Compact vision-language models (VLMs) now power a growing share of multimodal applications. The benchmarks used to compare them, however, inherit a frontier-centric design: each model is reduced to a single accuracy number, narrowing the inter-model gap on saturated suites and pressing models into low-score bands on harder ones. We introduce PRISM-VLM, a multi-axis discriminative benchmark that scores every item along seven axes covering the recurring failure modes (task quality, behavioral robustness, and capability bottlenecks) and combines them into a single PSCORE, with items recycled from fifteen public benchmarks. Across compact VLMs from the past two years, PSCORE separates model pairs more reliably than prior single-axis benchmarks under an item-level paired bootstrap, and surfaces behavioral differences these benchmarks average away. Even models with statistically indistinguishable PSCOREs diverge sharply along the per-axis profile, particularly on sycophancy, which is nearly orthogonal to single-prompt accuracy. We will release the full pipeline, prompts, and per-item annotations.

Math reasoning benchmarks have prolifer- ated, yet most lack a per-item difficulty sig- nal grounded in actual human performance. We introduce KCSAT-ML, a decade (2014– 2025) of Korean College Scholastic Ability Test (KCSAT; Suneung) mathematics: 664 problems with a 339-item core set carrying official per-item error rates from nationwide cohorts of hundreds of thousands of examinees. We pair the benchmark with Difficulty-aligned Reasoning Gain (DRG): a score-orthogonal metric that asks whether a model’s mistakes concentrate on the items humans found hard, or on items humans found easy. Together they expose, across a wide range of VLMs (and LLMs with OCR), three patterns: (i) low- budget accuracy collapses on the high-human- error tail at every model size; (ii) test-time scaling (TTS) raises token use roughly linearly with cohort error rate, while accuracy gains follow a non-monotonic curve; (iii) within a single family, TTS flips between anti-scaling on the hardest items and overthinking on eas- ier ones — two faces of the same alignment failure. With the proposed DRG metric, we find that models with near-identical accuracy can sit at near-opposite values: one model gets wrong what humans also find hard, while another solves the hardest items yet fails on items humans find easy, which is a critical con- trast that aggregate accuracy hides. Our code and dataset builder is fully open-sourced at https://github.com/naver-ai/KCSAT-ML.

Mobile GUI agents deployed in real settings must work across different applications that support the same functionality. Most existing benchmarks test each task in only one app, so a high score can mean the agent understands the task, or only that it knows that particular app. We introduce AnyAppBench, a category-controlled live Android benchmark that evaluates cross-application generalization while keeping the user goal fixed. It spans 10 functional categories, 100 task templates, and 520 task–application pairs over 52 applications. Agents run from raw instructions and with app-independent sub-goals, and a VLM judge labels every failed run under a fixed failure taxonomy whose reliability is measured by human annotation. We find that, across 13 agents, success on the original application does not transfer reliably to new applications with the same goal. Furthermore, providing high-level sub-goal decomposition produces only small, category-dependent changes that do not close the gap, and the mix of failure types changes with the target interface. Based on those insights, we believe the AnyAppBench benchmark provides an important stepping stone toward robust real-world deployment of mobile GUI agents. Our code, data and the leaderboard can be found at the project website https://anyappbench.github.io/.

KV caches are a major bottleneck in long-context inference and long-form generation with large language models. Existing training-free eviction policies largely rely on proxy importance signals, such as attention mass, to decide which past tokens to retain. We argue that cache compression should instead preserve the predictive behavior of the full-cache model, retaining entries whose removal would substantially change the model’s output distribution. We propose Behavior-Preserving KV Cache Compression, a training-free framework that scores candidate evictions by estimating the compressed-cache logits induced by their removal and evaluating the resulting KL to the full-cache next-token distribution. Using pre-eviction forward statistics, the method avoids running separate masked forward passes for each candidate. Across diverse architectures and both prefill-time and generation-time compression, our method delivers substantial gains in downstream task quality over lightweight attention-based heuristics at matched retained-KV budgets, with the largest gains under aggressive compression. It achieves these gains with additional compression-time computation while retaining an end-to-end speedup over full-cache inference in our evaluated settings.

Concurrent multi-agent LLM systems increasingly rely on role-specialized Workers that act in parallel, but effective coordination depends not only on what Workers generate, but also on what task state each Worker observes before generating. Existing systems usually fix this observation interface through workflow rules, retrieval heuristics, or full-history sharing, so information flow cannot adapt as task state and feedback evolve. We introduce Dynamic Context Allocation (DCA), which treats per-Worker observation control as a trainable coordination problem for a fixed-role team. An RL-trained Context Allocator maintains a bounded Context Board and decides before each round which evolving task-state entries should be retained, revised, removed, or exposed to each Worker, rather than learning agent selection, execution order, or workflow topology. We evaluate DCA on cooperative code generation, Customer Relationship Management (CRM) lead qualification, and MultiWebGen, a new benchmark for multi-agent full-stack web development. Across these settings, DCA improves over fixed-protocol coordination and fixed-context multi-agent RL baselines, and it outperforms all evaluated frontier single-agent LLMs on CRM lead qualification despite using smaller role-specialized Workers.

Large language models (LLMs) undergo extensive safety training to maximize both helpfulness and harmlessness in their responses. However, various jailbreak attacks jeopardize model safety, allowing malicious actors to bypass safety guidelines. Existing defense methods primarily focus on aligning the model’s output towards less harmful responses through post-processing or input perturbation. Consequently, these approaches are prone to general performance degradation and lack the ability to defend against a wide variety of attacks. In this paper, we propose goal-conditioned direct preference optimization (GC-DPO), which is trained to prioritize the system prompt over the user prompt through goal-conditioning, and thus enables a good balance between safety and performance. Empirically, we show that our approach significantly reduces the average Attack Success Rate (ASR) on a wide variety of jailbreak attacks. In particular, GC-DPO achieves a reduction of 67.1% to 5.0% in ASR for Vicuna-7B, a state-of-the-art result, without compromising the model’s general performance.

A critical component of the current generation of language models is preference alignment, which aims to precisely control the model’s behavior to meet human needs and values. The most notable among such methods is Reinforcement Learning with Human Feedback (RLHF) and its offline variant Direct Preference Optimization (DPO), both of which seek to maximize a reward model based on human preferences. In particular, DPO derives reward signals directly from the offline preference data, but in doing so overfits the reward signals and generates suboptimal responses that may contain human biases in the dataset. In this work, we propose a practical application of a diversity-seeking RL algorithm called GFlowNet-DPO (GDPO) in an offline preference alignment setting to curtail such challenges. Empirical results show GDPO can generate far more diverse responses than the baseline methods that are still relatively aligned with human values in dialog generation and summarization tasks.

With the advent of foundation models, prompt tuning has positioned itself as an important technique for directing model behaviors and eliciting desired responses. Prompt tuning regards selecting appropriate keywords included into the input, thereby adapting to the downstream task without adjusting or fine-tuning the model parameters. There is a wide range of work in prompt tuning, from approaches that directly harness the backpropagated gradient signals from the model, to those employing black-box optimization such as reinforcement learning (RL) methods. Our primary focus is on RLPrompt, which aims to find optimal prompt tokens leveraging soft Q-learning. While the results show promise, we have observed that the prompts frequently appear unnatural, which impedes their interpretability. We address this limitation by using sparse Tsallis entropy regularization, a principled approach to filtering out unlikely tokens from consideration. We extensively evaluate our approach across various tasks, including few-shot text classification, unsupervised text style transfer, and textual inversion from images. The results indicate a notable improvement over baselines, highlighting the efficacy of our approach in addressing the challenges of prompt tuning. Moreover, we show that the prompts discovered using our method are more natural and interpretable compared to those from other baselines.

Prompt tuning, in which prompts are optimized to adapt large-scale pre-trained language models to downstream tasks instead of fine-tuning the full model parameters, has been shown to be particularly effective when the prompts are trained in the multi-task transfer learning setting. These methods generally involve individually training prompts for each source task and then aggregating them to provide the initialization of the prompt for the target task. However, this approach critically ignores the fact that some of the source tasks could be negatively or positively interfering with each other. We argue that when we extract knowledge from source tasks via training source prompts, we need to consider this correlation among source tasks for better transfer to target tasks. To this end, we propose a Bayesian approach where we work with the posterior distribution of prompts across source tasks. We obtain representative source prompts corresponding to the samples from the posterior utilizing Stein Variational Gradient Descent, which are then aggregated to constitute the initial target prompt. We show extensive experimental results on the standard benchmark NLP tasks, where our Bayesian multi-task transfer learning approach outperforms the state-of-the-art methods in many settings. Furthermore, our approach requires no auxiliary models other than the prompt itself, achieving high degree of parameter-efficiency.

Although there have been remarkable advances in dialogue systems through the dialogue systems technology competition (DSTC), it remains one of the key challenges to building a robust task-oriented dialogue system with a speech interface. Most of the progress has been made for text-based dialogue systems since there are abundant datasets with written cor- pora while those with spoken dialogues are very scarce. However, as can be seen from voice assistant systems such as Siri and Alexa, it is of practical importance to transfer the success to spoken dialogues. In this paper, we describe our engineering effort in building a highly successful model that participated in the speech-aware dialogue systems technology challenge track in DSTC11. Our model consists of three major modules: (1) automatic speech recognition error correction to bridge the gap between the spoken and the text utterances, (2) text-based dialogue system (D3ST) for estimating the slots and values using slot descriptions, and (3) post-processing for recovering the error of the estimated slot value. Our experiments show that it is important to use an explicit automatic speech recognition error correction module, post-processing, and data augmentation to adapt a text-based dialogue state tracker for spoken dialogue corpora.

The Situated Interactive Multi-Modal Conversations (SIMMC) 2.0 aims to create virtual shopping assistants that can accept complex multi-modal inputs, i.e. visual appearances of objects and user utterances. It consists of four subtasks, multi-modal disambiguation (MM-Disamb), multi-modal coreference resolution (MM-Coref), multi-modal dialog state tracking (MM-DST), and response retrieval and generation. While many task-oriented dialog systems usually tackle each subtask separately, we propose a jointly learned multi-modal encoder-decoder that incorporates visual inputs and performs all four subtasks at once for efficiency. This approach won the MM-Coref and response retrieval subtasks and nominated runner-up for the remaining subtasks using a single unified model at the 10th Dialog Systems Technology Challenge (DSTC10), setting a high bar for the novel task of multi-modal task-oriented dialog systems.

The goal-oriented dialogue system needs to be optimized for tracking the dialogue flow and carrying out an effective conversation under various situations to meet the user goal. The traditional approach to build such a dialogue system is to take a pipelined modular architecture, where its modules are optimized individually. However, such an optimization scheme does not necessarily yield the overall performance improvement of the whole system. On the other hand, end-to-end dialogue systems with monolithic neural architecture are often trained only with input-output utterances, without taking into account the entire annotations available in the corpus. This scheme makes it difficult for goal-oriented dialogues where the system needs to integrate with external systems or to provide interpretable information about why the system generated a particular response. In this paper, we present an end-to-end neural architecture for dialogue systems that addresses both challenges above. In the human evaluation, our dialogue system achieved the success rate of 68.32%, the language understanding score of 4.149, and the response appropriateness score of 4.287, which ranked the system at the top position in the end-to-end multi-domain dialogue system task in the 8th dialogue systems technology challenge (DSTC8).

We present PyOpenDial, a Python-based domain-independent, open-source toolkit for spoken dialogue systems. Recent advances in core components of dialogue systems, such as speech recognition, language understanding, dialogue management, and language generation, harness deep learning to achieve state-of-the-art performance. The original OpenDial, implemented in Java, provides a plugin architecture to integrate external modules, but lacks Python bindings, making it difficult to interface with popular deep learning frameworks such as Tensorflow or PyTorch. To this end, we re-implemented OpenDial in Python and extended the toolkit with a number of novel functionalities for neural dialogue state tracking and action planning. We describe the overall architecture and its extensions, and illustrate their use on an example where the system response model is implemented with a recurrent neural network.

One of the crucial components of dialog system is the dialog state tracker, which infers user’s intention from preliminary speech processing. Since the overall performance of the dialog system is heavily affected by that of the dialog tracker, it has been one of the core areas of research on dialog systems. In this paper, we present a dialog state tracker that combines a generative probabilistic model of dialog state tracking with the recurrent neural network for encoding important aspects of the dialog history. We describe a two-step gradient descent algorithm that optimizes the tracker with a complex loss function. We demonstrate that this approach yields a dialog state tracker that performs competitively with top-performing trackers participated in the first and second Dialog State Tracking Challenges.