Papers for
customer support automation teams
Papers whose findings have a practical use for this group, as judged from the abstract. Open a paper to read what it means in practice.
EvoX improves success for multi-model AI agent tasks by 3 points
EvoX: Rethinking the Agent Harness as a Runtime Policy
Abstract: Large language model (LLM) agents increasingly interact with tools and environments, yet plausible model decisions do not always translate into successful execution due to tool-call mismatches, interrupted workflows, and unsupported completion claims. We present EvoX, a Rust-based, multi-model agent runtime guided by The Second Policy, which views the harness as an Operational Policy complementing the foundation model's Semantic Policy. EvoX combines model-facing guidance with deterministic runtime control through three complementary mechanisms: Action Realization, Execution Continuity, and Outcome Grounding. To evaluate harnesses without suppressing their native behaviors, we introduce Contrast, a controlled and auditable framework integrating explicit fairness constraints, provider-bound request verification, and evidence-based adjudication. Across five benchmarks, six foundation models, and six agent harnesses, EvoX achieves the highest numerical success rate of 56.2% over 691 common task--model pairs, exceeding the average of the other five harnesses by approximately 3.0 percentage points. Its larger advantages on GLM-5.3 (+10.7 points) and Kimi K3 (+7.9 points) highlight the model-dependent effectiveness of harness design and the potential for cross-provider adaptability. Achieved without memory, evolution, or extensions, these results establish EvoX as a competitive multi-model agent runtime and underscore the importance of harness-level operational control in realizing foundation-model capabilities.
Jev offers cost-effective multilingual text judging with some limits
Is Jev a Good Multilingual Judge?
Abstract: Generative large language models are widely used for multilingual text evaluation, yet whether decision models that directly output judgment probabilities can serve as multilingual judges has received limited attention. We evaluate Jev against generative LLM judges on four multilingual benchmarks with human annotations, covering up to 47 language varieties and three evaluation formats: binary decisions, pairwise comparisons, and graded ratings. At approximately $1/31$ to $1/61$ of GPT-6.1's evaluation cost, Jev is competitive on most multilingual judging tasks, but falls behind GPT-6.1 on knowledge-intensive answer verification. Further analysis shows that Jev is more likely to accept incorrect answers in low-resource languages than in English and that its probability outputs help identify errors across tasks. Based on these findings, we offer recommendations on evaluation instructions, confidence-based selection, and model routing, providing empirical guidance for using Jev as a multilingual judge.
Evidence linked memory improves reliability of LLM agents
Don't Just Trust Your Memory: Evidence-Linked Memory Reduces Memory-Induced Degradation in LLM Agents
Abstract: External memory enables LLM agents to improve from experience by distilling trajectories of past tasks into concise lessons for new tasks. However, memory can also lead to memory-induced degradation, where an agent fails on tasks it would solve without memory, often because it follows a lesson that is incorrect, or misapplies or misreads a valid one. This is because lessons are inferred from trajectories labeled only with task success or failure, yet are typically provided without the trajectories they came from, leaving the agent no way to check them. To address this, we introduce WikiTrace, an evidence-linked memory system designed to let agents check each lesson against its original trajectory. Specifically, motivated by the finding that a trajectory often still holds the evidence needed to catch an erroneous lesson, WikiTrace stores each lesson as a concise insight linked to a summary of why its trajectory succeeded or failed and to the raw trajectory itself. Then, on a new task, the agent starts from all insights and, whenever one lacks details needed to act, conflicts with another, or comes from a failed task, descends through its links with terminal tools to fill in missing details or correct the insight before acting. We validate WikiTrace on five agentic benchmarks, showing that it consistently outperforms existing memory approaches and a baseline given the same insights alone, while also reducing memory-induced degradation. Our code is available at https://github.com/yukyeongleee/wikitrace
Multilingual routing benchmark tests language models across 12 languages and mixes
PolyRoute: A Twelve-Language and Code-Mixed Benchmark for Typed-Decision Routing, with Baselines and a Leaderboard
Abstract: Typed-decision encoders answer a question about a text in one forward pass, but their behaviour outside English is rarely measured on a fixed yardstick. We release PolyRoute: 322 native-language support messages in 12 languages and 9 writing systems, 156 code-mixed or romanised messages (Hinglish, Spanglish, Arabizi), 967 optional native training rows, and a protocol for a three-way department-routing decision. Labels come from Qwen3.8-27B with a blind second pass (91% of candidates kept); an independent model, gemma4-31b, agrees on 97.2% of test items, and the largest character 5-gram similarity between any test text and any training text is 0.33. A leaderboard of 13 systems gives 61.2% for the English-root checkpoint, 77.3% for the shipped multilingual one and 96.6% for the best fine-tune; TF-IDF with logistic regression reaches 91.9%. On four withheld languages the three multilingual systems score 77%, 90% and 93%. On code-mixed text the English root (85%) is nominally ahead of the shipped multilingual checkpoint (79%; p=0.134). The split resolves differences only above about 4 points. All texts and labels are LLM-generated.
EmoHarness improves small models understanding mixed emotional signals
EmoHarness: Boosting Multimodal Emotion Understanding with Meta-Harness Evolving
Abstract: Small multimodal models can misjudge emotion when words, voice, and visual cues conflict. We present EmoHarness, which runs such a model, the frozen student, inside a harness: the prompts, settings, and verification code around its calls. A human-written starting harness, the scaffold, numbers the student's visual observations, transcript, and prosody descriptions as citable evidence and has the student weigh candidate emotions against it. In Meta-Harness Evolving (MHE), a stronger LLM, the Builder, reads the student's errors and execution traces offline and edits the harness; static checks and validation gates decide which edits are kept, and the Builder never sees evaluation items. On 5,507 MME-Emotion items, EmoHarness raises task-macro Recognition by 7.6-13.4 points over a single video-only call for all four students and both Builders. Transcripts and prosody descriptions supply most of this boost (+5.38 to +10.71); the scaffold changes the audio-augmented call by -0.79 to +2.54, partly through extra multi-label predictions, and evolved edits add at most +0.95, although no evolved increment survives Holm correction. The harness's emotion context also transfers to IntentQA, with small gains for three of four students, none of which survives correction. EmoHarness thus boosts small multimodal students chiefly by supplying and organizing audio-derived evidence.
Adaptive meta-learning improves noise robustness in retrieval augmented generation
ARMOR: Adaptive Curriculum Meta-Learning for Noise-Robust RAG Reasoning
Abstract: Retrieval-Augmented Generation (RAG) systems have demonstrated remarkable effectiveness in mitigating hallucinations by incorporating external knowledge. However, the retrieval process inevitably introduces noise, posing significant challenges to RAG robustness. Fundamentally, noise robustness is a composite competency encompassing noise identification, retrieval filtering, and robust reasoning. Unfortunately, existing methods often fail to capture this holistic nature, typically restricted to single-dimensional enhancements or hampered by noise-induced optimization interference. To address these limitations, we decompose RAG robustness into three sub-tasks: knowledge boundary perception, retrieval verification, and robust reasoning. Accordingly, we propose ARMOR, an Adaptive Curriculum Meta-Learning for Noise Robust RAG Reasoning. Inspired by the teaching process, ARMOR introduces an adaptive task scheduler as a teacher to curate training plans according to the absolute learning progress of RAG generators (Student). Distinct from GRPO, ARMOR utilizes inter-group advantage estimation during learning to effectively mitigate mode collapse caused by limited output spaces. Experiments on three complex reasoning datasets (2WikiMultiHopQA, MusiQue, and HotPotQA) indicate that ARMOR achieves synergistic improvements in retrieval noise robustness, significantly outperforming baselines in output quality.
Rule based agent wins bargaining games without real time language model calls
Beating the Pool, Not the Game: A White-Box Equilibrium Agent for Language-Based Economic Games
Abstract: We describe sudsho, a rule-based agent for the GLEE Competition that plays bargaining, negotiation, and persuasion with zero inference-time LLM calls. Games are scored as payoff percentiles against an empirical pool of prior play on the same configuration, so pool percentiles set our demand targets while clamped equilibrium shares bound demands and acceptance. Every move is a deterministic function of the visible state (persuasion mixing aside), so alignment is verified by replaying the post-freeze move log through the policy. Across 117,280 games the agent reached an overall rating of 2056 at $0 inference cost, and the account, ranked by its best agent (a same-policy replica), placed 26th of 71 ranked accounts. A replay against three frozen snapshots of the opponent pool credits the pool-aware clamp with 274 to 330 points of mean per-game bargaining rating, an effect that holds where logged opponent decisions fix the replay.
