LLMs are often trained with RL from human or AI feedback, yet such methods typically compress nuanced feedback into scalar rewards, discarding much of their richness and inducing scale imbalance. We propose treating verbal feedback as a conditioning signal. Inspired by language priors in text-to-image generation, which enable novel outputs from unseen prompts, we introduce the feedback-conditional policy (FCP). FCP learns directly from response-feedback pairs, approximating the feedback-conditional posterior through maximum likelihood training on offline data. We further develop an online bootstrapping stage where the policy generates under positive conditions and receives fresh feedback to refine itself. This reframes feedback-driven learning as conditional generation rather than reward optimization, offering a more expressive way for LLMs to directly learn from verbal feedback. Our code is available at https://github.com/sail-sg/feedback-conditional-policy.

While Transformer-based models have demonstrated remarkable language modeling performance, their high complexities result in high costs when processing long contexts. In contrast, recurrent neural networks (RNNs) such as linear attention and state space models have gained popularity due to their constant per-token complexities. However, these recurrent models struggle with tasks that require accurate recall of contextual information from long contexts, because all contextual information is compressed into a constant-size recurrent state. Previous works have shown that recall ability is positively correlated with the recurrent state size, yet directly training RNNs with larger recurrent states results in high training costs. In this paper, we introduce StateX, a training pipeline for efficiently expanding the states of pre-trained RNNs through post-training. For two popular classes of RNNs, linear attention and state space models, we design post-training architectural modifications to scale up the state size with no or negligible increase in model parameters. Experiments on models up to 1.3B parameters demonstrate that StateX efficiently enhances the recall and in-context learning ability of RNNs without incurring high post-training costs or compromising other capabilities. Recently, recurrent neural networks (RNNs) such as gated linear attention (GLA) (Y ang et al., 2024) and Mamba2 (Dao & Gu, 2024) have shown promising capabilities in language modeling. These architectures have constant per-token complexity, while the more popular Transformer architecture (V aswani et al., 2023) has per-token complexity that grows linearly with the context length.

Heuristic functions are central to the performance of search algorithms such as A-star, where admissibility - the property of never overestimating the true shortest-path cost - guarantees solution optimality. Recent deep learning approaches often disregard admissibility and provide limited guarantees on generalization beyond the training data. This paper addresses both of these limitations. First, we pose heuristic learning as a constrained optimization problem and introduce Cross-Entropy Admissibility (CEA), a loss function that enforces admissibility during training. On the Rubik's Cube domain, this method yields near-admissible heuristics with significantly stronger guidance than compressed pattern database (PDB) heuristics. Theoretically, we study the sample complexity of learning heuristics. By leveraging PDB abstractions and the structural properties of graphs such as the Rubik's Cube, we tighten the bound on the number of training samples needed for A-star to generalize. Replacing a general hypothesis class with a ReLU neural network gives bounds that depend primarily on the network's width and depth, rather than on graph size. Using the same network, we also provide the first generalization guarantees for goal-dependent heuristics.

Reinforcement Learning with V erifiable Rewards (RL VR) strengthens LLM reasoning but training often oscillates between entropy collapse and entropy explosion. We trace both hazards to the mean-baseline used in value-free RL (e.g., GRPO & DAPO), which improperly penalizes negative-advantage samples under reward outliers. We propose Quantile Advantage Estimation (QAE), replacing the mean with a group-wise K-quantile baseline. QAE induces a response-level, two-regime gate: on hard queries (p 1 K) it reinforces rare successes, while on easy queries (p > 1 K) it targets remaining failures. Under first-order soft-max updates, we prove two-sided entropy safety, giving lower/upper bounds on one-step entropy change that curb explosion and prevent collapse. Empirically, this minimal modification stabilizes entropy, sparsifies credit assignment (with tuned K, roughly 80% of responses receive zero advantage), and yields sustained pass@1 gains on Qwen3-8B/14B-Base across AIME'24/'25 and AMC'23. Reinforcement Learning with V erifiable Rewards (RL VR) (Lambert et al., 2024; DeepSeek-AI et al., 2025; Y ang et al., 2025a) enhances Large Language Models (LLMs) by rewarding verifiable correctness (Phan et al., 2025; Rein et al., 2023). Y et reward-driven optimization often triggers entropy collapse (Y u et al., 2025; Cui et al., 2025): the policy distribution sharpens prematurely, suppressing exploration and ultimately limiting performance.

Deliberation plays a crucial role in shaping outcomes by weighing diverse perspectives before reaching decisions. With recent advancements in Natural Language Processing, it has become possible to computationally model deliberation by analyzing opinion shifts and predicting potential outcomes under varying scenarios. In this study, we present a comparative analysis of multiple NLP techniques to evaluate how effectively models interpret deliberative discourse and produce meaningful insights. Opinions from individuals of varied backgrounds were collected to construct a self-sourced dataset that reflects diverse viewpoints. Deliberation was simulated using product presentations enriched with striking facts, which often prompted measurable shifts in audience opinions. We have given comparative analysis between two models namely Frequency-Based Discourse Modulation and Quantum-Deliberation Framework which outperform the existing state of art models. Deliberation is the structured process of reasoning, dialogue, and weighing evidence before decisions are made. Unlike ordinary conversation, it emphasizes logical argumentation, inclusivity, and critical reflection.

