ICLR 2026 Papers

Large language models (LLMs) have demonstrated strong persuasive capabilities comparable to those of humans, offering promising benefits while raising societal concerns about their deployment. However, systematically evaluating persuasive capabilities is inherently challenging, as the effectiveness of persuasion among humans varies significantly across different domains. In this paper, we take a theory-driven approach to provide a scalable and principled framework to measure the persuasive capabilities of LLMs in strategic interactions. Grounded in the Bayesian Persuasion (BP) framework, we repurpose existing human–human persuasion datasets to construct environments for evaluating and training LLMs in strategic persuasion. Our results reveal that frontier models can consistently achieve high persuasion gains and exhibit sophisticated persuasion strategies that align with theoretical predictions. Building on this, we use reinforcement learning to train LLMs for strategic persuasion in our environments. Our results also demonstrate that even small LLMs can obtain significantly higher persuasion gains through reinforcement learning.

In long-tailed semi-supervised learning (LTSSL), pseudolabeling often creates a vicious cycle of bias amplification, a problem that recent state-of-the-art methods attempt to mitigate using logit adjustment (LA). However, their adjustment schemes, inherited from LA, remain inherently hierarchyagnostic, failing to account for the semantic relationships between classes. In this regard, we identify a critical yet overlooked problem of intra-super-class imbalance, where a toxic combination of high semantic similarity and severe local imbalance within each super-class hinders effective LTSSL. This problem causes the model to reinforce on its errors, leading to representation overshadowing. To break this cycle, we propose Super-Class-Aware Debiasing (SCAD), a new framework that performs a dynamic, super-class-aware logit adjustment. SCAD leverages the latent semantic structure between classes to focus its corrective power on the most confusable groups, effectively resolving the local imbalances. Our extensive experiments validate that SCAD achieves new state-of-the-art performance, demonstrating the necessity of a super-class-aware approach for robust debiasing.

Large Reasoning Models (LRMs) have demonstrated remarkable performance on complex reasoning tasks by employing test-time scaling. However, they often generate over-long chains-of-thought that, driven by substantial reflections such as repetitive self-questioning and circular reasoning, lead to high token consumption, substantial computational overhead, and increased latency without improving accuracy, particularly in smaller models. Our observation reveals that increasing problem complexity induces more excessive and unnecessary reflection, which in turn reduces accuracy and increases token overhead. To address this challenge, we propose Adaptive Reflection and Length Coordinated Penalty (ARLCP), a novel reinforcement learning framework designed to dynamically balance reasoning efficiency and solution accuracy. ARLCP introduces two key innovations: (1) a reflection penalty that adaptively curtails unnecessary reflective steps while preserving essential reasoning, and (2) a length penalty calibrated to the estimated complexity of the problem. By coordinating these penalties, ARLCP encourages the model to generate more concise and effective reasoning paths. We evaluate our method on five mathematical reasoning benchmarks using DeepSeek-R1-Distill-Qwen-1.5B and DeepSeek-R1-Distill-Qwen-7B models. Experimental results show that ARLCP achieves a superior efficiency-accuracy trade-off compared to existing approaches. For the 1.5B model, it reduces the average response length by 53.1% while simultaneously improving accuracy by 5.8%. For the 7B model, it achieves a 35.0% reduction in length with a 2.7% accuracy gain.

Remote photoplethysmography (rPPG) enables non-contact physiological measurement but remains highly susceptible to illumination changes, motion artifacts, and limited temporal modeling. Large Language Models (LLMs) excel at capturing long-range dependencies, offering a potential solution but struggle with the continuous, noise-sensitive nature of rPPG signals due to their text-centric design. To bridge this gap, we introduce PhysLLM, a collaborative optimization framework that synergizes LLMs with domain-specific rPPG components. Specifically, the Text Prototype Guidance (TPG) strategy is proposed to establish cross-modal alignment by projecting hemodynamic features into LLM-interpretable semantic space, effectively bridging the representational gap between physiological signals and linguistic tokens. Besides, a novel Dual-Domain Stationary (DDS) Algorithm is proposed for resolving signal instability through adaptive time-frequency domain feature re-weighting. Finally, rPPG task-specific cues systematically inject physiological priors through physiological statistics, environmental contextual answering, and task description, leveraging cross-modal learning to integrate both visual and textual information, enabling dynamic adaptation to challenging scenarios like variable illumination and subject movements. Evaluation on four benchmark datasets, PhysLLM achieves state-of-the-art accuracy and robustness, demonstrating superior generalization across lighting variations and motion scenarios.

Reasoning LLMs (RLLMs) generate step-by-step chains of thought (CoTs) before giving an answer, which improves performance on complex tasks and makes reasoning transparent. But how robust are these reasoning traces to disruptions that occur within them? To address this question, we introduce a controlled evaluation framework that perturbs a model’s own CoT at fixed timesteps. We design seven interventions (benign, neutral, and adversarial) and apply them to multiple open-weight RLLMs across MATH, SCIENCE, and LOGIC tasks. Our results show that RLLMs are generally robust, reliably recovering from diverse perturbations, with robustness improving with model size and degrading when interventions occur early. However, robustness is not style-invariant: paraphrasing suppresses doubt-like expressions and reduces performance, while other interventions trigger doubt and support recovery. Recovery also carries a cost: neutral and adversarial noise can inflate CoT length by more than 200%, whereas paraphrasing shortens traces but harms accuracy. These findings provide new evidence on how RLLMs maintain reasoning integrity, identify doubt as a central recovery mechanism, and highlight trade-offs between robustness and efficiency that future training methods should address.

In this paper, we explore how to empower general-purpose Vision-Language Models (VLMs) to control humanoid agents. General-purpose VLMs (e.g., GPT-4) exhibit strong open-world generalization, and remove the need for additional fine-tuning data. To build such an agent, two key components are required: (1) an embodied instruction compiler, which enables the VLM to observe the scene and translate high-level user instructions into low-level control parameters; and (2) a motion executor, which generates human motions from these parameters while adapting to real-time physical feedback. We present BiBo, a VLM-driven humanoid agent composed of an embodied instruction compiler and a diffusion-based motion executor. The compiler interprets user instructions in context with the environment, and leverages a chain of visual question answering (VQA) to guide the VLM in specifying control parameters (e.g., motion captions, locations). The diffusion executor extends future joint trajectories from prior motion, conditioned on both control parameters and environmental feedback. Experiments demonstrate that BiBo achieves an interaction task success rate of 90.2\% in open environments, and improves the precision of text-guided motion execution by 16.3\% over prior methods. BiBo handles not only basic interaction but also diverse motions, and even dancing while striking at a sandbag. The code will be released upon publication.

A well-known pitfall of molecular generative models is that they are not guaranteed to generate synthesizable molecules. Existing solutions for this problem often struggle to effectively navigate exponentially large combinatorial space of synthesizable molecules and suffer from poor coverage. To address this problem, we introduce ReaSyn, an iterative generative pathway refinement framework that obtains synthesizable analogs to input molecules by projecting them onto synthesizable space. Specifically, we propose a simple synthetic pathway representation that allows for generating pathways in both bottom-up and top-down traversal of synthetic trees. We design ReaSyn so that both bottom-up and top-down pathways can be sampled with a single unified autoregressive model. ReaSyn can thus iteratively refine subtrees of generated synthetic trees in a bidirectional manner. Further, we introduce a discrete flow model that refines the generated pathway at the entire pathway level with edit operations: insertion, deletion, and substitution. The iterative refinement cycle of (1) bottom-up decoding, (2) top-down decoding, and (3) holistic editing constitutes a powerful pathway reasoning strategy, allowing the model to explore the vast space of synthesizable molecules. Experimentally, ReaSyn achieves the highest reconstruction rate and pathway diversity in synthesizable molecule reconstruction and the highest optimization performance in synthesizable goal-directed molecular optimization, and significantly outperforms previous synthesizable projection methods in synthesizable hit expansion. These results highlight ReaSyn's superior ability to navigate combinatorially-large synthesizable chemical space.

Serving large language models (LLMs) at scale necessitates efficient key-value (KV) cache management. KV caches can be reused across conversation turns via shared-prefix prompts that are common in iterative code editing and chat. However, stale caches consume scarce GPU memory, require offloading, or force recomputation. We present KVTC, a lightweight transform coder that compresses KV caches for compact on-GPU and off-GPU storage. Drawing on classical media compression, KVTC combines PCA-based feature decorrelation, adaptive quantization, and entropy coding. It requires only a brief initial calibration and leaves model parameters unchanged. By exploiting redundancies in KV caches, KVTC achieves up to 20x compression, while maintaining reasoning and long-context accuracy in Llama 3.1, Mistral-NeMo, and R1-Qwen 2.5. Across benchmarks including AIME25, LiveCodeBench, GSM8K, MMLU, Qasper, RULER, and MATH500, KVTC consistently outperforms inference-time baselines such as token eviction, quantization, and SVD-based methods, delivering substantially higher compression ratios. These results support KVTC as a practical building block for memory-efficient LLM serving with reusable KV caches.

Modern Recurrent Neural Networks have become a competitive architecture for 3D reconstruction due to their linear-time complexity. However, their performance degrades significantly when applied beyond the training context length, revealing limited length generalization. In this work, we revisit the 3D reconstruction foundation models from a Test-Time Training perspective, framing their designs as an online learning problem. Building on this perspective, we leverage the alignment confidence between the memory state and incoming observations to derive a closed-form learning rate for memory updates, to balance between retaining historical information and adapting to new observations. This training-free intervention, termed TTT3R, substantially improves length generalization, achieving a 2 $\times$ improvement in global pose estimation over baselines, while operating at 20 FPS with just 6 GB of GPU memory to process thousands of images. Code is available in https://rover-xingyu.github.io/TTT3R.

In this work, we introduce dual goal representations for goal-conditioned reinforcement learning (GCRL). A dual goal representation characterizes a state by "the set of temporal distances from all other states"; in other words, it encodes a state through its relations to every other state, measured by temporal distance. This representation provides several appealing theoretical properties. First, it depends only on the intrinsic dynamics of the environment and is invariant to the original state representation. Second, it contains provably sufficient information to recover an optimal goal-reaching policy, while being able to filter out exogenous noise. Based on this concept, we develop a practical goal representation learning method that can be combined with any existing GCRL algorithm. Through diverse experiments on the OGBench task suite, we empirically show that dual goal representations consistently improve offline goal-reaching performance across 20 state- and pixel-based tasks.

Autoregressive decoding is bottlenecked by its *sequential* nature. Speculative decoding has become a standard way to accelerate inference by using a fast *draft model* to predict upcoming tokens from a slower *target model*, and then verifying them *in parallel* with a single target model forward pass. However, speculative decoding itself relies on a *sequential* dependence between speculation and verification. We introduce *speculative speculative decoding* (SSD) to parallelize these operations. While a verification is ongoing, the draft model *predicts* likely verification outcomes and prepares speculations pre-emptively for them. If the actual verification outcome is then in the predicted set, a speculation can be returned immediately, thereby eliminating all speculation overhead. We identify three key challenges presented by speculative speculative decoding, and put forth principled methods to solve each after theoretical analysis. The result is **Saguaro**, an optimized SSD algorithm which is up to twice as fast as optimized speculative decoding baselines and up to 5× faster than autoregressive decoding with open source inference engines. **Saguaro** can be combined with existing methods like EAGLE and token tree speculation for further gains, and permits scaling draft compute to better predict verification outcomes, introducing new tradeoffs between compute and latency.

While the "deep reasoning" paradigm has spurred significant advances in verifiable domains like mathematics, its application to open-ended, creative generation remains a critical challenge. The two dominant methods for instilling reasoning—reinforcement learning (RL) and instruction distillation -- falter in this area; RL struggles with the absence of clear reward signals and high-quality reward models, while distillation is prohibitively expensive and capped by the teacher model's capabilities. To overcome these limitations, we introduce REverse-Engineered Reasoning (REER), a new paradigm that fundamentally shifts the approach. Instead of building a reasoning process "forwards" through trial-and-error or imitation, REER works "backwards" from known good solutions to computationally discover the latent, step-by-step deep reasoning process that could have produced them. Using this scalable, gradient-free approach, we curate and open-source DeepWriting-20K, a large-scale dataset of 20,000 deep reasoning trajectories for open-ended tasks. Our model, DeepWriter-8B, trained on this data, not only surpasses strong open-source baselines but also achieves performance competitive with, and at times superior to, leading proprietary models like GPT-4o and Claude 3.5.

Reliably composing Large Language Models (LLMs) for complex, multi-step workflows remains a significant challenge. The dominant paradigm---optimizing discrete prompts in a pipeline---is notoriously brittle and struggles to enforce the formal compliance required for structured tasks. We introduce Type-Compliant Adaptation Cascades (TACs), a framework that recasts workflow adaptation as learning typed probabilistic programs. TACs treat the entire workflow, which is composed of parameter-efficiently adapted LLMs and deterministic logic, as an unnormalized joint distribution. This enables principled, gradient-based training even with latent intermediate structures. We provide theoretical justification for our tractable optimization objective, proving that the optimization bias vanishes as the model learns type compliance. Empirically, TACs significantly outperform state-of-the-art prompt-optimization baselines. Gains are particularly pronounced on structured tasks, improving FinQA from $12.0\%$ to $24.7\%$ for a Qwen 3 8B model, MGSM-SymPy from $57.1\%$ to $75.9\%$ for a Gemma 2 27B model, MGSM from $1.6\%$ to $27.3\%$, and MuSR from $36.5\%$ to $62.6\%$ for a Gemma 7B model. TACs offer a robust and theoretically grounded paradigm for developing reliable, task-compliant LLM systems.

In this paper, we identify that the key criterion for determining whether a model is truly capable of novel view synthesis (NVS) is transferability: Whether any pose representation extracted from one video sequence can be used to re-render the same camera trajectory in another. We analyze prior work on self-supervised NVS and find that their predicted poses do not transfer: The same set of poses lead to different camera trajectories in different 3D scenes. Here, we present XFactor, the first geometry-free self-supervised model capable of true NVS. XFactor combines pair-wise pose estimation with a simple augmentation scheme of the inputs and outputs that jointly enables disentangling camera pose from scene content and facilitates geometric reasoning. Remarkably, we show that XFactor achieves transferability with unconstrained latent pose variables, without any 3D inductive biases or concepts from multi-view geometry — such as an explicit parameterization of poses as elements of SE(3). We introduce a new metric to quantify transferability, and through large-scale experiments, we demonstrate that XFactor significantly outperforms prior pose-free NVS transformers, and show that latent poses are highly correlated with real-world poses through probing experiments.

Reasoning abilities of large language models (LLMs) require explicit derivations compared to general question-answering, supervised fine-tuning (SFT) can empower multiple reasoning abilities in LLMs via learning from various datasets. However, neither training the datasets jointly (mix-up) nor continually can maintain the performance of single-dataset SFT, sometimes better while sometimes even worse, illustrating vanilla SFT can not only facilitate reasoning abilities but also introduce conflicts. In this paper, we propose a novel framework to mitigate the conflicts and preserve benefits among different reasoning tasks, and even surpass each task's single dataset SFT performance. We start by exploring the differences between reasoning fine-tuned and base LLMs by analyzing their parameter variations during model inference, and we discover that each reasoning capability has exclusive parameters that benefit it more evidently than others. In contrast, the overlapped parameters of tasks can bring benefits or conflicts. Inspired by the findings, we propose to update the exclusive and overlapped parameters according to specific reasoning task combinations differentially, thereby avoiding unnecessary conflicts while maintaining benefits. Consistent improvements in mix-up and continual SFT experiments demonstrate that the proposed SFT strategy can achieve better performance on various LLMs (Llama3-8B, Mistral-7B, and Qwen2.5-14B) and diverse reasoning tasks with fewer conflicts, showing the superiority and generality of our analysis findings and the proposed approach.

