70.0%), outperforming proprietary approaches like SeedProver (50.4%) and achieving a 422% improvement over the best publicly available baseline. Thus, Hilbert effectively narrows the gap between informal reasoning and formal proof generation.

Recent advances in text-to-video generation have produced increasingly realistic and diverse content, yet evaluating such videos remains a fundamental challenge due to their multi-faceted nature encompassing visual quality, semantic alignment, and physical consistency. Existing evaluators and reward models are limited to single opaque scores, lack interpretability, or provide only coarse analysis, making them insufficient for capturing the comprehensive nature of video quality assessment. We present VideoScore2, a multi-dimensional, interpretable, and human-aligned framework that explicitly evaluates visual quality, text-to-video alignment, and physical/common-sense consistency while producing detailed chain-of-thought rationales. Our model is trained on a large-scale dataset VideoFeedback2 containing 27,168 human-annotated videos with both scores and reasoning traces across three dimensions, using a two-stage pipeline of supervised fine-tuning followed by reinforcement learning with Group Relative Policy Optimization (GRPO) to enhance analytical robustness. Extensive experiments demonstrate that VideoScore2 achieves superior performance with 44.35 (+5.94) accuracy on our in-domain benchmark VideoScore-Bench-v2 and 50.37 (+4.32) average performance across four out-of-domain benchmarks (VideoGenReward-Bench, VideoPhy2, etc), while providing interpretable assessments that bridge the gap between evaluation and controllable generation through effective reward modeling for Best-of-N sampling. Project Page: https://tiger-ai-lab.github.io/VideoScore2/

Open-source software projects are foundational to modern software ecosystems, with the Linux kernel standing out as a critical exemplar due to its ubiquity and complexity. Although security patches are continuously integrated into the Linux mainline kernel, downstream maintainers often delay their adoption, creating windows of vulnerability. A key reason for this lag is the difficulty in identifying security-critical patches, particularly those addressing exploitable vulnerabilities such as out-of-bounds (OOB) accesses and use-after-free (UAF) bugs. This challenge is exacerbated by intentionally silent bug fixes, incomplete or missing CVE assignments, delays in CVE issuance, and recent changes to the CVE assignment criteria for the Linux kernel. While fine-grained patch classification approaches exist, they exhibit limitations in both coverage and accuracy. In this work, we identify previously unexplored opportunities to significantly improve fine-grained patch classification. Specifically, by leveraging cues from commit titles/messages and diffs alongside appropriate code context, we develop DUALLM, a dual-method pipeline that integrates two approaches based on a Large Language Model (LLM) and a fine-tuned small language model. DUALLM achieves 87.4% accuracy and an F1-score of 0.875, significantly outperforming prior solutions. Notably, DUALLM successfully identified 111 of 5,140 recent Linux kernel patches as addressing OOB or UAF vulnerabilities, with 90 true positives confirmed by manual verification (many do not have clear indications in patch descriptions). Moreover, we constructed proof-of-concepts for two identified bugs (one UAF and one OOB), including one developed to conduct a previously unknown control-flow hijack as further evidence of the correctness of the classification.

Large language models (LLMs) can adapt to new tasks via in-context learning (ICL) without parameter updates, making them powerful learning engines for fast adaptation. While extensive research has examined ICL as a few-shot learner, whether it can achieve long-term retention and cross-task knowledge accumulation when multitasks arrive sequentially remains underexplored. Motivated by human memory studies, we investigate the retention characteristics of ICL in multitask settings and extend it to in-context continual learning (ICCL), where continual learning ability emerges through task scheduling and prompt rearrangement. Experiments on Markov-Chain benchmarks demonstrate that, for specific large-language models, ICCL benefits from distributed practice (DP) in a manner analogous to humans, consistently revealing a spacing "sweet spot" for retention. Beyond retention performance, we propose a human-retention similarity metric to quantify how closely a continual-learning (CL) method aligns with human retention dynamics. Using this metric, we show that linear-attention models such as MAMBA and RWKV exhibit particularly human-like retention patterns, despite their retention performance lagging behind that of Transformer-based LLMs. Overall, our results establish ICCL as both cognitively plausible and practically effective, providing an inference-only CL paradigm that mitigates catastrophic forgetting and addresses the stability-plasticity dilemma in conventional CL methods.