Chat
-
What Looks Like a Capability Limit in Vision-Language Models Is a Readout Limit
2026-09-23 06:21:46Alfredo F. Frontera Del ValleAbstract
Benchmarks for vision-language models offer their answer choices in some convention: a letter, a color name, a pixel coordinate. That convention is treated as neutral. We find it is not, and that the limits a benchmark reports can belong to the readout rather than to the model. On 200 COCO photographs, Qwen3-VL-4B picks the correct one of nine locations for a named object 68.5% of the time when the locations are given in English and 20.0% when the same locations are given as pixel coordinates. Chance is 11.1%. The cost arises when the answer options are coordinates; giving the model a coordinate in the question instead costs 3.5 points and is not significant. The gap holds on a 4x4 grid, under 8-bit rather than 4-bit quantization, and in every slice by object size, boundary distance and category. It also decides which model wins. Two models that tie under English names differ by 39 points in one coordinate system and by 54 in the other, in opposite directions. On the color task, three of the four open models capable of the task show the penalty; on photographs, two of three open models do, and so does Gemini, at 11.1 points on parseable answers (p = 1e-4). GPT-4o does not. To ask whether a model reads a coordinate at all, we attach the wrong name to each one and record which the model follows. Color options written as hue angles are followed below chance; a normalized pixel convention is followed at four times chance. This tells apart conventions a model can use from ones it cannot, though it did not predict accuracy on two untried conventions. Five models also name the same color wheel five different ways, so a fixed answer vocabulary is not neutral across models either. Five times during this work we measured a capable model as incapable because our scorer and the model disagreed about what an answer looks like. We report each case. They are the phenomenon in miniature.
Abstract (translated)
URL
https://arxiv.org/abs/2609.27408
PDF
-
When Parallel Drafter Meets Parallel Speculative Decoding
2026-09-23 05:53:48Fuliang Liu, Xue Li, Kun Qian, Zhibin Wang, Wanchun Dou, Wenyuan Yu, Chen TianAbstract
DSpark-style parallel drafters have made speculative decoding highly effective, yet their draft phase remains serialized on the critical path of every round. Parallel speculative decoding (PSD) overlaps drafting with verification, yet existing methods must guess the accepted prefix and bonus token in advance: a wrong guess reverts the whole batch to serial drafting. We present DPara, a PSD framework that reuses effective parallel drafters yet guarantees backbone--verification overlap in every round, thereby eliminating this probabilistic fallback altogether. While the target verifies, DPara's diffusion backbone precomputes draft representations for every acceptance boundary with the bonus left unspecified; a lightweight autoregressive head then combines the revealed verification outcome with the matching precomputed representation to emit the next round's draft tokens almost instantly---fully parallelizing the dominant backbone forward with verification and leaving only the negligible head cost serial. Experiments on Qwen3-8B and Qwen3-14B across seven math, coding, and chat benchmarks show that DPara achieves average speedups of $3.21\times$ and $3.52\times$ over autoregressive decoding, surpassing the strongest serial and parallel speculative decoding baselines alike.
Abstract (translated)
URL
https://arxiv.org/abs/2609.27396
PDF
-
Forecast Workflow Bench: Evaluating Language-Model Decisions with Budgeted Forecast Tools
2026-09-23 05:35:40Shunya NagashimaAbstract
Time-series foundation models (TSFMs) provide forecasts for operational decisions, but accuracy alone does not determine their value. Evaluating agents that use these models requires measuring decision quality and forecast cost. FWBench evaluates this capability on 1,251 electricity and cycle-hire cases using fixed forecast tools and simulated capacity contracts. Agents select models, histories and horizons, then submit capacities to minimize a stated loss-cost objective. We evaluated two hosted and eight local configurations, including small language models, and tested local models with and without TSFMs. GPT-6 Astra bought inexpensive short-horizon forecasts selectively, using 2.5% of the budget, and outperformed fixed policies when the saved decisions were scored with three loss-cost weightings. FWBench enables reproducible evaluation of how language models select and use time-series forecasts to make decisions under cost constraints.
Abstract (translated)
URL
https://arxiv.org/abs/2609.27385
PDF
-
Psychoacoustically Aligned Latent Smoothing for Adversarial Robustness of Full-Duplex Speech-to-Speech Dialogue Models
2026-09-23 05:29:13Kian Shamsaie, Iman ModarressiAbstract
End-to-end speech-to-speech dialogue models listen and speak simultaneously, so a continuously open acoustic channel is exposed to adversarial manipulation. We formalize imperceptible attacks on full-duplex agents as optimization over additive perturbations confined beneath the psychoacoustic masking threshold of the carrier speech, under three goals: targeted semantic hijacking, response suppression, and policy jailbreaking. Against an undefended Moshi-style agent, white-box attacks succeed in up to 91.7% of trials. We then introduce psychoacoustically aligned latent smoothing (PALS), which injects anisotropic Gaussian noise shaped by local codebook covariance at the residual-vector-quantized latent interface, with input noise shaped by the masking threshold constraining the attacker and trained by a Kullback--Leibler consistency objective. Deployed with no inference-time cost, PALS reduces hijack to 8.3%, mute to 11.2%, and jailbreak to 9.1% at clean quality within 2.3%. A Monte Carlo-smoothed variant certifies an ellipsoidal latent radius up to 0.616, a guaranteed floor that the empirical robustness far exceeds.
Abstract (translated)
URL
https://arxiv.org/abs/2609.27378
PDF
-
Neither Silence nor Overlap Is Failure: Intent-Conditioned Evaluation of Turn-Taking in Full-Duplex Spoken Dialogue Models
Abstract
Benchmarks for full-duplex spoken dialogue models score turn-taking with binary fixed-window rules that reward immediate response or silence by completeness of the prior turn. We argue that the appropriateness of a response offset, whether delayed silence or anticipatory overlap, is conditional on the speaker's latent intent, identifiable only from that speaker's behavior. We introduce TACT, a benchmark of 9,728 episodes and 73.2 hours from five dyadic corpora; each episode carries dialogue history, a per-speaker memory profile, and an annotator-derived posterior over six intent classes. