Chat
-
When Reasoning Traces Conflict with Evidence: How Answer Configuration Shapes Evidence Following
Abstract
A reasoning trace retained at answer time plays two roles: it can raise accuracy, and it can compete with the current evidence when the two disagree. We disentangle these roles by crossing verified evidence versions with the traces generated from each and answering from the same traces and evidence under two configurations: continuation of the model's own thinking and a separate answer call. Across three model families (gpt-oss, Gemma4, Qwen3) on HotpotQA and MuSiQue, the separate call raises evidence-following over continuation by 29.9 to 60.6 points. On the same pairs, removing the trace lowers accuracy by similar amounts under both configurations. Yet continuation repeats the stale conclusion in most conflicts, and the separate call follows the current evidence far more often. Benefiting from a trace and deferring to it are therefore distinct properties. Further analyses show that the final line of a trace carries most of its value, gold evidence helps most before the trace is written, distractors weaken correction under a separate call, and renewed reasoning, unlike a verification instruction, follows the current evidence in most conflicts. Because trace deletion alone cannot reveal which source an answer follows, evaluations should report both properties together with the answer configuration.
Abstract (translated)
URL
https://arxiv.org/abs/2610.20805
PDF
-
Calibration-Preserving Pruning: Compression as a Reliability Contract
2026-10-14 18:40:25Ibne Farabi Shihab, Adria Binte Habib, Joyanta Jyoti MondalAbstract
Split conformal prediction, not the pruning rule, supplies finite-sample marginal coverage once a pruned model is fixed independently of the conformal calibration split. We study the separate efficiency problem: can pruning preserve score geometry well enough to obtain smaller valid prediction sets? Calibration-Preserving Pruning (CPP) augments a base pruning score with nonconformity-gradient saliency and uses disjoint pruning, validation-selection, conformal-calibration, and test splits. Bounded score perturbations imply bounded conformal-quantile shifts and controlled set inflation, but do not make the generic coverage theorem CPP-specific. Final five-seed Qwen2.5-1.5B results at 50\% sparsity show the largest gains on large-label tasks. On DBpedia-14, CPP-SparseGPT reduces mean set size from \(10.1\) to \(8.6\) while changing accuracy from \(0.347\) to \(0.366\); CPP-Wanda reduces \(11.2\) to \(9.0\) with an accuracy trade-off from \(0.310\) to \(0.295\). Across 15 dataset--sparsity cells, CPP-SparseGPT produces smaller sets in 13 and higher accuracy in 11. Matched controls show that generic supervised gradients explain much of the gain: true-label CPP is not statistically resolved from matched Wanda+SNIP, whereas threshold-aware candidate-label CPP reaches \(7.8\) mean set size at explicit accuracy and offline-compute costs. RoBERTa-base and Llama-3-8B diagnostics support transfer, but our claims remain limited to reliability-sensitive classification.
Abstract (translated)
URL
https://arxiv.org/abs/2608.23744
PDF
-
INFUSE: A Unified Surface Benchmark for Indirect Prompt Injection in Agentic Systems
2026-10-14 18:26:15Thuy Dung Nguyen, Jingyi Guo, Rana Seda Dubauskas, Sreepriya Damuluru, Taylor T. Johnson, Kevin LeachAbstract
LLM-based agents increasingly execute long-horizon, end-to-end workflows involving shell commands, code modification, dependency installation, and interaction with remote services. This growing autonomy expands their exposure to indirect prompt injection (IPI), where adversarial instructions embedded in external content are interpreted as legitimate task directives. Existing evaluations, however, often focus on isolated malicious strings, single-step tool calls, or narrowly scoped agent configurations. We introduce INFUSE, an end-to-end benchmark for evaluating IPI throughout the lifecycle of long-horizon agent workflows. INFUSE comprises 40 realistic tasks and a structured taxonomy of 56 attack scenarios spanning five threat categories. Attacks are injected through realistic interaction surfaces encountered during task execution, extending evaluation beyond attacks placed only in initial prompts or isolated tool interactions. We evaluate 19 LLMs from the GPT, Claude, Gemini, Llama, Gemma, and Qwen families, producing 41,824 adversarial trajectories (26,220 without defenses) and 720 benign reference runs. Across the benchmark, mean attack success reaches 19.6% under LLM-based evaluation and 15.4% under heuristic verification, with per-model ASR-L ranging from 3.2% to 41.9%. Existing guard models provide limited protection, while prompt-level instructional hardening reduces mean attack success from 58.2% to 14.3%, revealing substantial but model-dependent security-utility trade-offs. Our code is available at this https URL.