Large language models edited with geometry aware knowledge protection
Editing Large Language Models with Geometry-Aware Regularization
Abstract: Knowledge editing has become a key technique for updating knowledge in large language models. However, existing methods neglect the geometric structure of the knowledge representation space, leading to two key limitations: they cannot distinguish which pre-trained knowledge should be preferentially preserved (core vs. non-core) nor which new edits are inherently difficult (hard vs. easy). To address this, we propose GEdit, a geometry-aware framework built on a Low-Rank Plus Diagonal (LR+D) factor model that captures the intrinsic low-rank structure of Transformer FFN hidden representations. GEdit features two complementary mechanisms. First, the low-rank regularization applies targeted protection on a low-rank core knowledge subspace. Second, the adaptive regularization dynamically adjusts the regularization strength based on each edit request's editability: relaxing constraints for hard edits to ensure success, and tightening them for easy edits to better preserve pretrained knowledge. Extensive experiments on GPT2-XL, GPT-J, and LLaMA-3-8B under large-scale editing scenarios, including both mini-batch and fully sequential editing settings with up to 10,000 continuous updates, show that GEdit consistently and significantly outperforms state-of-the-art baselines in both editing success and knowledge preservation.
ReplayRCA measures how well models fix agent failures live and after the fact
ReplayRCA: Measuring Post-hoc Diagnosis and Live Oversight of Agent Failures by Counterfactual Replay
Abstract: When a language-model agent fails a multi-step task, can another model find the decisive mistake and fix it with a single action? We introduce ReplayRCA, a benchmark that scores such diagnoses by intervention rather than by agreement with annotations: the environment is restored to the state preceding the chosen step, the proposed action replaces the agent's own, and the same agent continues. The score, causal repair, is the continuation success rate minus that of a paired placebo that replays the agent's original action through the same mechanism. Two tracks share this protocol: a post-hoc autopsy, where the diagnoser sees the whole failed run and is told that it failed, and live oversight, where a supervisor watches the run unfold and may intervene once, before an irreversible action, without knowing whether the run will fail. Tasks come from OpsWorld, a deterministic, procedurally generated back-office environment with six difficulty levels, the last of which is multimodal (QR codes, handwriting). On 62 calibrated systematic failures, the best of 13 frontier models (Grok 4.7) reaches a causal repair of +51.3 points; the top eight models are statistically indistinguishable, the flagship models of the two largest vendors rank 8th and 12th, and 5 items are repaired by no model. Forcing a strong model to intervene at the last step costs it 17.1 points, so the last-step shortcut of annotation-based attribution benchmarks does not transfer. On 87 runs, the two best live supervisors are statistically tied and one of them is a small open-weight model (DeepSeek V4.1 Flash); the supervisor that is among the best post-hoc diagnosers (GPT-6.1 Sol) is the weakest live one (-14.5 points on the same runs, p = 0.002), so post-hoc diagnostic performance can be a poor proxy for live oversight. The environment and the evaluation code are released at https://github.com/jeremy1392/RCAreplay.
Framework reveals patterns in large language model agent actions
Behaviour analysis for large language model-driven agents
Abstract: Understanding language-agent behaviour requires learning how actions are organised across a trajectory and how this organisation relates to context and outcomes. We propose a general framework that connects data-driven behavioural discovery with analysis and experimental investigation. Categories developed from recorded activity define what is measured. Analyses preserve the behaviours' positions in the process and express their relationships as interpretable hypotheses. Statistical assessment and tests in held-out contexts examine the strength and scope of the evidence. Experiments, including prompt steering, then test how behaviour and outcomes change under deliberately varied conditions. A study of iterative self-reflection illustrates how discovery, dependency learning and prompt comparisons can inform one another. We develop this approach for trajectories of varying length, event order and agent roles, and identify advances in coding, analysis and experimentation needed to support its wider use.
Pos signatures improve demonstration choice for dependency parsing
POS-Signature Demonstration Retrieval for In-Context Dependency Parsing
Abstract: Demonstrations for in-context learning are often retrieved by semantic similarity, which may be a poor proxy for relevance in structured prediction. We study demonstration selection for Universal Dependencies parsing with a POS-signature policy that represents sentences with weighted part-of-speech statistics and retrieves three length-compatible, gold-parsed examples from the labeled training split. We evaluate two open-weight language models without fine-tuning on all 2,077 sentences of the UD English Web Treebank test set, comparing the policy with shot- and format-matched fixed and semantic three-shot baselines, as well as zero-shot, chain-of-thought, and critique-refinement systems. POS-signature retrieval yields the highest labeled attachment score among the tested prompting systems for both models (72.4 and 73.2), outperforming semantic retrieval by 1.9 and 2.7 points; paired-bootstrap 95% confidence intervals exclude zero for both contrasts. Chain-of-thought and refinement comparisons vary across models. These exploratory results favor the tested POS-signature retrieval approach on English-EWT. We release the prompts, archived outputs, and evaluation code.
Gui agent improves task success by reusing subgoal memories
Subgoal-aware Memory Graph for GUI Agent
Abstract: GUI agents aim to automate complex user tasks by interacting with mobile interfaces, but ef- fective decision-making in long-horizon set- tings requires reusing prior experience rather than relying solely on the environment ob- servations. Existing memory-based methods typically store experience either at the single- action level or as entire trajectories. Single- step actions are often too fine-grained to carry meaningful semantics, since the success of an individual action does not necessarily imply progress towards the overall goal. In contrast, whole trajectories are often too coarse-grained and difficult to reuse across different tasks. We propose a subgoal-aware memory graph that distils past trajectories into structured, reusable evidence at the level of subgoals. We define a subgoal as consecutive actions shar- ing a single intent, the granularity at which progress toward the task becomes meaning- ful. To reduce redundancy and support cross- task reuse, we organise subgoal experiences as nodes in a graph, where related memories can be merged, extended, and retrieved based on the environment observation. On SPA-Bench, our method improves success rate by 11.11% over the baseline.
Generative reward models improve rationale quality and reduce bias
Judging the Justification: Decision Certificate Supervision for Generative Reward Models
Abstract: Generative reward models (GenRMs) generate natural language rationales to justify preference judgments, yet a correct verdict does not guarantee a sound justification. Existing approaches often rely on reference explanations to supervise rationales, but obtaining such references can be costly and does not necessarily establish whether generated rationales support the resulting verdict. To address these limitations, we propose DeCert, a framework that evaluates the validity and verdict consistency of generated rationales without requiring reference explanations. Crucially, DeCert translates these assessments into preference supervision that accounts for both verdict correctness and justification quality. In particular, when multiple sampled judgments share the same correct verdict, differences in certificate quality enable DeCert to construct rationale preference pairs that distinguish stronger from weaker justifications. Extensive empirical studies across two model families and multiple in- and out-of-distribution benchmarks demonstrate that DeCert consistently outperforms strong generative judge baselines while improving rationale quality. Moreover, DeCert effectively mitigates positional bias, improving position-consistent accuracy by up to 14.8 points. These results highlight the importance of supervising not only which verdict is correct, but also how that verdict is justified.
System one models continuously monitor and guide system two reasoning
SOMA: Continuous Monitoring and Intervention for System Two Reasoning with System One Models
Abstract: Reasoning models often run into problems without noticing them, even when symptoms are visible in their own text. In humans, metacognition separates monitoring from control: fast, intuitive signals keep watch while slow, deliberate reasoning unfolds and help decide whether to continue, change course, or stop. System One (S1) decision models, such as Jev, are a natural candidate for this role: they return a calibrated probability for a natural-language question about a text in one fast call, yet they are used only as gates before reasoning. We introduce SOMA (System One Metacognitive Awareness), in which an S1 model continuously monitors a System Two (S2) reasoning model and intervenes when needed. SOMA has three layers: a scenario layer states each objective in natural language with predefined actions, an awareness layer lets S1 judge whether the described behavior appears, and a control layer accumulates these judgments, first reminds S2, and takes over only if the behavior persists. Sharing one S1 model and one controller across tasks, without training, SOMA addresses two different failures on two reasoning models: on overthinking, it cuts reasoning tokens by up to 36\% while improving average accuracy, and on unanswerable questions, it raises abstention from 70.5\% to 87.2\%.
Tool using agents improved to handle weaker user instructions better
Not Every User Says It All: Training Tool-Using Agents Beyond Ideal Simulated Users
Abstract: Tool-using agents are trained and evaluated almost exclusively with strong simulated users who state every detail clearly, yet real users are often weaker. We construct weak users by masking sentences of a simulated user's replies, holding the task and the user's knowledge fixed, and cross-evaluate agents trained with strong and weak users on tau2-bench and ToolWOZ. Weak users lower the success of every policy, and training with them does not help: no policy trained with weak users, through fixed masks, mixtures or a curriculum, serves weak users better than the strong-trained policy beyond run-to-run noise, and prompting the agent to ask more does not help either. We propose StaGPO (State-Grouped Policy Optimization): the agent writes what the task needs, what is missing and how it is being filled, and GRPO uses this state to compare turns of different rollouts that reach the same situation, assigning step-level credit. Within the standard recipe of a small SFT cold start followed by RL, StaGPO makes weak-user training useful: on ToolWOZ, the Qwen3.5-2B policy trained with weak users reaches 76.8% with weak users, the highest of any Qwen3.5-2B policy we trained, 4.8 points above GRPO and 2.8 above the strongest baseline; it is also 5.4 points above GRPO with strong users, and generalises to an unseen mask rate; trained with strong users instead, it serves weak users at least as well as GRPO trained with weak users. On tau2-bench it scores highest with weak users for both models, by a small margin for Qwen3.5-4B.
Splitting test-time compute optimizes model answer verification accuracy
Sample, Mix, or Verify? Allocating Test-Time Compute under Imperfect Verifiers
Abstract: A common way to make language models and agents more accurate is to spend more compute at inference time: generate many candidate solutions and submit the one a verifier prefers. Success then requires that some candidate is correct and that the verifier picks it. The first grows predictably with the number of samples; the second does not, because real verifiers (unit tests, majority votes, LLM judges, reward models) accept wrong answers, reject right ones, and make correlated mistakes. With a fixed budget, a practitioner must decide how many samples to draw, which models to draw them from, and how much to spend on verification, and today these choices are made by rules of thumb. We introduce SPLIT, which makes them from a small labelled pilot. It rests on an exact formula for the probability of submitting a correct answer under accept/reject and top-score selection that accounts for varying task difficulty, blind spots shared across models, and correlated repeated verifier calls. SPLIT fits the formula by empirical Bayes and uses it to plan budgets on unseen tasks. Across seven benchmarks spanning math, science QA, code and software-engineering agents, 26 models, 175 agent systems and six kinds of real verifiers, all from public releases, SPLIT lands within 0.02-0.8 percentage points of the best allocation in hindsight, while every fixed rule we test loses at least 2.5 points in some setting. Regret depends only on how well SPLIT ranks configurations, and it stays near-optimal under compute, price and latency costs and online. The formula also explains when asking more LLM judges lowers precision. As a forecaster far beyond the pilot, SPLIT beats power-law fitting when the verifier caps success, and we report where it loses.
Graph method reveals where AI agents act less consistently
GraphACM: Graph-based Action Consistency Measurement for Uncertainty Quantification in LLM Agents
Abstract: Behavioral consistency across repeated executions provides an informative signal for assessing the reliability of large language model (LLM) agents. However, existing approaches typically summarize consistency with a single task-level score, without identifying where behavioral differences occur or whether actions are compared under equivalent decision contexts. We propose a graph-based framework for fine-grained behavioral consistency measurement and uncertainty quantification. Given multiple rollouts of the same task, we merge equivalent states into a directed, action-labeled graph, enabling action comparisons at shared decision states. We introduce node-level and global measures of action consistency, using the global consistency score as a proxy for uncertainty to predict task success or failure. Experiments across three LLMs on ALFWorld, AppWorld, and WebShop demonstrate that our measure is positively associated with task success and achieves an average AUROC of 0.811 for task-success prediction, outperforming the strongest behavioral-consistency baseline (0.729) and verbalized confidence (0.769). These results highlight the effectiveness of state-conditioned behavioral consistency and pinpoint where agent actions agree.
Large language model agents share experience to improve task success
XferAgent: Cross-Agent Experience Transfer for Self-Improving LLM Agents
Abstract: Large language model (LLM) agents can reuse experience from prior task executions to improve performance on subsequent tasks. Although existing methods enable cross-agent experience sharing, differences in agents' execution histories and task requirements make shared experiences difficult to reuse effectively. We propose XferAgent, a framework for cross-agent experience transfer that organizes reusable experience in local and shared experience graphs. XferAgent reconciles compatible contributions from heterogeneous agents while avoiding unreliable merges. For each new task, it selects shared experiences and integrates them with the target agent's local experience to support task execution. Experiments across diverse task domains show improvements over representative memory-based methods and effective experience transfer between heterogeneous agent frameworks. Cumulative task success also increases over successive task passes as agents update and reuse their experiences.
LLM agents struggle with user emotions in long tasks
$τ$-Emotion-Bench: Exposing the Interaction Fragility of LLM Agents through Real-World User Emotions
Abstract: In long-horizon tool-calling tasks, LLM agents can fail not only because of tool-use errors, but also because of ineffective interaction with users. Yet existing benchmarks largely assume cooperative and behaviorally stable users, leaving the effect of emotional variation on agent reliability underexplored. We introduce $τ$-Emotion-Bench, a ten-domain benchmark for evaluating tool-calling agents under controlled emotion-conditioned user interactions while preserving the underlying task objective. To scale both benchmark construction and agent training, we introduce Tracer, a scalable agentic data-generation pipeline that constructs grounded, policy-valid tasks and collects trajectories through an interactive agent--user--environment harness. By allowing dissatisfied users to terminate inefficient interactions, the harness provides an interaction-aware signal that filters trajectories with excessive user dependence and favors effective tool use. Experiments with two user simulators show that frontier agents remain unreliable, with substantial gaps between average task success and consistent task completion. Tracer-generated supervision consistently improves Qwen3 models across scales, generalizes to unseen domains, and transfers to $τ$-Bench and $τ$-Trait. Behavioral analyses further show that higher task success is associated with a better balance between tool execution and necessary user follow-up. Overall, $τ$-Emotion-Bench and Tracer provide a scalable framework for evaluating agent reliability under diverse user behaviors and improving tool-calling agents through interaction-aware supervision.
Large language models adjust creativity based on conversations context
Read the Room Before Your Model Generates
Abstract: Large language models (LLMs) generally rely on sampling temperature to regulate the amount of randomness in their outputs. Mainstream methods for selecting temperature typically involve fixed, heuristic global settings and context or token-level calibration, rather than a geometrically interpretable representation of the communicative scenario. In this work, we introduce a simple and interpretable temperature selection framework, the \textbf{Scenario-Actor}, which adapts decoding to the inferred situational characteristics. Scenario-Actor extracts scenario-relevant structures from Transformer hidden states and determines the temperature from scenario embedding, a structured geometric representation of the communicative setting. Empirical results show that on Llama-3.1-8B-Instruct, Scenario-Actor achieves strong goal-completion performance in interactive social scenarios, significant gains on open-ended instruction following, and competitive performance on multi-turn conversation compared to multiple baselines.
Agent reflex improves language agents reflection and recovery in noisy settings
Agent-Reflex: Teaching Language Agents to Act Reflectively
Abstract: Large language models have evolved from single-turn chatbots into interactive agents capable of long-horizon reasoning. In real deployments, MCP tool services can return noisy feedback and users may intervene to correct the agent, but existing methods provide limited supervision for reflection learning and failure recovery. To bridge this gap, we propose Agent-Reflex, a general training framework for incentivizing proactive reflection in agents. Specifically, Agent-Reflex comprises three stages. First, Environment Reflection Skill Discovery (RSD) injects grounded tool-feedback perturbations into environment scaling, samples a capability-diverse LLM pool, mines matched success-failure pairs, and summarizes reusable reflection meta-skills. Building on this skill library, we then apply Reflective Supervised Fine-Tuning (ReSFT) that converts each skill into supervision for self-reflection, skill recovery, and user-centric correction, strengthening complementary reflective behaviors. Finally, we propose On-Policy Reflective Policy Optimization (ORPO) to combine agentic reinforcement learning under noisy interactions with reflection-guided on-policy distillation to further elicit proactive reflection. Built on Qwen3-8B, Agent-Reflex reaches 58.6 overall on five long-horizon tool-use benchmarks and 29.1 on AgentNoiseBench, surpassing strong environment-scaling and reflection baselines. Experiments on six challenging agent benchmark suites demonstrate the effectiveness of Agent-Reflex. Further analyses show that Agent-Reflex improves agents' robustness to noise while maintaining stable training dynamics, supporting scalable agent training. Our code is available at https://github.com/RUC-NLPIR/Agent-Reflex.
PaceReasoner boosts short answers accuracy using long reasoning
PaceReasoner: Label-Free Learning to Answer Briefly from Long Reasoning
Abstract: Extended chain-of-thought reasoning improves problem-solving performance but increases inference cost and latency, motivating more accurate answers with shorter reasoning. We introduce PaceReasoner, a label-free test-time reinforcement learning framework that uses a model's LONG responses to improve its SHORT-mode accuracy. Both modes share a trainable policy, creating a need for reliable supervision, sufficient SHORT training samples, and continued LONG responses to guide learning. PaceReasoner combines all-response and LONG-only answer consensus through progressive, confidence-weighted supervision. A behavior reward gradually shifts from encouraging SHORT exploration to rewarding agreement with LONG answers, while a ratio controller regulates the balance between modes. Across three model backbones and four benchmarks, PaceReasoner improves average SHORT accuracy over the baselines by 7.4--13.5 percentage points, with fewer than 25 reasoning tokens per response on average. Ablations support the key components, and cross-dataset results demonstrate transfer beyond the adaptation set.
Visual language models learn better by using their own mistakes
The Value of Being Wrong: Self-Mined Visual In-Context Learning from Errors
Abstract: Visual in-context learning (ICL) enables vision-language models (VLMs) to adapt at test time via demonstration examples, but its effectiveness depends critically on the examples selected. Recent counterfactual selection methods improve ICL by constructing demonstrations that expose how individual attribute changes affect the answer. However, they require multi-stage pipelines involving attribute extraction, caption engineering, and composed image matching. We propose a simpler alternative grounded in a key observation: a model's own prediction errors are natural counterfactual demonstrations that directly encode its failure modes. Our method, SMILE (Self-Mined Visual In-Context Learning from Errors), builds a model-specific pool of hard negatives offline by recording the target VLM's own mistakes. At test time, the prediction serves as a direction signal that surfaces, in the semantic embedding space, the past mistakes most likely to recur for the current query. This prediction-conditioned identification provides a unified mechanism for both closed-set classification and open-ended visual question answering without task-specific design. Across four benchmarks and four VLMs, SMILE delivers an average task-score gain of +3.09 points over the SOTA baseline at 5x lower latency. Project page: https://tree-shu-zhao.github.io/SMILE/
Pid gates improve prompt updates under noisy feedback
PID-Gated Continual Prompt Adaptation Under Noisy Feedback
Abstract: Deploying LLM agents in non-stationary environments requires continually revising their instructions as the task drifts, yet rewriting after every reported mistake can turn noisy feedback into persistent errors. This raises an often-overlooked scheduling question in prompt adaptation: when should feedback trigger a lasting instruction update, and at what editor-call cost? Taking inspiration from proportional-integral-derivative (PID) control, we introduce a family of lightweight gates that react to the current error, accumulate evidence over time, and roll back deterioration after an edit. We benchmark these gates against both simple baselines and advanced prompt optimizers (TextGrad and GEPA), over a diverse suite of drifting tasks, LLM parameter sizes, feedback-noise levels, and random seeds. PID-gated continual prompt adaptation yields a more favorable Pareto trade-off between task performance (high accuracy, low error) and editor-call budget than the baselines considered, pointing to control-theoretic scheduling as a simple, general tool for robust prompt adaptation under unreliable feedback.
Benchmark shows agents struggle with real business task delivery
FDE-Bench: Evaluating End-to-End Delivery from Underspecified Real-World Business Requests
Abstract: Real-world business requests often leave key requirements unstated because frontline practitioners struggle to articulate rules they know from experience. Forward-deployed engineers (FDEs) must elicit these requirements through dialogue and iteratively optimize solutions until they meet customer acceptance criteria. Existing optimization benchmarks typically start from explicit objectives, while clarification benchmarks often introduce information gaps into existing tasks. We introduce FDE-Bench, the first benchmark built from real customer projects to evaluate agents on end-to-end delivery tasks that require interactive requirement clarification and iterative optimization. Its 49 cases span combinatorial optimization and machine learning. Clarification targets are distilled from discussions between frontline practitioners and FDEs. Agents elicit missing requirements from a simulated customer, and case-specific evaluators assess final artifacts against acceptance criteria using customer-accepted FDE solutions as quality references. Across 18 agent configurations, the highest mean delivery score is only 0.5510, compared with an FDE reference of 1.0. The highest delivery success rate (DSR), defined as the fraction of planned runs reaching this reference, is only 6.12%. These results show that reliable FDE-level end-to-end delivery remains out of reach for the evaluated agents on underspecified real-world business tasks. The benchmark project page is available at https://baidubce.github.io/FM-Agent/FDE-Bench/.
Latent world model improves gui agent speed and success rates
Latent World Action Model for GUI Agents
Abstract: Pixel-space world models have recently emerged as a promising approach for GUI agents, predicting future screens prior to execution and subsequently re-perceiving them for action selection. However, this two-stage paradigm faces two key challenges in GUI environments: (L1) the high computational cost of pixel reconstruction and (L2) the need to reconstruct large and diverse appearance-level details beyond what is required for action selection. In practice, this screen-generation objective can induce an identity shortcut problem, where future-state prediction largely preserves the current state rather than capturing action-induced changes. In this work, we reformulate GUI world modeling by shifting the prediction target from full-screen reconstruction to action-induced future information in latent space. Concretely, we leverage a large multimodal model as the latent world model, whose shared representations support two dedicated heads for both action selection and predictive transition modeling. The action head directly translates these latent representations into action selection, while the transition head captures action-conditioned future information. To mitigate the identity shortcut, we further introduce an action-aware contrastive objective that encourages the transition representations to capture discriminative action-induced changes. Extensive experiments across offline and online GUI settings show that our method achieves up to 31.9% and 40.6% relative gains in success rate over pixel-space world models, respectively, while substantially reducing inference latency by up to 98.9%. Our code is available at https://github.com/sjyu001/latent2world.
Evaluating how AI agents admit uncertainty when facing conflicting knowledge
Accurate but Not Humble: Evaluating Epistemic Humility in LLM Agents under Knowledge Conflict
Abstract: When retrieved evidence contradicts an agent's prior beliefs, does it revise its answer, acknowledge uncertainty, or persist with an incorrect conclusion? Existing evaluations of agentic systems focus primarily on task success, offering limited insight into how agents handle such conflicts. We propose to evaluate agents on epistemic humility (EH): the agent's willingness to recognize, act on, and communicate uncertainty during task execution. We operationalize EH through three trajectory-level behavioral dimensions: Identify, Solve, and Escalate (ISE). Through knowledge conflict, situations where the backbone language model's parametric knowledge contradicts the evidence it encounters, or where two contextual sources disagree, we evaluate two conflict settings: (1) controlled conflict and (2) naturally occurring conflict during multi-step agentic execution, each paired with matched no-conflict controls. Evaluating four agents, we find that higher task accuracy does not necessarily correspond to greater epistemic humility: some high-accuracy configurations recognize conflicts during execution but do not communicate unresolved uncertainty in their incorrect final answers. Trajectory-level analysis further reveals that agents frequently detect conflicts in early steps of execution but fail to maintain or resolve them in later steps. Finally, we show that model-level interventions can improve EH, but often at the cost of task accuracy, suggesting that epistemic humility emerges from the interaction among the backbone model, agent harness, and evaluation environment.
Skill reranking improves with difference-focused text selection
SkillContrast: Difference-Guided Text Selection for Agent Skill Reranking
Abstract: Similar agent skills can share instructions but differ in their conditions of use. Query-based text selection may retain shared instructions and omit these distinctions. We introduce SkillContrast, a training-free selector that compares retrieved skills and retains their differing text with local context for a pretrained reranker. On 1,235 requests from SameCapRisk-Bench, it yields 54-72 more clean hits (requests that retrieve a helpful skill without its marked risky sibling) than TF-IDF query selection at identical per-candidate input lengths, across 2 retrievers and 2 reranker sizes. Length-matched component replacements identify differing text as the main contributor in the primary setting, with smaller, mixed context effects. Relative to full skill bodies, SkillContrast uses 51.1-58.8% fewer model-input tokens, with 10-18 fewer clean hits at 0.6B and matching or higher observed clean-hit counts at 4B. Candidate-relative differences thus complement query relevance in selecting compact reranking inputs.
Multi-agent ai systems improve negotiation reliability and deadlock recovery
Constitutional Gating and Deterministic Recovery for Multi-Agent LLM Negotiation: Ablations Against a Stateful Adversarial Gatekeeper
Abstract: Multi-agent LLM systems negotiating with a stateful counterpart waste model calls in three ways: polite loops that never meet the counterpart's hidden acceptance condition, malformed outputs that trigger retries, and compliance deadlocks in which the counterpart demands something the agent must refuse. We study a three-part control stack - a 5-Pillar runtime constitution, a 4-tier swarm (Director, three-agent majority vote, Monitor, schema hard gate) and Cognitive Annealing (deterministic deadlock detection, atomic purge of the agent-side context, a canonical recovery message) - against a released adversarial Gatekeeper whose acceptance rules are fixed regular expressions and whose LLM only renders reply text. The testbed has a known solution: it measures whether the stack executes a constitution-aligned strategy against swarm drift and recovers from deadlock, not whether it discovers anything. In five runs per configuration (30 runs; Gemini 2.5 Pro agents, Claude Haiku 4.5 Gatekeeper) we find: (i) the constitution and Director make an acceptable framing possible but not reliable - 0/5 baseline unlocks versus 1/5 and 2/5 with the constitution; when the swarm unlocks it does so in one turn with 7-8 calls and about 15k tokens (67-73% below baseline); when it does not, it costs 17-38% more; (ii) the Monitor and hard gate do not reduce unlocks and leave an audit trail; (iii) under a honeytrap-to-compliance deadlock, LLM-only steering escapes 0 of 5 times while atomic purge plus a canonical strike escapes 5 of 5 (Fisher $p = 0.008$) at the same call budget, with zero calls for the strike. LLM-written strikes failed the deterministic pre-flight 5 of 5 times although an LLM Monitor had approved four. Pre-registered hypotheses on average call and token reduction were not supported. Cost is bounded in every arm by deterministic stop rules; the stack adds recovery at no extra model cost.
Latent space evaluation cuts AI judgment time by six times
Judging in Latent Space: Efficient Generative Reward Modeling via Semantics-Preserving Compression
Abstract: Reward modeling often requires jointly representing and reasoning over multiple evaluation criteria, yet verbalizing this process token by token can incur substantial inference cost. Recent work on latent reasoning suggests that continuous states may support this computation more compactly. We introduce LatentGRM, a latent evaluation framework built on semantic chunking, compression, and reconstruction. By using the structure of rubric-guided evaluations to guide compression, LatentGRM learns compact continuous trajectories that support autonomous pairwise judgments without generating textual assessments. A separate interpreter reconstructs evaluation text from these trajectories, providing an offline view of the information retained under compression. Under matched training data and backbones, LatentGRM achieves competitive aggregate preference accuracy relative to explicit Supervised Fine-Tuning (SFT) judges at both 4B and 8B scales. Across four benchmark domains, LatentGRM-8B compresses evaluation trajectories by 8.9--9.2x and reduces total judge inference time by 6.1--7.0x at vote@5. Controlled rubric interventions show that criterion-dependent preference information is carried through the latent sequence. Together, these results demonstrate that continuous latent evaluation can substantially reduce inference cost while preserving competitive judgment quality.
SearchJev speeds up search agents with better calibrated fast decisions
SearchJev: A Fast and Calibrated System-1 Model for Search Agents
Abstract: Search agents repeatedly make short decisions about relevance, evidence sufficiency, and search actions. Using generative language models for these decisions introduces latency and unreliable confidence. We present SearchJev, a fast and calibrated System-1 model that separates search decisions from System-2 reasoning and generation. Given a search state and a decision schema, SearchJev directly scores legal options without autoregressive output generation. We propose Soft-Label Learning for Calibrated Decisions (SLCD) to learn decision probabilities from uncertain supervision and calibrate their confidence. In a dual-system search agent, SearchJev handles short decisions and delegates uncertain judgments to System 2, which retains planning, query generation, and answer composition. We also introduce SearchDecision-Bench, a benchmark unifying six types of search decisions for training and evaluation. On SearchDecision-Bench, SEARCHJEV improves decision quality over same-size Qwen3.5 autoregressive models, achieves 5.2-5.3 times faster decisions, and reduces average expected calibration error by 41-74%. On BrowseComp-Plus, the dual-system agents achieve a 3.7-4.7 times speedup in active search time while improving answer accuracy from 45% to up to 54%.
Vision based web agent improves accuracy on live websites interactions
WebFovea: When the Model Is Right but the Click Is Wrong -- Reliable Round Trips for Vision-Based Web Agents on Live Websites
Abstract: We present WebFovea, a vision-based web agent that placed 2nd in the WebRetriever Challenge 2026 with a final score of 57.0 out of 100. The challenge evaluates agents end to end on Protocol III of the WebRetriever benchmark (arXiv:2607.06118): starting from an entry URL on a live website, the agent must operate the site's own interface and return a verifiable answer. A capable multimodal large language model (LLM) is necessary for this, but not sufficient. The model's decisions reach the browser through the harness, the code between the model and the page. At every step, four things must go right: the model's reply must be parsed into the intended action, the action must take effect on the page, the result must be reported back accurately, and the model must be shown the information it needs. On real websites, many of the failures we observed occurred at one of these four stages rather than in the model's reasoning. A coordinate-space mismatch placed every click at 3/4 of its intended coordinates; actions on native dropdowns, inside iframes, and in text boxes failed silently; and self-generated chat-template tokens contaminated 4.9% of task episodes. WebFovea hardens each stage and surrounds the loop with guardrails that keep the agent within the rules and its budget. The four-stage view does not depend on the model, although some individual fixes do. Because we used the same model in all four submissions, the rise of our official hidden-set score from 31.0 to 57.0 reflects changes to the harness, up to run-to-run variance on live sites. We describe the design, the evidence for each component (including negative results), a failure analysis, the limitations, and a roadmap that includes routing different steps to different models.
Agentic text to sql systems improve question handling to reduce hidden errors
Beyond Correctness: Resolving Underspecification in Agentic Text-to-SQL
Abstract: Agentic Text-to-SQL systems can interact with users to clarify underspecified queries before generating SQL. However, a correct execution result does not necessarily imply that the agent has adequately resolved the underlying underspecification: the agent may silently make unverified assumptions that happen to match the intended answer. We show that this behavior is driven in part by premature clarification termination. Although forcing an agent to ask more questions improves execution accuracy, ambiguities are concentrated in earlier interactions, making brute-force questioning inefficient. More importantly, even when explicitly prompted to plan its clarification process, the agent frequently abandons questions that it has already identified as relevant. To address this failure mode, we introduce PlanPool, which externalizes the clarification plan as a mutable question pool. Every planned question must be explicitly asked or dropped before submission, while newly discovered ambiguities can be added during interaction. Across three benchmarks derived from BIRD-Interact and Spider, PlanPool consistently improves ambiguity coverage and reduces silent failures over unconstrained and prompt-based alternatives, while maintaining competitive execution accuracy. Our results highlight an important distinction in agentic reasoning: identifying missing information is not sufficient, and the agent must also reliably maintain and resolve it before committing to an answer.
Agent control cuts memory use for assistant queries by 92 percent
APDMem: Agent-Controlled Progressive Disclosure for Query-Adaptive Long-Term Memory
Abstract: Personalized LLM assistants must recover sparse evidence from long conversation histories across queries of varying complexity. We introduce APDMem (Agent-controlled Progressive Disclosure Memory), a hierarchical long-term memory architecture that applies progressive disclosure to memory retrieval. Rather than relying on a flat memory store or fixed retrieval granularity, APDMem represents conversation history as four progressively detailed layers: thematic summaries, personalized key facts, turn-level evidence notes, and raw messages. At inference time, a controller applies progressive disclosure to the memory hierarchy: it first reads high-level summaries and drills into finer evidence only when needed. This creates an adaptive cost-fidelity trade-off: simple queries can terminate early, while complex temporal, multi-hop, or exact-evidence queries trigger deeper inspection. A note synthesizer converts retrieved evidence into a query-focused structure that consolidates facts, orders events, and flags contradictions before final answer generation. Experiments on LongMemEval show that APDMem achieves strong performance for long-context memory reasoning while accessing only 8% of the total conversations.
Bayesian method improves confidence and trust in multiple AI models’ decisions
Counting Moves, Weighing Voices: Bayesian Dialectical Argumentation for Calibrated Multi-LLM Councils under Persistent Adversaries
Abstract: A multi-LLM \emph{council} lets several large language models (LLMs) deliberate on a question and return an answer together with a confidence estimate. As these systems become increasingly used for reasoning, that confidence should represent a calibrated \emph{probability of being correct}, and the decision should remain robust when some agents are persistently unreliable. Existing \emph{council aggregation} methods fail on both fronts: their confidence estimates measure decisiveness rather than correctness, and they cannot identify or discount persistently unreliable agents. We introduce Bayesian Dialectical Argumentation (BDA), which treats the council's \emph{typed} moves---who proposed, challenged, or conceded which answer---as observations of a classical annotator model with \emph{per-agent} reliabilities. This formulation recasts multi-agent deliberation as a reliability estimation problem, using the deliberation trace to infer agent reliability under persistent adversarial behavior. By weighting evidence according to inferred agent reliability, BDA yields calibrated posterior probabilities over candidate answers while allowing persistently unreliable agents to be inverted rather than merely outvoted. Across binary and multi-class benchmarks, BDA achieves the best calibration among zero-cost council aggregation methods, requiring no additional LLM calls, and improves robustness under persistent adversarial coalitions while remaining competitive in clean settings.
RAGScope controls risk and cuts cost in AI answer review processes
RAGScope: A Leakage-Controlled, Cost-Aware Evidence-Gating Protocol for RAG Hallucination Triage
Abstract: Retrieval-augmented generation (RAG) systems need inexpensive ways to route generated answers: accept low-risk outputs, review uncertain ones, and reserve strong verifiers for the expensive tail. We present RAGScope, a leakage-controlled protocol for evaluating local evidence gates that use only the task input, retrieved context, and answer text. The protocol combines context-grouped splits, fold-scoped preprocessing, group bootstrap intervals, deployment operating points, end-to-end runtime, and explicit source-shift stress tests. On three RAGTruth tasks, the enhanced gate RAGScope-E reaches 0.798 AUROC and 0.660 average precision (AP) in pooled grouped cross-validation. Its pooled AP exceeds ROUGE-L by 0.034 with a 95% context-group interval of [0.002, 0.064], although the AUROC gain is not significant and ROUGE-L remains stronger on data-to-text. At a top-10% review budget, RAGScope-E attains 0.748 precision; accepting the lowest-risk 50% yields 0.141 residual unfaithfulness. RAGScope-E runs in 6.22 ms/example on CPU, versus 145.75 and 223.07 ms/example for the tested DeBERTa-NLI and HHEM settings. A 14,900-example HaluBench stress test exposes the deployment boundary: an in-domain calibrated gate reaches 0.879 AUROC, but leave-source-out calibration averages only 0.466. Target-only calibration recovers to 0.675 AUROC with 100 labels per source and 0.685 with 200. Cheap evidence gates are therefore useful routing components, but learned calibration must be validated and adapted within the target domain.
Hierarchical reinforcement learning improves coding agent efficiency
Doing More with Less Tokens: Hierarchical Reinforcement Learning for Efficient Coding Agents
Abstract: Recently, coding agents have emerged as a dominant paradigm for real-world software engineering (SWE) scenarios, which solve complex tasks through multi-turn interactions with development environments. However, frequent interactions with environments would inevitably introduce substantial token overhead, leading to high usage costs and latency. Although recent studies have explored reducing token usage by context manipulation and interaction limits at inference time, these approaches focus on improving token efficiency while overlooking the risk of discarding task-relevant information, thus struggling to balance the trade-off between resolution rate and token efficiency. In this paper, we study a more general paradigm without suffering from the limitation, i.e., training token-efficient coding agents with promising resolution performance, which is a highly-practical yet less-explored problem. To this end, we reveal two core observations in SWE scenarios: i) Efficiency Variation: successful resolution could be achieved with fewer tokens; ii) Entropy Correlation: unproductive behaviors are associated with turn-level entropy. Motivated by observations, we propose a novel reinforcement learning framework, dubbed HERO. Specifically, HERO prioritizes task resolution over token efficiency during policy optimization and encourages efficient reasoning patterns at both trajectory and turn levels. Extensive experiments on SWE-bench Verified and SWE-bench Multilingual demonstrate that HERO achieves a favorable trade-off between resolution rate and token efficiency compared with state-of-the-art coding agents and reinforcement learning methods.
Llm agents improve understanding despite user miscommunication and changing goals
Beyond Oracle Communication: Benchmarking Interactive Intent Alignment Under Miscommunication and Evolving User Intent
Abstract: Modern LLM agents increasingly tackle complex tasks through interactive, long-horizon exchanges with users, while existing benchmarks generally assume that users always accurately and sufficiently communicate a fixed intent. However, this oracle communication assumption rarely holds in practice: users may miscommunicate, change their goals, and run out of patience. We define this task setting as Interactive Intent Alignment, where agents must recover and continuously track the user's current intent despite imperfect communication and evolving goals. To study this setting, we introduce Drift-Bench++, a principled benchmark construction pipeline for verified executable tasks with controlled misalignment and intent shifts, along with an interaction protocol featuring finite patience, diverse simulated users, and silent interaction-conditioned shifts. We further develop GRIP, a comprehensive evaluation protocol covering task grounding, user realism, inquiry effectiveness, and adaptation to evolving intent. Across diverse environments, models, and interaction conditions, stronger interaction consistently helps but remains far from oracle performance; Validation on deployed ProdAgent sessions further shows that the modeled failures are prevalent and consequential in deployment. By providing a unified, executable benchmark for interactive intent alignment, Drift-Bench++ offers a foundation for evaluating and advancing agents under realistic communication and evolving intent.
Advisors learn better advice for language models using self-distillation
AdviSD: Learning to Advise Frontier LLMs via Targeted Multi-Turn Self-Distillation
Abstract: A small trainable advisor can steer a frozen language-model executor using natural-language advice. In addition to learning from task rewards, the advisor can use feedback from completed interactions to improve its advice. However, a plausible correction need not change execution, yet learning from such corrections can still affect the advisor's future decisions in other contexts. In a shared-parameter model, we prove that such corrections can limit learning if their targets favor useful advice less strongly than those of other corrections. Keeping them less often than the rest improves the model's eventual performance compared to learning from every correction. Motivated by this, our method, Advisor Self-Distillation (AdviSD), pairs outcome-based reinforcement learning with self-distillation from a feedback-conditioned copy of the advisor selectively. Reflection proposes corrections, and the advisor scores the same recorded executor response with and without its issued advice, using the magnitude of the difference to select decisions for supervision. This approach does not require executor likelihoods or additional executor rollouts. Experiments with Qwen3-8B advisors for Gemini and Claude show that AdviSD outperforms advisor-GRPO by 4.2-6.4 percentage points on BFCL-v3 and by 3.9-5.1 score points on EnvScaler. The trained advisors generalize to out-of-domain tasks and transfer across different executor versions and model families. AdviSD also beats matched-count random selection, supporting the value of its selection rule.
SkillCome improves large language model skills with group contrast learning
SkillCome: Group Contrast Skill Optimization with Dual Memory
Abstract: Skill evolution improves the capabilities of large language models by analyzing trajectories generated under a given skill and modifying the skill accordingly. Existing approaches typically generate a single trajectory per question. However, this provides insufficient optimization signals since it requires inferring effective skill edits from a solitary path. It is difficult to pinpoint which actions caused the failure in a failed trajectory, or to determine which actions in a successful one should be incorporated into the skill. Furthermore, they rely on a local batch of trajectories for analysis, making the optimization direction susceptible to noisy evidence. To address these, we propose SkillCome, a Skill-evolution method based on group Contrast optimization with dual memory. For each question, SkillCome generates trajectories and performs group contrast analysis to precisely identify key behavioral divergences between successful and failed trajectories, offering reliable optimization signals. The dual memory system further accumulates evidence from historical steps to track patterns shared across different groups, leading to more generalized optimization directions. Together, SkillCome builds a systematic optimization process that transforms experience from observed successful trajectories into reusable skills. Extensive experiments on six benchmarks spanning question answering, reasoning, and agentic tasks demonstrate the effectiveness of our method. SkillCome consistently outperforms baselines across five models of varying families and scales, with gains up to +5.69 points.
Spatial credit assignment improves gui agent training from click feedback
SCA: Spatial Credit Assignment for Reinforcement Learning of GUI Agents
Abstract: GUI agents automate tasks on digital devices by grounding language instructions in visual interfaces. Existing group-relative reinforcement learning improves GUI action prediction by comparing the rewards of multiple responses sampled from the same GUI state. However, binary evaluation treats spatially different failed clicks as identical and provides no relative signal when all sampled clicks fail. To address these limitations, we propose Spatial Credit Assignment (SCA), which uses the screen coordinates of sampled clicks to refine group-relative credit. Specifically, SCA predicts each held-out response's reward from the other responses in groups containing both successes and failures, then uses the prediction residual to adjust credit. When all sampled clicks fail, SCA instead orders them by distance to the annotated target. These spatial references are used only to construct the training update; the deployed policy remains unchanged. We evaluate whether this correction improves the policy update itself by comparing its error and directional alignment with the exact return gradient in a controlled synthetic study. Across GUI grounding and offline action-prediction benchmarks, SCA improves grounding across professional domains and achieves the strongest results among reinforcement-fine-tuned models on most action-prediction metrics, with consistent gains across the reported GUI suites.
Failure-transparent agents reduce false success claims in AI tools
Failure-Transparent Agents: Benchmarking Post-Failure Reporting in Tool-Using Language Models
Abstract: Tool-using agents can fail twice: a required tool can fail, and the agent can then report success without the evidence needed to justify it. Existing benchmarks often entangle this reporting failure with tool selection, recovery, and environment dynamics. We introduce Failure-Transparent Agents (FTA), a controlled benchmark that fixes the failed observation and required evidence state before generation, making post-failure claims directly auditable. FTA contains 100 tasks with deterministic failure traces spanning five failure families, a neutral control, and four user-pressure conditions, and evaluates unsupported claims alongside useful recovery. Across six models, three response policies, and 3,600 human-annotated responses, false-success rates are 22.8% under the baseline policy, 9.3% with a transparency instruction, and 0.8% with a structured evidence contract. Fabricated-detail rates decrease from 28.3% to 14.3% and 0.8%, while useful responses increase from 74.9% to 89.2% and 98.8%, respectively. The tested evidence-contract policy is associated with substantially lower post-failure reporting errors while useful-response rates remain high within this blocked-task benchmark.
Model uses confident answers to stop reading early and save time
The Model Knows When to Stop: Training-Free Early Stopping for Long-Context Reading
Abstract: Language models often process long inputs sequentially in chunks, but continuing to read after sufficient evidence has been acquired wastes computation. Existing stopping mechanisms either learn sufficiency from internal activations or train an exit gate, while a simpler alternative asks the model whether it has read enough. We introduce Answer-Convergence Stopping (ACS), a training-free stopping rule that measures rather than asks. After each chunk, it probes the frozen model's current answer state and stops when that state is both confident and stable. The rule requires only output-side generation and token log probabilities, has no trained components, and uses one shared configuration across models and benchmarks. Because a stopping policy can save computation simply by stopping too early, we evaluate the stopping decision itself using evidence position where available. On the full LongBench-v2 with two frontier models, ACS is the only stopping policy that matches or exceeds full-reading accuracy. Furthermore, across 250 S-NIAH questions, the premature stopping rate for ACS across five models from two families ranges from 0% to 12%, compared to 8.4% to 45.6% for the verbalized gate. Taken together, ACS reveals that by properly utilizing the output signals of frozen models, we can achieve favorable behaviors like adaptive stopping without the need for additional training.
Memory system for AI agents improves itself before errors happen
Remember Before You're Asked: MemDream for Self-Probing Memory Evolution
Abstract: Memory is essential for enabling LLM-based agents to maintain coherent, personalized behavior over long-horizon interactions. However, existing memory systems share a fundamental limitation: they never proactively test their own memory, repairing it only after real queries expose weaknesses. This reactive paradigm means every retrieval failure corresponds to a real interaction in which the cost has already been paid. We propose MemDream, a framework that enables self-probing memory evolution for LLM agents. Our framework periodically enters offline dream cycles where three specialized agents (Dreamer, Analyst, Consolidator) collaboratively probe, diagnose, and repair the memory graph before failures occur. A policy trained via Group Relative Policy Optimization learns which repair operations produce durable retrieval improvements, while a soft decay mechanism provides reversible forgetting driven by the same anticipatory signal. Experiments on LoCoMo and MemoryAgentBench demonstrate that MemDream improves answer F1 by 4.5 points on LoCoMo and achieves a 9.1-point higher overall score on MAB over the strongest reactive-evolution baselines.
Fixed language model judges make version-dependent mistakes evaluating agents
Frozen Judges, Moving Agents: Version-Dependent LLM-Judge Error and the Limits of Judge-Assisted Agent Evaluation
Abstract: Language-model judges compare agent upgrades with their predecessors, but a fixed judge can make version-dependent mistakes. We analyze 35 public coding-agent submissions (20 prespecified version pairs on 250 SWE-bench Verified issues), two customer-service agents (155 tau-bench tasks), and 1,106 expert-labeled AgentRewardBench trajectories. An upstream outage left three judges for the primary SWE-bench analysis (8,743 aligned agent-task cells); the fourth is descriptive. All three coding-agent judges and all four tau-bench judges reject task-conditioned error invariance after multiplicity adjustment. On SWE-bench, 32 of 60 judge-by-pair units have a detectable differential comparison component; eight judge-only intervals declare improvements that execution-based intervals cannot establish, despite rank correlations of 0.71-0.79. In tau-bench, one judge confidently reverses a nine-point reference-reward gap by penalizing a procedural habit the reward ignores. False acceptance of failed coding patches rises with agent capability conditional on task and reference outcome, while a task-solvability prediction from AgentRewardBench reverses sign in SWE-bench. Transporting old-version calibration raises mean absolute comparison error on SWE-bench from 3.8 to 19.5 percentage points; 24.6% of ratio-bootstrap draws are undefined near the correction boundary. A tuned paired audit narrows a classical interval by only about 5% at 80 labeled tasks. A randomized three-arm test does not support the predicted increase in false acceptance from showing the agent's final report (all three Holm-adjusted p-values = 1.0). These results favor explicit reference standards and paired audits of current outputs over judge-only release decisions or transported old-version calibration.
Recursive skill evolution improves large language model agents performance
R$^2$ Flow: Recursive Self-Improvement via Recursive Skill Evolution
Abstract: LLM-based agents can improve themselves across tasks by reusing and revising the skills they orchestrate into executable procedures. Flow-based training fits this loop: it samples procedures in proportion to reward, and the flow through each skill credits it for the next library revision. Three obstacles stand in the way of making this self-improvement reliable: flow training suffers strategy collapse over tree-structured histories; nonnegative flow-based credit rewards frequent use as if it were benefit; and library edits rest on the task reward the policy optimizes. We introduce R$^2$ Flow, a recursive self-improvement framework that alternates policy learning, independent verification, and versioned skill-library updates on a shared-state orchestration graph. The graph merges histories that differ only in the order of independent steps, allowing flow training to pool evidence across equivalent executions. A flow-share readout of the trained flow, invariant to the backward policy, and a separate signed utility rank which skills to change, verifier evidence decides whether an edit is warranted, and a residual-variance plateau sets when to update. Committed edits reshape the graph the next policy learns on, realizing recursive skill evolution. Across question answering, mathematical reasoning, interactive decision making, and code generation, R$^2$ Flow improves task accuracy and library-edit precision over heuristic orchestration, reinforcement learning, and skill-evolution baselines, and transfers across executors. Code is available at https://github.com/beita6969/r2flow.
Evolution-aware memory improves long-term AI agent interactions
EMIR$^2$: Evolution-Aware Memory with Intent-Guided Multi-Round Retrieval
Abstract: Long-term memory enables large language model (LLM) agents to leverage historical interactions for future tasks. However, existing memory systems struggle to utilize continuously evolving historical information, as they often rely on static memory representations and single-round retrieval strategies, failing to track factual changes or integrate distributed evidence across long-term interactions. To address these challenges, we propose \textsc{EMIR}$^{2}$, an \textbf{E}volution-Aware \textbf{M}emory framework with \textbf{I}ntent-Guided Multi-\textbf{R}ound \textbf{R}etrieval, enabling LLM agents to maintain evolving historical knowledge and adaptively retrieve relevant evidence. Specifically, \textsc{EMIR}$^{2}$ constructs a State-Evolving Memory Graph (SEMG) that represents long-term memory as evolving knowledge states supported by temporal event trajectories and evidential associations. By maintaining semantic states through evidence-based updates, SEMG preserves historical evolution and enables evidence tracing under complex and conflicting scenarios. Building upon this, we introduce an intent-guided multi-round retrieval mechanism that iteratively identifies missing evidence and expands retrieval based on accumulated information. Experiments on LoCoMo and MemConflict demonstrate that \textsc{EMIR}$^{2}$ improves long-term memory utilization, dynamic and static conflict handling, and complex retrieval performance, achieving relative improvements of more than 12\% in certain categories. These results highlight the effectiveness of jointly modeling memory evolution and adaptive evidence acquisition for long-term agent interactions.
Multi-agent systems manage changing tasks with issue tracking
RepoMAS: Solving Progressively Specified Tasks with Issue-Driven Multi-Agent Systems
Abstract: LLM-based multi-agent systems (MASs) have shown strong potential for solving complex tasks, but most assume that task requirements are sufficiently specified before execution. In practice, user requests are often incomplete, and additional requirements may only become clear during reasoning, tool use, or execution. We refer to such problems as progressively specified tasks. To systematically study this setting, we introduce ProgSpec, a benchmark that evaluates final outputs against requirements explicitly stated in the initial request and additional requirements supported by the available task evidence. We further propose RepoMAS, an issue-driven multi-agent framework inspired by open-source project management. RepoMAS records newly discovered requirements, conflicts, and failures as structured Issues and uses them to revise the task specification and execution structure during problem solving. Across ProgSpec and five existing benchmarks, RepoMAS achieves the best performance. Further analyses show that its issue-driven revision and repository maintenance mechanisms consistently contribute to performance. These results highlight the importance of allowing MASs to revise not only how a task is solved, but also revise their explicit representation of task requirements during execution.
Enterprise agents struggle with hidden facts in company data questions
Era by Eon: Benchmarking Enterprise Agents on Hidden Knowledge
Abstract: In the Era by Eon benchmark, each question states the rules for its answer, and code computes the answer from a generated company's data. When agents can run code, the four strongest models each answer 22 to 25 of 27 such questions, so the benchmark barely separates them. We add eight question templates that depend on hidden facts. No question or document states a hidden fact, and the records that seem to hold it show something else. Other data implies it. For example, the sales system says a customer dropped a purchase because of timing. On a recorded call, the customer blames an outage. For each generated company, code fills each template and computes an exact answer without a language model. We evaluate 12 agents. Each pairs a model with an agent program, which connects it to the company's systems. The best agent answers 18 of its 24 attempts, three per question, correctly. Four of the six models answer at most 6 of 24 with any program. The hardest questions require picking one of several similar records, such as which of three renewal offers a customer signed. All agents together answered two such questions correctly in only 1 of 84 attempts.
Framework improves math tutoring by learning from past conversations
REAT: A Reflective Experience-Augmented Tutoring Framework for Multi-turn Mathematical Instruction
Abstract: Current Large Language Models (LLMs) excel at solving complex mathematical problems, yet this proficiency does not inherently translate into effective tutoring. While advanced LLM tutors may leverage multi-agent frameworks or fine-tuning, most still lack a mechanism to systematically accumulate and reuse pedagogical experience over time, limiting their adaptability to diverse student needs during fluid, multi-turn interactions. To bridge this gap, we propose the Reflective Experience-Augmented Tutoring (REAT) framework, which couples experience distillation from historical dialogues with real-time adaptive retrieval. Driven by a multi-agent Observer-Critic-Mentor (OCM) distillation pipeline, REAT reviews past conversational trajectories and distills raw interactions into structured, problem-agnostic pedagogical experiences. During live tutoring, a state-aware retrieval module injects these curated experiences to provide adaptive scaffolding based on the student's cognitive state. Experiments demonstrate that the proposed framework significantly outperforms both prompt-only and supervised fine-tuning (SFT) baselines, particularly in improving complex, low-scoring tutoring scenarios. Crucially, the distilled experiences exhibit robust generalization across diverse model architectures and mathematical datasets.
