Evaluating large language models typically requires thousands of benchmark items, making the process expensive and increasingly impractical. Existing protocols rely on average accuracy over fixed item sets, treating all items as equally informative—yet many items are too easy or too hard to distinguish between models, making evaluation both wasteful and imprecise. Item Response Theory (IRT) with Computerized Adaptive Testing (CAT) offers a principled alternative—separately estimating model ability and item difficulty, then adaptively selecting the most informative items. However, practical deployment is bottlenecked by costly item calibration, and standard greedy selection causes severe over-exposure of frequently picked items while leaving the rest of the bank under-utilized, which risks data contamination over time. We propose FAIR-Eval with two components: Feature-Driven Calibration (FDC), which jointly learns item parameters from text features and sparse responses, enabling low-cost bank expansion requiring only minimal model responses; and Maximum Fisher Information Plus (MFI+), which counteracts greedy selection by incorporating exposure control, while simultaneously enforcing difficulty coverage and domain diversity to systematically activate under-utilized items. On a 6,796-item bank (193 calibration models, 50 held-out), FAIR-Eval achieves Pearson r = 0.9882 and Spearman 𝜌 = 0.9828 against full evaluation using around 3% of items. Crucially, it broadens difficulty coverage from 60.0% to 68.7%, achieving a more comprehensive and robust assessment of model capabilities with minimal test length.

Tangible control interfaces (TCIs), such as appliance panels, remotes, elevators, and embedded GUIs, are a fundamental component of everyday human-built environments. Interacting with these interfaces requires agents not only to ground language in visual observations, but also to execute actions, track temporally evolving state changes, and verify whether intended outcomes have been achieved. However, existing benchmarks predominantly evaluate open-loop perception or single-step action execution, failing to capture this continuous cycle of interaction, feedback, and correction. We introduce SWITCH, a benchmark for closed-loop interactive reasoning with TCIs inrealistic egocentric environments1. SWITCH comprises 1,170 temporally interactive videos across diverse functional categories, providing structured annotations of instructions, actions,state transitions, outcomes, and recovery behaviors over time. To probe generative world modeling, SWITCH also evaluates video generation models on interaction-centered tasks using both LLM-as-judge and human evaluation.Experiments with frontier proprietary and opensource multimodal models reveal persistent weaknesses in fine-grained visual-temporal perception, outcome verification, and error recovery, highlighting SWITCH as a testbed for closed-loop embodied intelligence

Reinforcement learning with outcome-based objectives such as GRPO enables LLM-based agents to solve complex, long-horizon tasks, yet the reusable exploration patterns embedded in interaction trajectories are largely discarded after a single policy update. Existing experience-augmented approaches retrieve historical guidance at inference time, but they apply experiences without accounting for the policy’s evolving capability and create persistent dependencies on external retrieval. We propose EDGE (Experience-Distillation for Guided Exploration), a framework that treats retrieved experiences as temporary training-time scaffolds and progressively internalizes their benefits into the parametric policy. Concretely, EDGE partitions each rollout group into experience-conditioned and experience-free trajectories to estimate and admit only positive marginal gains without extra sampling, then distills the induced behavior into the base policy via a reverse-KL objective on its own empirical support. A co-evolutionary experience bank further synthesizes guidance from emerging failure modes and prunes obsolete entries as the policy evolves. Across embodied, web, and search-based QA tasks, EDGE improves over strong RL baselines by up to 12.5 points and remains effective without inference-time scaffolds or a proprietary reflector. The code is available at https://github.com/xvolcano02/EDGE.

Language-model agents increasingly rely on skills: reusable procedural knowledge for reasoning, tool use, and interaction. Existing work studies how skills are acquired, retrieved, compressed, or composed, but often assumes that once a skill is selected, its interface to the agent is fixed. We argue that this overlooks a key source of skill utility: the same skill can help, distract, or mislead depending on how it is exposed. We propose SkillAlign, a provider-agnostic framework that represents candidate skills as multi-view procedural cards and renders them through alternative exposure interfaces, including full instructions, hints, compressed summaries, workflows, or no exposure. This enables counterfactual evaluation where the task, agent, and candidate skills are fixed while only the exposure interface varies. Across ALFWorld and SkillsBench, we show that exposure form substantially affects task success and rendered context cost, and that compact top-k exposure can outperform full-library injection. We further conduct a replay-based policy-learning analysis on ALFWorld, showing that adaptive exposure contains learnable signal but remains far from oracle selection. Our results suggest that skill-augmented agents should optimize not only which skills to use, but also how those skills are presented.

Complex multi-step reasoning remains challenging for large language models (LLMs). While parallel inference-time scaling methods, such as step-level beam search, offer a promising solution, existing approaches typically depend on either domain-specific external verifiers, or self-evaluation which is brittle and prompt-sensitive. To address these issues, we propose Collaborative Beam Search (CBS), an iterative framework that harnesses the collective intelligence of multiple LLMs across both generation and verification stages. For generation, CBS leverages multiple LLMs to explore a broader search space, resulting in more diverse candidate steps. For verifications, CBS employs a perplexity-based collective consensus among these models, eliminating reliance on an external verifier or complex prompts. Between iterations, CBS leverages a dynamic quota allocation strategy that reassigns generation budget based on each model’s past performance, striking a balance between candidate diversity and quality. Experimental results on six tasks across arithmetic, logical, and commonsense reasoning show that CBS outperforms single‐model scaling and multi-model ensemble baselines by over 4 percentage points in average accuracy, demonstrating its effectiveness and general applicability.

Recent advances in text-only “slow-thinking” reasoning have prompted efforts to transfer this capability to vision-language models (VLMs), for training visual reasoning models (VRMs). However, such transfer faces critical challenges: Effective “slow thinking” in VRMs requires visual reflection, the ability to check the reasoning process based on visual information. Through quantitative analysis, we observe that current VRMs exhibit limited visual reflection, as their attention to visual information diminishes rapidly with longer generated responses. To address this challenge, we propose a new VRM Reflection-V, which enhances visual reflection based on reasoning data construction for cold-start and reward design for reinforcement learning (RL). Firstly, we construct vision-centered reasoning data by leveraging an agent that interacts between VLMs and reasoning LLMs, enabling cold-start learning of visual reflection patterns. Secondly, a visual attention based reward model is employed during RL to encourage reasoning based on visual information. Therefore, Reflection-V demonstrates significant improvements across multiple visual reasoning benchmarks. Furthermore, Reflection-V maintains a stronger and more consistent reliance on visual information during visual reasoning, indicating effective enhancement in visual reflection capabilities.

In visual question answering (VQA) context, users often pose ambiguous questions to visual language models (VLMs) due to varying expression habits. Existing research addresses such ambiguities primarily by rephrasing questions. These approaches neglect the inherently interactive nature of user interactions with VLMs, where ambiguities can be clarified through user feedback. However, research on interactive clarification faces two major challenges: (1) Benchmarks are absent to assess VLMs’ capacity for resolving ambiguities through interaction; (2) VLMs are trained to prefer answering rather than asking, preventing them from seeking clarification. To overcome these challenges, we introduce ClearVQA benchmark, which targets three common categories of ambiguity in VQA context, and encompasses various VQA scenarios. Furthermore, we propose an automated pipeline to generate ambiguity-clarification question pairs, enabling VLMs to ask reasonable clarification questions and generate more accurate and specific answers based on user feedback, as demonstrated by experimental results.

Bilingual lexicon induction induces the word translations by aligning independently trained word embeddings in two languages. Existing approaches generally focus on minimizing the distances between words in the aligned pairs, while suffering from low discriminative capability to distinguish the relative orders between positive and negative candidates. In addition, the mapping function is globally shared by all words, whose performance might be hindered by the deviations in the distributions of different languages. In this work, we propose a novel ranking-oriented induction model RAPO to learn personalized mapping function for each word. RAPO is capable of enjoying the merits from the unique characteristics of a single word and the cross-language isomorphism simultaneously. Extensive experimental results on public datasets including both rich-resource and low-resource languages demonstrate the superiority of our proposal. Our code is publicly available in https://github.com/Jlfj345wf/RAPO.

Motivated by the success of T5 (Text-To-Text Transfer Transformer) in pre-trained natural language processing models, we propose a unified-modal SpeechT5 framework that explores the encoder-decoder pre-training for self-supervised speech/text representation learning. The SpeechT5 framework consists of a shared encoder-decoder network and six modal-specific (speech/text) pre/post-nets. After preprocessing the input speech/text through the pre-nets, the shared encoder-decoder network models the sequence-to-sequence transformation, and then the post-nets generate the output in the speech/text modality based on the output of the decoder. Leveraging large-scale unlabeled speech and text data, we pre-train SpeechT5 to learn a unified-modal representation, hoping to improve the modeling capability for both speech and text. To align the textual and speech information into this unified semantic space, we propose a cross-modal vector quantization approach that randomly mixes up speech/text states with latent units as the interface between encoder and decoder. Extensive evaluations show the superiority of the proposed SpeechT5 framework on a wide variety of spoken language processing tasks, including automatic speech recognition, speech synthesis, speech translation, voice conversion, speech enhancement, and speaker identification.

Though some recent works focus on injecting sentiment knowledge into pre-trained language models, they usually design mask and reconstruction tasks in the post-training phase. This paper aims to integrate sentiment knowledge in the fine-tuning stage. To achieve this goal, we propose two sentiment-aware auxiliary tasks named sentiment word selection and conditional sentiment prediction and, correspondingly, integrate them into the objective of the downstream task. The first task learns to select the correct sentiment words from the given options. The second task predicts the overall sentiment polarity, with the sentiment polarity of the word given as prior knowledge. In addition, two label combination methods are investigated to unify multiple types of labels in each auxiliary task. Experimental results demonstrate that our approach consistently outperforms baselines (achieving a new state-of-the-art) and is complementary to existing sentiment-enhanced post-trained models.

While pre-training techniques are working very well in natural language processing, how to pre-train a decoder and effectively use it for neural machine translation (NMT) still remains a tricky issue. The main reason is that the cross-attention module between the encoder and decoder cannot be pre-trained, and the combined encoder-decoder model cannot work well in the fine-tuning stage because the inputs of the decoder cross-attention come from unknown encoder outputs. In this paper, we propose a better pre-training method for NMT by defining a semantic interface (SemFace) between the pre-trained encoder and the pre-trained decoder. Specifically, we propose two types of semantic interfaces, including CL-SemFace which regards cross-lingual embeddings as an interface, and VQ-SemFace which employs vector quantized embeddings to constrain the encoder outputs and decoder inputs in the same language-independent space. We conduct massive experiments on six supervised translation pairs and three unsupervised pairs. Experimental results demonstrate that our proposed SemFace can effectively connect the pre-trained encoder and decoder, and achieves significant improvement by 3.7 and 1.5 BLEU points on the two tasks respectively compared with previous pre-training-based NMT models.

The commonly used framework for unsupervised machine translation builds initial translation models of both translation directions, and then performs iterative back-translation to jointly boost their translation performance. The initialization stage is very important since bad initialization may wrongly squeeze the search space, and too much noise introduced in this stage may hurt the final performance. In this paper, we propose a novel retrieval and rewriting based method to better initialize unsupervised translation models. We first retrieve semantically comparable sentences from monolingual corpora of two languages and then rewrite the target side to minimize the semantic gap between the source and retrieved targets with a designed rewriting model. The rewritten sentence pairs are used to initialize SMT models which are used to generate pseudo data for two NMT models, followed by the iterative back-translation. Experiments show that our method can build better initial unsupervised translation models and improve the final translation performance by over 4 BLEU scores. Our code is released at https://github.com/Imagist-Shuo/RRforUNMT.git.

Unsupervised bilingual lexicon induction is the task of inducing word translations from monolingual corpora of two languages. Recent methods are mostly based on unsupervised cross-lingual word embeddings, the key to which is to find initial solutions of word translations, followed by the learning and refinement of mappings between the embedding spaces of two languages. However, previous methods find initial solutions just based on word-level information, which may be (1) limited and inaccurate, and (2) prone to contain some noise introduced by the insufficiently pre-trained embeddings of some words. To deal with those issues, in this paper, we propose a novel graph-based paradigm to induce bilingual lexicons in a coarse-to-fine way. We first build a graph for each language with its vertices representing different words. Then we extract word cliques from the graphs and map the cliques of two languages. Based on that, we induce the initial word translation solution with the central words of the aligned cliques. This coarse-to-fine approach not only leverages clique-level information, which is richer and more accurate, but also effectively reduces the bad effect of the noise in the pre-trained embeddings. Finally, we take the initial solution as the seed to learn cross-lingual embeddings, from which we induce bilingual lexicons. Experiments show that our approach improves the performance of bilingual lexicon induction compared with previous methods.

Pre-training has proven to be effective in unsupervised machine translation due to its ability to model deep context information in cross-lingual scenarios. However, the cross-lingual information obtained from shared BPE spaces is inexplicit and limited. In this paper, we propose a novel cross-lingual pre-training method for unsupervised machine translation by incorporating explicit cross-lingual training signals. Specifically, we first calculate cross-lingual n-gram embeddings and infer an n-gram translation table from them. With those n-gram translation pairs, we propose a new pre-training model called Cross-lingual Masked Language Model (CMLM), which randomly chooses source n-grams in the input text stream and predicts their translation candidates at each time step. Experiments show that our method can incorporate beneficial cross-lingual information into pre-trained models. Taking pre-trained CMLM models as the encoder and decoder, we significantly improve the performance of unsupervised machine translation.

Neural Machine Translation (NMT) performs poor on the low-resource language pair (X,Z), especially when Z is a rare language. By introducing another rich language Y, we propose a novel triangular training architecture (TA-NMT) to leverage bilingual data (Y,Z) (may be small) and (X,Y) (can be rich) to improve the translation performance of low-resource pairs. In this triangular architecture, Z is taken as the intermediate latent variable, and translation models of Z are jointly optimized with an unified bidirectional EM algorithm under the goal of maximizing the translation likelihood of (X,Y). Empirical results demonstrate that our method significantly improves the translation quality of rare languages on MultiUN and IWSLT2012 datasets, and achieves even better performance combining back-translation methods.

In order to alleviate data sparsity and overfitting problems in maximum likelihood estimation (MLE) for sequence prediction tasks, we propose the Generative Bridging Network (GBN), in which a novel bridge module is introduced to assist the training of the sequence prediction model (the generator network). Unlike MLE directly maximizing the conditional likelihood, the bridge extends the point-wise ground truth to a bridge distribution conditioned on it, and the generator is optimized to minimize their KL-divergence. Three different GBNs, namely uniform GBN, language-model GBN and coaching GBN, are proposed to penalize confidence, enhance language smoothness and relieve learning burden. Experiments conducted on two recognized sequence prediction tasks (machine translation and abstractive text summarization) show that our proposed GBNs can yield significant improvements over strong baselines. Furthermore, by analyzing samples drawn from different bridges, expected influences on the generator are verified.