Abstract (translated)
URL
https://arxiv.org/abs/2610.20776
PDF
-
WorldCast: Exploring Intelligent Visual Generation in World State Prediction
2026-10-14 18:00:07Wenhao Yan, Haodong Li, Xiaotong Li, Yaxin Luo, Zhengrong Yue, Jialv Zou, Xu Huang, Haojin Yang, Nan Jia, Yan Feng, Peng Pei, Xunliang CaiAbstract
Humans interpret multimodal observations of the world and forecast how it will evolve. Recent multimodal generative systems have demonstrated similar capabilities: they can integrate multimodal context and synthesize visual outputs that depict changes in the world. To characterize this emerging trajectory, we organize model capabilities into a four-level hierarchy and discover that higher-level tasks remain relatively underexplored despite model progress. To bridge this gap, we explore model performance on multi-step world state prediction, a high-level task that requires models to forecast successive future world states from multimodal observations and diverse references. Such a task requires models to possess entangled understanding and generation capabilities, as well as extensive world knowledge to translate human action sequences, physical transformations, and abstract visual reasoning into coherent visual outputs. We introduce WorldCast, a benchmark comprising 320 manually designed cases across eight categories, and evaluate 12 representative models. The conducted experiments yield three key insights: (a) advances in frontier models are rapidly pushing the boundaries of intelligent world modeling, with the GPT-Image-2.5 System scoring 72.8 and demonstrating early-stage visual state prediction capabilities on real-world tasks; (b) WorldCast reveals substantial gaps between models: even Nano Banana Pro reaches only 42.8 and the best open-source model scores 20.1, with deficits in handling complex context, entangled references, and reasoning-heavy tasks; (c) strong generative models effectively draw on the capabilities of their underlying language models. These findings highlight the need for multimodal generative models that are capable of dynamic planning and reasoning while also being deeply unified to faithfully simulate real-world states.
Abstract (translated)
URL
https://arxiv.org/abs/2610.20731
PDF
-
PerSo: Investigating Facets of LLMs' Social Intelligence in Real-World Persuasion Scenarios
2026-10-14 17:39:35Ali Zamani, Soroush Esfahanian, Erfan Shahabi, Mohammad Taher Pilehvar, Azadeh Shakery, Yadollah YaghoobzadehAbstract
Understanding the social intelligence of large language models (LLMs) is increasingly important for applications involving human-AI interaction and multi-agent environments. Social intelligence comprises two aspects: cognitive intelligence, the ability to infer others' mental states, and behavioral intelligence, the ability to select appropriate social actions based on these inferences to achieve social goals. Persuasion provides a natural testbed in which both aspects are simultaneously at play, yet prior benchmarks typically evaluate them in isolation using synthetic scenarios. We introduce PerSo, a benchmark that evaluates both aspects on real-world persuasion dialogues from Reddit's ChangeMyView community, using an LLM-in-the-loop pipeline to construct the evaluation tasks, validated through human annotation. Our evaluation of diverse LLMs shows that performance generally scales with model size but varies across dimensions. The strongest LLM slightly surpasses the best observed human performance on the behavioral task, while the best human participant remains ahead on the cognitive task.
