Multi-agent debate (MAD) improves the reasoning capabilities of large language models by having multiple agents iteratively refine their responses through discussion. However, MAD suffers from a critical vulnerability known as shared misconception: when a majority of agents initially converge on an incorrect answer, the debate process tends to amplify rather than correct the error. Existing methods primarily address peer skew but leave the agents’ inherently biased concept priors unaddressed. To mitigate this systematic weakness, we propose **R²-MAD** (Remember and Reweight for Multi-Agent Debate), a framework that equips agents with an experience memory accumulated from past debates. R²-MAD intervenes on both failure modes through two complementary mechanisms: A debate-state-aware retrieval policy dynamically calibrates the concept prior by retrieving relevant historical evidence based on the current consensus level. Then these retrieved experiences provide a basis for estimating per-agent reliability, yielding confidence weights to modulate peer influence. Experiments on various benchmarks show that R²-MAD achieves consistent improvements over existing single-agent and MAD baselines.

The significance of tasks entrusted to LLM-based assistants (agents) and the associated societal risks are increasing each year. Agents are being explored in critical domains such as medicine, finance, law, infrastructure, and other sensitive applications that require system transparency and high user trust. The quality of these agents is typically evaluated by accuracy, sometimes extended to partial correctness. In this position paper, we argue that this focus on outcomes is insufficient as it can obscure risky agent behaviours such as skipping critical steps, hallucinating tool use, relying on outdated parametric knowledge and other means of bypassing recommended processes. Our core position is that a holistic agent evaluation must include process evaluation, especially for critical applications. We conduct a small-scale study to assess the feasibility of automatic process evaluation, present a compliance score, analyse use cases of bad and good behaviours, and offer recommendations for more holistic evaluation.

Motivated by the observed human-like behaviours in Large Reasoning Models (LRMs), this paper introduces a comprehensive taxonomy to characterise atomic reasoning steps and analyse the reasoning behaviours of LRMs. Grounded in human cognitive processes, we propose a taxonomy comprising five groups and seventeen categories. Through this taxonomy, we conduct an in-depth analysis of contemporary LRMs and distil four actionable takeaways for model optimisation. Most notably, we reveal that prevailing post-answer “double-checks” are largely superficial and rarely yield substantive revisions. A targeted intervention further shows that explicitly eliciting richer reflection processes can substantially improve failed self-correction. To support this large-scale study, we propose CAPO, an automated annotation method used to construct a dataset of 277,534 reasoning steps with strong agreement with human expert annotations. We further validate the main behavioural patterns on a newer reasoning model and a coding domain, demonstrating the broader applicability of the proposed taxonomy. All source code and data are available at https://github.com/hehepig4/psyche.

Tool-augmented vision-language agents can acquire external perceptual evidence through OCR, detection, segmentation, and other tools, but executing every proposed tool call is costly and sometimes unnecessary. We study the pre-call control problem: after a ReAct-style VLM agent proposes a perceptual tool call, should the call be executed or skipped before its output enters the context? Across five benchmarks, we find that the baseline agent exhibits poor local selectivity: helpful and harmful calls occur at similar rates (11.8% vs. 9.9%), while most calls do not change the immediate forced-answer prediction. We introduce ToolGate, a lightweight external controller that predicts execute/skip decisions from trajectory text and simple structural features. Across two Qwen3-VL backbones, ToolGate reduces token cost to 64–69% of the unrestricted ReAct baseline while preserving average accuracy in cross-domain settings. With matched-domain trajectory training on Qwen3-VL-30B, ToolGate additionally improves average accuracy by 1.65 points. These results suggest that tool-augmented VLM agents benefit not only from better perceptual tools, but also from explicit inference-time control over when tool outputs are worth paying for.

Memory is essential for LLM agents to accumulate task experience and reuse task-specific execution strategies. However, real-world deployment over boundary-agnostic and evolving task streams exposes a fundamental stability-plasticity dilemma. External retrieval-based memory can rapidly absorb new evidence, but it often fails to internalize recurring execution patterns and incurs inference-time retrieval overhead. Parametric memory enables stable and efficient execution once learned, but typically relies on explicit task boundaries and fixed parameter budgets. Inspired by the human brain, which balances plasticity and stability through complementary episodic storage and gradual consolidation, we propose UniMem, a self-routing framework for autonomous memory management. UniMem uses learnable routing tokens as memory controllers, enabling adaptive coordination between complementary memory pathways: novel or sparse tasks are retained in an episodic buffer for retrieval-augmented execution, while recurring and reliable patterns are consolidated into expandable parametric memory. By decoupling task identification from task execution with routing tokens and parametric memory blocks, UniMem expands memory on demand without task labels during deployment or uncontrolled parameter growth. Experiments on long-horizon streaming task sequences show that UniMem typically achieves strong aggregate results while maintaining routing and execution stability, with an average gain of 4.0 EM over Replay LoRA.

Large language models (LLMs) excel at natural language tasks but face deployment challenges due to computational demands. We introduce Dual Activation-Weight Sparsity (DAWS), a training-free framework that jointly exploits activation and weight sparsity through magnitude-based routing. Systematic analysis of pretrained transformers reveals two key observations: (1) the activation energy is concentrated in a few neurons, and (2) activation and weight sparsity patterns are complementary between attention and FFN layers. DAWS employs a three-tier routing strategy: high-magnitude activations pass through full-precision weights to preserve critical pathways, medium-magnitude activations use magnitude-pruned sparse weights for efficiency, and low-magnitude activations are directly discarded. Unlike prior work that uses activation-aware pruning methods like WANDA, our approach uses direct magnitude-based pruning, which we show is more robust to sample-level variations. Experiments on Llama and Mistral models demonstrate that DAWS maintains >98% of dense model performance at 50% sparsity, outperforming WANDA, TEAL, and R-Sparse.

Direct alignment algorithms have proven an effective step for aligning language models to human-desired behaviors. Current variants of the Direct Preference Optimization objective have focused on a strict setting where all tokens are contributing signals of KL divergence and rewards to the loss function.However, human preference is not affected equally by each word in a sequence but is often dependent on specific words or phrases, e.g. existence of toxic terms leads to non-preferred responses. Based on this observation, we argue that not all tokens should be weighted equally during PO and propose a flexible objective termed SparsePO, that aims to automatically learn to weight the KL divergence and reward corresponding to each token during PO training. We propose two different variants of weight-masks that can either be derived from the reference model itself or learned on the fly. Notably, our method induces sparsity in the learned masks, allowing the model to learn how to best balance reward and KL divergence contributions at the token level, learning an optimal level of mask sparsity.Extensive experiments illustrate the effectiveness of our approach at aligning to preference proxies, including sentiment control, helpfulness and harmless, and summary quality.Our method obtains +10% and +3% win-rate points in summarization and dialogue scenarios, respectively,without compromising the reasoning capabilities of the model, or the relevancy and faithfulness of the summary response.

The Spiral of Silence (SoS) theory holds that individuals with minority views often refrain from speaking out for fear of social isolation, enabling majority positions to dominate public discourse. When the “agents” are large language models (LLMs), however, the classical psychological explanation is not directly applicable, since SoS was developed for human societies. This raises a central question: can SoS-like dynamics nevertheless emerge from purely statistical language generation in LLM collectives? We propose an evaluation framework for examining SoS in LLM agents. Specifically, we consider four controlled conditions that systematically vary the availability of “History” and “Persona” signals. Opinion dynamics are assessed using trend tests such as Mann–Kendall and Spearman’s rank, along with concentration measures including kurtosis and interquartile range. Experiments across open-source and closed-source models show that history and persona together produce strong majority dominance and replicate SoS patterns; history signals alone induce strong anchoring; and persona signals alone foster diverse but uncorrelated opinions, indicating that without historical anchoring, SoS dynamics cannot emerge. The work bridges computational sociology and responsible AI design, highlighting the need to monitor and mitigate emergent conformity in LLM-agent systems.