Efficiently detecting human intent to interact with ubiquitous robots is crucial for effective human-robot interaction (HRI) and collaboration. Over the past decade, deep learning has gained traction in this field, with most existing approaches relying on multimodal inputs, such as RGB combined with depth (RGB-D), to classify time-sequence windows of sensory data as interactive or non-interactive. In contrast, we propose a novel RGB-only pipeline for predicting human interaction intent with frame-level precision, enabling faster robot responses and improved service quality. A key challenge in intent prediction is the class imbalance inherent in real-world HRI datasets, which can hinder the model's training and generalization. To address this, we introduce MINT-RVAE, a synthetic sequence generation method, along with new loss functions and training strategies that enhance generalization on out-of-sample data. Our approach achieves state-of-the-art performance (AUROC: 0.95) outperforming prior works (AUROC: 0.90-0.912), while requiring only RGB input and supporting precise frame onset prediction. Finally, to support future research, we openly release our new dataset with frame-level labeling of human interaction intent.

Test-Time Scaling (TTS) enhances the reasoning ability of large language models (LLMs) by allocating additional computation during inference. However, existing approaches primarily rely on output-level sampling while overlooking the role of model architecture. In mainstream Mixture-of-Experts (MoE) LLMs, we observe that varying the number of activated experts yields complementary solution sets with stable accuracy, revealing a new and underexplored source of diversity. Motivated by this observation, we propose Dynamic Experts Search (DES), a TTS strategy that elevates expert activation into a controllable dimension of the search space. DES integrates two key components: (1) Dynamic MoE, which enables direct control of expert counts during inference to generate diverse reasoning trajectories without additional cost; and (2) Expert Configuration Inheritance, which preserves consistent expert counts within a reasoning path while varying them across runs, thereby balancing stability and diversity throughout the search. Extensive experiments across MoE architectures, verifiers and reasoning benchmarks (i.e., math, code and knowledge) demonstrate that DES reliably outperforms TTS baselines, enhancing accuracy and stability without additional cost. These results highlight DES as a practical and scalable form of architecture-aware TTS, illustrating how structural flexibility in modern LLMs can advance reasoning.

The rapid progress of Large Multimodal Models (LMMs) and cloud-based AI agents is transforming human-AI collaboration into bidirectional, multimodal interaction. However, existing codecs remain optimized for unimodal, one-way communication, resulting in repeated degradation under conventional compress-transmit-reconstruct pipelines. To address this limitation, we propose UniMIC, a Unified token-based Multimodal Interactive Coding framework that bridges edge devices and cloud AI agents. Instead of transmitting raw pixels or plain text, UniMIC employs compact tokenized representations as the communication medium, enabling efficient low-bitrate transmission while maintaining compatibility with LMMs. To further enhance compression, lightweight Transformer-based entropy models with scenario-specific designs-generic, masked, and text-conditioned-effectively minimize inter-token redundancy. Extensive experiments on text-to-image generation, text-guided inpainting, outpainting, and visual question answering show that UniMIC achieves substantial bitrate savings and remains robust even at ultra-low bitrates (<0.05bpp), without compromising downstream task performance. These results establish UniMIC as a practical and forward-looking paradigm for next-generation multimodal interactive communication.

Asynchronous patient-clinician messaging via EHR portals is a growing source of clinician workload, prompting interest in large language models (LLMs) to assist with draft responses. However, LLM outputs may contain clinical inaccuracies, omissions, or tone mismatches, making robust evaluation essential. Our contributions are threefold: (1) we introduce a clinically grounded error ontology comprising 5 domains and 59 granular error codes, developed through inductive coding and expert adjudication; (2) we develop a retrieval-augmented evaluation pipeline (RAEC) that leverages semantically similar historical message-response pairs to improve judgment quality; and (3) we provide a two-stage prompting architecture using DSPy to enable scalable, interpretable, and hierarchical error detection. Our approach assesses the quality of drafts both in isolation and with reference to similar past message-response pairs retrieved from institutional archives. Using a two-stage DSPy pipeline, we compared baseline and reference-enhanced evaluations on over 1,500 patient messages. Retrieval context improved error identification in domains such as clinical completeness and workflow appropriateness. Human validation on 100 messages demonstrated superior agreement (concordance = 50% vs. 33%) and performance (F1 = 0.500 vs. 0.256) of context-enhanced labels vs. baseline, supporting the use of our RAEC pipeline as AI guardrails for patient messaging.

Electroencephalography (EEG), with its broad range of applications, necessitates models that can generalize effectively across various tasks and datasets. Large EEG Models (LEMs) address this by pretraining encoder-centric architectures on large-scale unlabeled data to extract universal representations. While effective, these models lack decoders of comparable capacity, limiting the full utilization of the learned features. To address this issue, we introduce ECHO, a novel decoder-centric LEM paradigm that reformulates EEG modeling as sequence-to-sequence learning. ECHO captures layered relationships among signals, labels, and tasks within sequence space, while incorporating discrete support samples to construct contextual cues. This design equips ECHO with in-context learning, enabling dynamic adaptation to heterogeneous tasks without parameter updates. Extensive experiments across multiple datasets demonstrate that, even with basic model components, ECHO consistently outperforms state-of-the-art single-task LEMs in multi-task settings, showing superior generalization and adaptability.