Abstract (translated)
URL
https://arxiv.org/abs/2610.20548
PDF
-
ReaDy: Recurrent States as Dynamic Context for Speculative Drafting in Hybrid Linear Attention Large Language Models
2026-10-14 17:08:52Bohan Sun, Zunhai Su, Xialie Zhuang, Jianjin Li, Ngai Wong, Chuan-Wei KuoAbstract
Speculative decoding accelerates autoregressive generation by drafting multiple tokens per step, yet feature-based drafters rely on target hidden features whose contextual emphasis may not align with the needs of subsequent draft positions. Hybrid linear attention (HLA) models maintain compact recurrent states that accumulate historical associations, offering an efficient source of complementary context for speculative drafting. We introduce ReaDy, the first speculative decoding method, to the best of our knowledge, to leverage HLA recurrent states for position-specific context retrieval during drafting. ReaDy equips each draft position with a learned query to retrieve supplementary context directly from the target's committed recurrent states. A lightweight state reader integrates this context into draft representations through residual updates, while low-rank state reads and fused operators keep additional computation and I/O overhead low. Building on DSpark's parallel backbone and sequential prediction head, ReaDy employs a lightweight two-stage adaptation strategy, with reader capacity, state sources, and residual placement examined through ablation studies. Across seven benchmarks spanning mathematics, code generation, and dialogue, ReaDy achieves an average $3.16\times$ decoding speedup over autoregressive decoding with Qwen3.8-27B on a single NVIDIA B200 GPU. It outperforms DSpark on every benchmark, improving throughput by 7.97% and acceptance length by 8.50% on average. Our code is available at this https URL.
Abstract (translated)
URL
https://arxiv.org/abs/2610.20503
PDF
-
Same Claim, Different Interpretations: How LLM Agents Resolve Ambiguity in Scientific Claims
2026-10-14 16:42:05Delip Rao, Ahmed Muharram, Chris Callison-BurchAbstract
Co-scientists and agentic scientific claim verifiers are increasingly used to assess claims across the sciences. Natural-language scientific claims, written for human experts, are often underspecified and so admit several defensible interpretations. Such verifiers usually pick one, often tacitly, and benchmarks score their verdicts against expert labels without recording the expert's interpretation or the verifier's in a comparable form. This leads to disagreements with expert labels that cannot be traced to the evidence or the interpretation, and to compute spent on unchecked interpretations. We introduce Claim Cards to make such interpretations explicit in the reasoning itself. A Claim Card records a verifier's interpretation as nine fields before evidence gathering, each marked stated, inferred, or default. To test whether an agent's interpretation can change its verdict, we ran a GPT-5.4 agent five times on each of 60 expert-labeled materials-science and biomedical claims with a card, without one, and with one card frozen across runs (900 runs). Over 40% of card fields were not marked stated, and a card added about a tenth to a run's GPT-5.4 tokens and time without measurably changing agreement with expert labels. The agent's own interpretations varied between runs, although within a claim this variation did not measurably track the verdict. A deliberate change to one field of the frozen card, however, making the quantifier universal or the threshold for support stricter, changed the verdict on about a third of the claims, against about one in seven expected by chance (p < 0.001 for each edit). On 24 materials-science claims run with computational tools, the stricter threshold also changed what the agent asked those tools to compute. We therefore argue that verdicts should be reported with their interpretations and that benchmarks could publish the interpretation behind each label.
Abstract (translated)
URL
https://arxiv.org/abs/2610.20457
PDF
-
From Unreliable to Usable: Decomposing Failure Modes and Confidence Signals in LLM Entity Extraction - Evidence from Cyber Threat Intelligence
2026-10-14 16:13:09Jiahong QianAbstract
Audits of large language model (LLM) extraction from full-length documents report severe unreliability, leaving two questions open: what the failures actually consist of, and whether confidence signals can turn unreliable output into usable output. We run a preregistered, fully frozen evaluation of two new-generation low-cost LLMs (DeepSeek-V4-Flash, GLM-5.3-Flash) with a flagship-tier control (GPT-6.1-Sol) on a public corpus of 354 incident campaigns (315 unique reports, 630 model-report evaluations), scoring extraction quality against gold, confidence calibration, and robustness to single-sentence perturbations. Three findings emerge. (1) Unreliability persists across tiers: micro precision is 0.03-0.34 and verbalized calibration error (ECE) 0.53-0.94 across both splits; the gap does not close at the flagship tier. (2) The dominant failure is not hallucination: a full mention census plus a three-tier annotation of 360 sampled candidates (two independent LLM annotators from vendors outside the evaluated set, Cohen's kappa = 0.79 on the preregistered acceptance line, fully disclosed) back-projects to 65.6% background mentions and 19.9% gold-scope gaps, with only 1.2% unsupported candidates. (3) Confidence ranks but does not verify: zero-cost self-consistency-frequency tiering passes a preregistered top-30% precision-gain bar in all six test-split model-type cells (4/6 on train), the zero-cost text-frequency x SC frequency combination wins 3, ties 2, loses 1 across cells, yet relation-level counterfactuals on all 45 targets show that a single misattribution sentence flips high-confidence judgments in 95.6-100% of cases, and even a background sentence explicitly marked as unrelated still binds a fake vulnerability in up to 26.7% of them. We conclude with a deployment boundary: confidence tiering may set review priorities, never truth.
Abstract (translated)
URL
https://arxiv.org/abs/2610.20417
PDF
-
Evaluating LLMs as Agentic Workflow Executors in Real-World Customer Support Skills
2026-10-14 16:00:21Md Tahmid Rahman Laskar, Gundeep Singh, Xue-Yong Fu, Karol Chang, Kevin Sanders, Shi Zong, Tania Habib, Julien Bouvier Tremblay, Shayna Gardiner, Harsh Saini, Matthias Lee, Elena Khasanova, Quinten McNamara, Shashi Bhushan TNAbstract
Enterprise customer-support agents increasingly rely on Large Language Models (LLMs) to execute bounded, tool-using workflows. We study this problem through the lens of Skill LLMs: specialized executors that operate within a selected skill context and must complete a narrow, stateful workflow under policy and tool constraints. We present a benchmark construction pipeline that mines recurrent workflows from real-world transcripts, converts them into structured skill descriptions, maps them to tool requirements, simulates environment state with Model Context Protocol (MCP)-compatible mocks, and generates multi-turn tool-using dialogues. Using this benchmark, we conduct a large-scale zero-shot evaluation of 17 language models spanning closed model families and deployment-oriented open-weight models from 1.7B to 14B parameters, each judged independently by five strong LLM judges, plus a deterministic rule-based evaluator that checks tool coverage, hallucinated tool use, and unsupported claims. We find a consistent ranking in which GPT-5.4 leads by a wide margin and QWEN3.5-9B is the strongest open-weight model (below 20B), outperforming several closed baselines under every judge. We also find that judge choice is not a negligible methodological detail: pairwise judge agreement on the pass/fail verdict averages Kappa=0.524 across 10 judge pairs. These results suggest that comprehensive zero-shot evaluation of skill execution requires judge reliability as carefully as model capability.
Abstract (translated)
URL
https://arxiv.org/abs/2610.14346
PDF
-
Anuvani-LID: Robust and Calibrated Spoken Language Identification for Indian Languages
2026-10-14 14:29:35Stephen Fernandes, Narayan ChoudharyAbstract
Every multilingual speech pipeline begins with a language decision, an error there propagates silently through recognition, translation and routing. In India it spans 22 scheduled languages across four language families, heard over short, noisy, narrowband and code-mixed audio. Foundational speech models increasingly present language identification (LID) as an inherent latent, generalized capability. Autoregressive decoders in speech models emit a special language ID token, and modern LLMs with audio capabilities can name the language when prompted. We test this assumption on identical clips and find that whisper and similar speech models has tokens for only 15 of the 23 languages, on one second inputs Gemini 3.5 Flash reaches 0.52 macro-F1 and GPT-audio 0.14. We present Anuvani LID, a dedicated 115M parameter LID conformer model trained on 54,643 hours of Indian speech spanning 22 Indian languages and Indian accented English. Along with designed and controlled studies on 102 condition benchmark with speaker and corpus held out tests. Across 37 conditions, it reaches 0.914 mean macro-F1, 5.2 points above Gemini 3.5 Flash and ahead of 14 other systems. It is 24 points ahead of Gemini at one second, scores 0.897 on all 23 languages where no other system exceeds 0.713, and is calibrated to an ECE of 0.008. Reliable Indian LID still requires a robust, dedicated and calibrated model.
Abstract (translated)
URL
https://arxiv.org/abs/2610.20263
PDF
-
Natural Code Transformation Confuses Agents in Vulnerability Analysis
2026-10-14 13:20:51Jie Zhu, Ronik Bhaskar, Heather Zheng, Ben Y. Zhao, Kexin PeiAbstract
Having demonstrated a strong potential for automating vulnerability discovery, AI agents can also be used to attack software systems at scale. Today's software systems need mechanisms to level the playing field by slowing down potential attacks, giving them time to patch vulnerabilities and strengthen software. Along this line, our study of agent trajectories in vulnerability analysis shows a surprising finding: agents produce dramatically different results across real-world code commits that carry the same functionality and the same vulnerability. This phenomenon appears consistently in multiple code domains and vulnerability categories. It motivates us to study natural and functionality-preserving code transformations derived from real-world software development, and their potential as a defense against attacker agents. We build a corpus of natural transformations by mining recurring development patterns from real-world code commits, and apply them to evaluate impact on agentic vulnerability analysis. Our results show that even a single form of transformation can effectively "confuse" frontier agents, e.g., Codex with GPT-6 Sol, reducing successes in vulnerability discovery. We also study different agent failures, whether transformations can be composed to provide stronger protection, and adaptive attackers who seek to reverse them. Overall, our study of natural code transformation significantly expands the research space of using code transformations to defend against malicious agents.
Abstract (translated)
URL
https://arxiv.org/abs/2610.14350
PDF
-
MultiPartyTurnBench: Benchmarking Floor Control in Multi-Party Spoken Dialogue
2026-10-14 13:20:14Puneet Mathur, Dinesh ManochaAbstract
Full-duplex speech systems must decide when to enter, hold, or relinquish the conversational floor, even as multiple speakers overlap or backchannel without taking it. Existing evaluations study dyadic turn-taking or end-to-end agent behavior, but lack a unified, controlled evaluation of genuine floor transitions against confusable non-events in multi-party settings. We introduce MultiPartyTurnBench (MPTB), a benchmark of 240 synthetically generated, channel-separated conversations spanning two to five speakers and six interaction types. MPTB evaluates two complementary tasks: End-of-Turn (EOT) detection distinguishes genuine turn completion from temporary holds, while Interruption (INT) detection distinguishes successful floor takeovers from non-floor-taking speech. Exact event-conditioned annotations enable causal evaluation using false-positive-constrained recall and signed latency. Across acoustic, semantic, and turn-projection baselines, we find a stark asymmetry: learned systems achieve 82.5-92.6% EOT recall at 4.3-11.8% false-positive rates, while INT detection suffers from missed takeovers or excessive false activations. This asymmetry extends to five full-duplex agents which achieve 86.1-95.8% balanced accuracy on turn entry, yet perform near chance (46.9-54.7%) on selective interruption handling. Gemini-3.8-Live yields to 97.7% of backchannels, whereas MiniCPM-o-4.5 yields to only 1.4% of genuine interruptions. These findings establish selective floor control as a distinct, unresolved challenge for reliable multi-party voice interaction.
Abstract (translated)
URL
https://arxiv.org/abs/2610.20172
PDF
-
Gap Detection: Mitigating Homogenization through Demand and Defensibility
2026-10-14 13:16:03Xucheng Yu, Haibo Jin, Xiaoqi Wang, Runze Kong, Haohan WangAbstract
Language models often produce semantically similar answers to open-ended questions, even when many plausible alternatives exist. Generating useful alternatives requires identifying both what the query calls for and what existing responses already cover. We introduce Gap Detection, a framework that searches for relevant but unoccupied directions in a shared text embedding space. Its objective, GapScore, combines query relevance (Demand) with dissimilarity to the nearest existing item (Defensibility). The framework optimizes this product on the unit sphere, maps candidate directions to concept words with a language model, and uses task-specific filtering and constrained prompting to turn those concepts into text. On five curated topics and 90 INFINITY-CHAT queries, the reported mean relative reductions in output Collision are 12.8% and 10.8%, respectively. We also instantiate the framework for generative engine optimization over four product markets and for iterative synthetic-data augmentation. In the latter setting, with 100 training examples per method, Gap Detection reduces Collision from 0.324 to 0.298 relative to undirected accumulation and increases the Vendi Score from 27.16 to 29.20; the accuracy difference is one example on a 40-example test set. Component ablations expose the trade-off between semantic separation and relevance, while also showing that better direction scores do not uniformly yield better downstream metrics. These results support explicit gap search as a practical mechanism for diversifying content, with task validity checked separately from geometric novelty.
Abstract (translated)
URL
https://arxiv.org/abs/2610.20162
PDF
-
Gray-Box Hallucination Detection in LLMs with Token Identities
2026-10-14 11:12:12Shaked Goldshtein, Yitzchak Grunbaum, Fabrizio Frasca, Guy Bar-Shalom, Haggai MaronAbstract
Large language models often produce fluent but factually incorrect output. Gray-box detectors identify such hallucinations from the token-probability statistics emitted during generation, avoiding the need for white-box access to hidden states that frontier models do not provide. These detectors, however, typically rely only on probability-derived representations and ignore the identities of the candidate tokens. They can therefore observe how uncertain the model is, but not what it is uncertain about. Consequently, they cannot distinguish between a model that is torn between "big" and "small" and one that is torn between "big" and "huge" when the two pairs receive the same probability distribution. We introduce GURI (Gray-box Uncertainty Representations with Identity), a lightweight gray-box hallucination detector that combines candidate-token probabilities with per-token vector identifiers. This exposes which alternatives the model considered while preserving the accessibility and transferability of the gray-box setting. Across three open-weight models and three question-answering benchmarks, GURI increases mean test ROC-AUC from 72.09% to 79.09% over a probability-only baseline. On the closed-source GPT-4.1 model, mean ROC-AUC increases from 67.39% to 71.84%. Among the three open-weight models, GURI outperforms transferred probability-only baseline in 17 of 18 directed cross-model transfers without fine-tuning. For the three open-weight models, cross-dataset transfer between HotpotQA and TriviaQA with the model held fixed improves mean ROC-AUC over transferred LOS-Net by 3.76 points.
Abstract (translated)
URL
https://arxiv.org/abs/2610.19996
PDF
-
MolHop: Scaffold-Hopping Molecular Design with Agentic Reinforcement Learning
2026-10-14 10:11:10Yanyi Su, Chenyan Wang, Yulei Gong, Yeyun Chen, Xiaohui Yang, Jun ChengAbstract
Molecular design is often formulated as an optimization problem over existing molecular spaces. However, molecular optimization and the discovery of structurally distinct replacement molecules represent two fundamentally different search problems. LLM-based molecular design agents typically perform the former, conducting local refinement within the neighborhood of existing structures to improve target properties through structural modifications. In contrast, the latter requires exploration of a broader chemical space to discover molecules with entirely different scaffolds while preserving the functional characteristics of reference molecules. We propose MolHop, an agentic framework for cross-scaffold molecular discovery that enables models to navigate and edit within a pre-organized chemical space. MolHop organizes approximately 1.6 million molecular scaffolds from commercial molecules into a coarse-to-fine hierarchical action space. The agent selects new exploration regions through scaffold navigation actions and adjusts candidate properties through molecular editing. The framework maintains historical trajectory memory to support long-horizon search processes. On drug molecule and commercial compound tasks, connecting Grok-4.6 and DeepSeek-Flash-V4.1 to the MolHop framework substantially improves their cross-scaffold joint success rates. We develop MolHop-9B through agentic reinforcement learning, and MolHop-9B outperforms generative baselines including GPT-5.6-sol, Claude-5-Opus, and Gemini-3.1-Pro, while generating candidate molecules with larger scaffold differences from the reference molecules. These results demonstrate that effective molecular discovery beyond the traditional optimization paradigm requires structured chemical and action spaces, enabling agents to explore new structural regions rather than merely refine existing molecules.