Typed classifier judges answers cheaper and faster than llms with similar mistakes
JEV vs. LLMs as Rubric Judges: Cheaper, Faster, and Wrong in the Same Places
Abstract: We ask whether Jev, a typed classifier that returns probabilities over permitted answers without generating text, can replace an LLM rubric judge. We compare it with three flash-tier LLM judges on nine panels drawn from seven benchmarks, giving every judge identical criterion texts. Jev's accuracy differs significantly from an LLM judge's in only 8 of 27 paired comparisons, ahead mostly on binary criteria and behind only on graded ones, and most of the other comparisons are inconclusive. Summed over the nine panels, the LLM judges, called once per criterion, cost 29 to 325 times as much as Jev and took 30 to 220 times as long. On graded criteria all four judges agree more with one another than with the labels and mostly assign lower levels than the raters. One of several observational accounts is that raters followed scale conventions our criterion texts omit. Jev's confidence ranks its own errors on most panels, which should make a cheap classifier the ideal first stage of a cascade that defers its uncertain verdicts to an LLM judge. Correlated errors undo that advantage. The LLM judges repeat nearly all of Jev's most confident errors, so a cascade replayed on the recorded verdicts lowers cost but gains at most 1.5 points over the best single judge with cross-fitted thresholds, and at most 2.0 even with oracle thresholds.
SkillAA improves AI skill updating with precise graph-based editing
SkillAA: Attribution-Guided Skill-Graph Updating with Targeted Validation and Rollback
Abstract: External skills provide domain procedures without parameter updates, but existing methods often edit skills directly from failed rollouts without structured routing from an observed failure to an editable location; existing skill graphs also underuse semantic boundaries, object addresses, and topological dependencies for skill retrieval, targeted updating, and scoped validation. We introduce SkillAA (Skill Abductive Attribution), a structured skill-optimization framework for frozen language models. It represents skill applicability, execution, and composition in a unified graph, allowing the same structure to support skill selection, attribution-guided repair, and update validation. SkillAA contrasts successful and failed executions to route candidate repairs to specific graph objects, updates only the selected local structure, and uses Local and Big Gates to screen candidate changes before commitment. With gpt-5.6-sol, SkillAA reaches 81.5%, 66.7%, and 91.2% on SearchQA, LiveMath, and DocVQA, respectively, and attains the highest observed mean in every main setting. These results support the utility of attribution-guided graph editing and graph-scoped validation.
Multi-turn agents refined to cut cost and boost accuracy
Dependency-Aware Trajectory Refinement for Efficient Multi-Turn Agent Fine-Tuning
Abstract: Multi-turn agent trajectories often contain redundant rounds (failed tool calls, parallel sub-queries, verification-only steps) that inflate both training and inference cost. We propose viewing each trajectory as a \emph{round-level dependency DAG} that exposes which rounds are globally load-bearing for the final answer, and fine-tune agents on trajectories refined through this DAG. Given an LLM-annotated DAG, these edits are deterministic and interpretable, with optional rephrasing. Models trained on these refined trajectories consistently outperform those trained on the original trajectories at lower inference cost. Specifically, across four multi-modal QA benchmarks, our refinements improve downstream accuracy by up to $1.7$\,pp over vanilla SFT (and $5.7$\,pp over an LLM-deletion baseline) while reducing per-sample inference messages by up to approximately $40\%$ and inference tokens by up to approximately $48\%$, translating to substantial savings in compute and serving cost. Code is available.
Speech language models improve reasoning accuracy with real time self correction
RetroThinker: Enabling Retrospective Thinking in Speech LLMs
Abstract: Speech large language models (SpeechLLMs) offer reduced latency and retain paralinguistic nuances that are typically lost in cascaded automatic speech recognition (ASR) and text-based LM architectures. However, they continue to lag behind text-only LLMs on complex reasoning tasks, while real-time spoken interaction imposes strict latency constraints. Although prior works employ Chain-of-Thought (CoT) and concurrent reasoning to enhance reasoning capabilities without inducing prohibitive delays, an inherent accuracy-latency trade-off persists. In this paper, we investigate whether a streaming SpeechLLM can dynamically revise its reasoning traces on the fly. We introduce RetroThinker, a multi-stage post-training framework that equips the Moshi model to self-verify and forward-correct CoT steps during inference. RetroThinker combines supervised fine-tuning (SFT) on curated retrospective thinking data with length-based direct preference optimization (DPO) to optimize retrospective during early reasoning (i.e., reasoning concurrently while the user speaks). Evaluated on the GSM8K benchmark, RetroThinker significantly improves the accuracy-latency trade-off over non-retrospective baselines, achieving an 11% absolute accuracy gain at a comparable latency.
Weaker AI models improve task success with expert-guided corrections
Co-Evolving Harnesses and Models: On-Policy Correction Helps Weaker Models Catch Up Where Imitation Fails
Abstract: Agent harnesses (the system prompt, tool set, execution hooks, and context-management scaffolding around a model) are a critical determinant of agentic task success. Automated harness evolution can enable smaller models to perform well on domain-specific tasks at a fraction of frontier-model cost. Since both the harness and model weights shape behavior, we ask how harness evolution and lightweight fine-tuning should be combined. Across seven enterprise agent tasks, we first evolve a harness with the weaker model, then find that a stronger expert often uses it more effectively, suggesting expert supervision could close the remaining gap. However, training the weaker model on the expert's complete trajectories under the evolved harness backfires: performance regresses on all seven tasks by 4 to 30 points across Qwen3-Coder and Gemma 4, even though the same procedure helps under the unevolved harness. Our analysis shows that imitation transfers knowledge and increases scaffold usage, but disrupts model-harness fit: the weaker model adopts the expert's planning strategy without the competence to execute it and no longer matches the harness evolved around its native planning style. We therefore develop an on-policy expert-correction pipeline, automated by a meta-level MLE agent, that localizes the failing turn in the weaker model's own rollout and asks the expert to rewrite only that turn. This preserves the model's planning style and combines the gains of harness evolution and model adaptation. Our results identify and resolve a source of contention between harness and weight updates, yielding a compatibility-preserving recipe for economical co-evolution on domain-specific enterprise tasks.
Memory cleaning improves task success for long running AI conversations
MeClear: Cooperative Game-Theoretic Attribution and Risk-Aware Memory Clearance for Long-Horizon LLM Agents
Abstract: Long horizon Large Language Model (LLM) agents rely on external memory systems to preserve user preferences and task knowledge across extended interactions. Conventional retrieval mechanisms optimize semantic compatibility rather than downstream utility, frequently introducing outdated, misleading, or conflicting evidence into the active context. We present MeClear, a task conditioned memory clearance framework that identifies memories featuring negative downstream utility through cooperative attribution and selectively suppresses them from agent execution. MeClear combines Leave One Out screening with sampled cooperative Shapley attribution to distribute utility across interacting evidence, effectively resolving redundant conflict masking where single removal evaluations fail. Utilizing attribution rankings, MeClear executes a query scoped minimal clearance strategy over a nested filtration, verifying task recovery on the cleared context without permanently altering the persistent memory bank. Comprehensive experimental evaluations across ten long dialogue memory pools demonstrate that MeClear achieves a target recall of 85.9% and an overall task recovery rate of 82.3%, representing a 25.5 percentage point improvement over Leave One Out (LOO) baselines.
AgentGrad improves multi agent system prompts with guided fixes
AgentGrad: Intervention-guided Prompt Optimization for Multi Agent Systems
Abstract: Large language model (LLM)-based multi-agent systems (MAS) achieve strong performance by employing specialized multiple agents, yet their performance depends on the prompt design of each agent. For MAS prompt optimization, textual gradient methods that guide prompt updates using natural-language feedback have emerged as a leading paradigm. In this paper, we identify limitations in two stages of existing textual gradient approaches: gradient extraction and gradient aggregation. In gradient extraction, previous works select a target prompt without verifying whether modifying it resolves the failure, and derive gradients without agent-level supervision over the corresponding agent's intermediate output. In gradient aggregation, individual gradients are randomly grouped and concatenated, often mixing unrelated failure modes and producing prompts that fail to generalize. To address these limitations, we propose \textbf{AgentGrad}, a prompt optimization framework for multi-agent systems based on sequential intervention and semantic textual gradient abstraction. For each failure, sequential intervention modifies the behavior of one agent at a time to identify the target agent whose modification resolves the failure. The modified output of the target agent then serves as agent-level supervision for extracting a fine-grained gradient. Semantic textual gradient abstraction clusters semantically similar gradients to prevent mixing unrelated failure modes, and abstracts each cluster into a generalized gradient that captures the shared corrective pattern. Experimental results show that AgentGrad achieves state-of-the-art performance across five MAS benchmarks and reduces wall-clock optimization time by $2.5\times$ on average compared to the next-fastest baseline.