While large language models (LLMs) are increasingly used as automated heuristic designers for vehicle routing problems (VRPs), current state-of-the-art methods predominantly rely on prompting massive, general-purpose models like GPT-4. This work challenges that paradigm by demonstrating that a smaller, specialized LLM, when meticulously fine-tuned, can generate components that surpass expert-crafted heuristics within advanced solvers. We propose RFTHGS, a novel Reinforcement learning (RL) framework for Fine-Tuning a compact LLM to generate high-performance crossover operators for the Hybrid Genetic Search (HGS) solver, applied to the Capacitated VRP (CVRP). Our method employs a multi-tiered, curriculum-based reward function that progressively guides the LLM to master generating first compilable, then executable, and finally, superior-performing operators that exceed human expert designs. This is coupled with an operator caching mechanism that discourages plagiarism and promotes diversity during training. Comprehensive experiments show that our fine-tuned LLM produces crossover operators which significantly outperform the expert-designed ones in HGS. The performance advantage remains consistent, generalizing from small-scale instances to large-scale problems with up to 1000 nodes. Furthermore, RFTHGS exceeds the performance of leading neuro-combinatorial baselines, prompt-based methods, and commercial LLMs such as GPT-4o and GPT-4o-mini.

Test-time scaling offers a promising path to improve LLM reasoning by utilizing more compute at inference time; however, the true promise of this paradigm lies in extrapolation (i.e., improvement in performance on hard problems as LLMs keep "thinking" for longer, beyond the maximum token budget they were trained on). Surprisingly, we find that most existing reasoning models do not extrapolate well. We show that one way to enable extrapolation is by training the LLM to perform in-context exploration: training the LLM to effectively spend its test time budget by chaining operations (such as generation, verification, refinement, etc.), or testing multiple hypotheses before it commits to an answer. To enable in-context exploration, we identify three key ingredients as part of our recipe e3: (1) chaining skills that the base LLM has asymmetric competence in, e.g., chaining verification (easy) with generation (hard), as a way to implement in-context search; (2) leveraging "negative" gradients from incorrect traces to amplify exploration during RL, resulting in longer search traces that chains additional asymmetries; and (3) coupling task difficulty with training token budget during training via a specifically-designed curriculum to structure in-context exploration. Our recipe e3 produces the best known 1.7B model according to AIME'25 and HMMT'25 scores, and extrapolates to 2x the training token budget. Our e3-1.7B model not only attains high pass@1 scores, but also improves pass@k over the base model.

Fine-tuning pretrained models is a standard and effective workflow in modern machine learning. However, robust fine-tuning (RFT), which aims to simultaneously achieve adaptation to a downstream task and robustness to adversarial examples, remains challenging. Despite the abundance of non-robust pretrained models in open-source repositories, their potential for RFT is less understood. We address this knowledge gap by systematically examining RFT from such non-robust models. Our experiments reveal that fine-tuning non-robust models with a robust objective, even under small perturbations, can lead to poor performance, a phenomenon that we dub _suboptimal transfer_. In challenging scenarios (eg, difficult tasks, high perturbation), the resulting performance can be so low that it may be considered a transfer failure. We find that fine-tuning using a robust objective impedes task adaptation at the beginning of training and eventually prevents optimal transfer. However, we propose a novel heuristic, _Epsilon-Scheduling_, a schedule over perturbation strength used during training that promotes optimal transfer. Additionally, we introduce _expected robustness_, a metric that captures performance across a range of perturbations, providing a more comprehensive evaluation of the accuracy-robustness trade-off of diverse models at test-time. Extensive experiments on wide range of configurations (six pretrained models and five datasets) show that _Epsilon-Scheduling_ successfully prevents _suboptimal transfer_ and consistently improves expected robustness.

Large language models are trained on massive corpora of web data, which may include private data, copyrighted material, factually inaccurate data, or data that degrades model performance. Eliminating the influence of such problematic datapoints on a model through complete retraining---by repeatedly pretraining the model on datasets that exclude these specific instances---is computationally prohibitive. To address this, unlearning algorithms have been proposed, that aim to eliminate the influence of particular datapoints at a low computational cost, while leaving the rest of the model intact. However, precisely unlearning the influence of data on a large language model has proven to be a major challenge. In this work, we propose a new algorithm, MSA (**M**odel **S**tate **A**rithmetic), for unlearning datapoints in large language models. MSA utilizes prior model checkpoints--- artifacts that record model states at different stages of pretraining--- to estimate and counteract the effect of targeted datapoints. Our experimental results show that MSA achieves competitive performance and often outperforms existing machine unlearning algorithms across multiple benchmarks, models, and evaluation metrics, suggesting that MSA could be an effective approach towards more flexible large language models that are capable of data erasure.