Abstract (translated)
URL
https://arxiv.org/abs/2610.19920
PDF
-
BluePixel: Toward Native 4K Agentic Image Editing
2026-10-14 09:37:09Shanyan Guan, Shiyu Li, Xin Li, Zhizhou Chen, Yunzhe Xu, Yuehao Liu, Ruixue Shen, Shiju Zhang, Min Jin, Qiang Liu, Junbo Liu, Penggang Shi, Sheng Sun, Xiao Wang, Yong Zhang, Zhongang Qi, Wei Li, Chao Ma, Ying Tai, Yanhao GeAbstract
Instruction-driven image editing has reached native 4K, yet how such editors are trained remains undisclosed, and pretraining one from scratch is beyond most teams. The task itself compounds the difficulty: a single end-to-end editor cannot make precise parametric adjustments, match specialized experts, or interpret the vague instructions users actually write. We argue that 4K editing is a system problem rather than a single-model problem, and present BluePixel, which couples a native-4K editing model with an agent harness; the complete training recipe is disclosed in this paper. The model makes 4K training practical: it turns a 1K-pretrained editor into a native-4K one through progressive VAE adaptation that keeps 4K training as affordable as 1K, absorbs a long tail of vertical editing skills through continual learning, and refines the aesthetic-fidelity trade-off with reinforcement learning. Distillation via source-anchored target projection (SATP) then compresses 50 denoising steps to 4, completing a 4K edit in 8.7 s on a single Kunlunxin P800 XPU while preserving task success rate and background consistency. The harness, BluePixel Agent, makes editing controllable: it decomposes user instructions into specialist and general operations, verifies every result with the same reward model that guided training, and retries within bounded attempts. On VINS-4KEval++ (3,305 native-4K cases), BluePixel ranks second among seven models, behind only GPT Image 2 in non-edited-region consistency.
Abstract (translated)
URL
https://arxiv.org/abs/2610.19882
PDF
-
The World Won't Wait: Evaluating and Teaching Agents Under a Live Clock
2026-10-14 09:35:54Ningxin Pan, Hanyu Li, Yeqing Shen, Ning Ding, Yehui Tang, Wentao ZhangAbstract
Large language model (LLM) agents can solve complex interactive tasks, yet reasoning and tool calls take time. In continuously changing environments, an otherwise correct action may arrive too late. We introduce RealtimeBench, with 180 tasks across six game types, to evaluate agents while games continue running during inference. Paired evaluations show that GPT-6.1-sol outperforms the faster Jev when games are paused, but falls behind when they keep running. We propose Two-Timescale Offline Recursive Self-Improvement (TORSI): a slow teacher studies failures and tests corrections in paused games between attempts, then writes reusable instructions for a fast student. The student acts without teacher calls during gameplay, and validation scores determine which updates are adopted. Two rounds raise Jev's average test score from 13.51 to 24.19 on a 0--100 progress scale, improving all six game types. The instructions also improve average scores for three other models without retuning. These findings point toward real-time intelligence: enabling powerful LLM agents to remain effective in a continuously changing world by organizing deliberation, learning, and action across timescales.
Abstract (translated)
URL
https://arxiv.org/abs/2610.20665
PDF
-
EGCForest: Structured Agent Memory Organization for Enterprise Group Chats
2026-10-14 09:07:53Lingxi Cui, Jinhao Ruan, Wenqing Wei, Guo Yi, Li Rao, Ke Chen, Lidan ShouAbstract
Enterprise group chats record project progress, decisions, task ownership, and other organizational information valuable to enterprise agents. Reliable retrieval is challenging because evidence spans interleaved threads, channels, and time, with meaning and authority conditioned on the speaker. Neither RAG's independent chunk retrieval nor memory systems' summary-based consolidation can reliably recover such distributed evidence. We present EGCForest, a novel multi-view hierarchical memory framework for enterprise group chats. EGCForest organizes an immutable message log into three complementary forests: a Channel forest for temporal workflows, an Actor forest for speaker-conditioned context, and an Artifact forest for deliverable lifecycles. Cross-view connections are enabled by a Placement Map and sparse structural bridges. Built on this memory backbone, EGCForest retrieves and assembles evidence by jointly leveraging message-level relevance, hierarchical context from forest summaries, and cross-view links, recovering distributed evidence. Experiments on three enterprise-chat benchmarks show that EGCForest consistently outperforms RAG and memory-system baselines, improving answer accuracy by up to 20.85 points over MemOS while supporting efficient online retrieval.
Abstract (translated)
URL
https://arxiv.org/abs/2610.19836
PDF
-
Delta Tokenizer: Reference-Identical Tokenization of Session Deltas and Segment-Parallelism for Long Prompts on Multicore CPUs
Abstract
An agent that drives an LLM sends the model the same conversation again and again, and each turn adds only a little text at the end. Inference engines reuse the KV cache of the shared prefix, so only the new tokens pass through the model. However, tokenization has no such shortcut. The CPU front end of the serving engine converts every request back into token IDs from its first character, which lengthens the time to first token (TTFT) and consumes CPU on every turn. We present Delta Tokenizer, a CPU tokenization front end that tokenizes only the delta between a request and the previous request of the same session. It retains the earlier token sequence, encodes again only the appended text and the short tail whose tokens the append could affect, and joins the two only at a position the tokenizer's splitting rules themselves cannot move, which keeps the result exact. For long requests with no state to reuse, it cuts the text where practical tokenizers already keep neighboring pieces independent, encodes the pieces on several cores at once, and concatenates the outputs into the reference encoding, exactly. A shared CPU pool schedules these chunks together with the short incremental updates, so an update runs between two chunks rather than behind an entire long job. The incremental path shortens the tokenization of a session turn by roughly 40x relative to the Hugging Face reference tokenizer, and on 16 cores the chunk-parallel path encodes a four-million-character request 8.2x faster than a serial encode and 1.6-2.1x faster than LoPT. Inside SGLang, Delta Tokenizer reduces the median TTFT through the chat interface by 59% at 100K tokens and brings the P99 TTFT of agent sessions down 9.5x under load and 23x when long requests tie up the front end. Every output we checked matches the reference, among them 41.8 billion documents encoded by its fast encoder.
Abstract (translated)
URL
https://arxiv.org/abs/2610.19801
PDF
-
Learning Perception to Action Rules through Reinforcement Learning for Theory-of-Mind Reasoning
2026-10-14 08:40:06Soyoung Oh, Nursulu Sagimbayeva, Nicholas Jennings, Paulina Hussein, Mehwish Alam, Vera DembergAbstract
In Theory-of-Mind (ToM) reasoning, large language models may correctly infer the agent's mental state but fail to use that inference when predicting or judging the agent's subsequent actions. In this work, we investigate whether explicitly connecting perception, belief, and action with rules improves ToM reasoning. We generate concise \textit{if--then} rules that express relations between perception to action and use Group Relative Policy Optimization (GRPO) to train Qwen2.5 and Llama3 models across four ToM benchmarks. GRPO with rule generation outperforms prompting, supervised fine-tuning, and think reasoning. Moreover, rules generated by larger models ($\geq$3B) provide transferable knowledge that improves downstream performance of GPT-4o-mini. These findings suggest that explicit rules help models bridge the reasoning gap in ToM and support knowledge transfer across models.
Abstract (translated)
URL
https://arxiv.org/abs/2610.19796
PDF