To address these limitations, we propose BDC, a novel framework that Boosts reasoning exploration via multi-agent collaboration, Disentangles heterogeneous data into specialized experts, and Customizes solutions through dynamic model composition. BDC integrates a Monte Carlo Tree-of-Agents algorithm, where multiple LLMs mutually verify and refine reasoning paths through reflection-guided pruning, enabling efficient exploration of high-quality solutions. To handle data diversity, we cluster problems by latent semantics, train composable LoRA experts on each cluster, and deploy an input-aware hypernetwork to dynamically merge these experts into tailored solvers. Experiments on APPS and CodeContest benchmarks demonstrate BDC’s superiority: it achieves up to 73.8% accuracy on hard problems, outperforming state-of-the-art methods like LATS and RethinkMCTS by 9–15%. This work lays the groundwork for advancing LLM capabilities in complex reasoning tasks, offering a novel System2-to-System1 solution.

While large language models (LLMs) have significantly advanced mathematical reasoning, Process Reward Models (PRMs) have been developed to evaluate the logical validity of reasoning steps. However, PRMs still struggle with out-of-distribution (OOD) challenges. This paper identifies the OOD issues including step OOD, arising from differences in reasoning patterns across model types and sizes, and question OOD, due to dataset shifts between training and real-world problems. To address these issues, we introduce Retrieval-Augmented Process Reward Model (RetrievalPRM), a novel framework designed to tackle these OOD issues. By utilizing a two-stage retrieval-enhanced mechanism, RetrievalPRM retrieves semantically similar questions and steps for PRM as a warmup to stimulate its potential to judge target steps, improving generalization and reasoning consistency across different models and problem types. Our extensive experiments demonstrate that RetrievalPRM outperforms existing baselines across multiple real-world datasets. Our open-source contributions include a retrieval-enhanced dataset, a tuning framework for PRM training, and the RetreivalPRM model, establishing a new standard for PRM performance.

We introduce Direct Value Optimization (DVO), an innovative offline reinforcement learning framework for enhancing large language models in complex reasoning tasks. Unlike traditional methods relying on preference labels, DVO utilizes value signals at individual reasoning steps, optimizing models via a mean squared error loss. The key benefit of DVO lies in its fine-grained supervision, circumventing the need for labor-intensive human annotations. Target values within the DVO are estimated using either Monte Carlo Tree Search or an outcome value model. Our empirical analysis on 3 math reasoning, 4 commonsense reasoning, and 3 coding tasks shows that DVO consistently outperforms existing offline preference optimization techniques by a significant margin of 4% to 6%, and is competitive to online GRPO but with higher sample efficiency. These findings underscore the importance of value signals in advancing reasoning capabilities and highlight DVO as a superior methodology under scenarios lacking explicit human preference information.

Modelling prosody variation is critical for synthesizing natural and expressive speech in end-to-end text-to-speech (TTS) systems. In this paper, a cross-utterance conditional VAE (CUC-VAE) is proposed to estimate a posterior probability distribution of the latent prosody features for each phoneme by conditioning on acoustic features, speaker information, and text features obtained from both past and future sentences. At inference time, instead of the standard Gaussian distribution used by VAE, CUC-VAE allows sampling from an utterance-specific prior distribution conditioned on cross-utterance information, which allows the prosody features generated by the TTS system to be related to the context and is more similar to how humans naturally produce prosody. The performance of CUC-VAE is evaluated via a qualitative listening test for naturalness, intelligibility and quantitative measurements, including word error rates and the standard deviation of prosody attributes. Experimental results on LJ-Speech and LibriTTS data show that the proposed CUC-VAE TTS system improves naturalness and prosody diversity with clear margins.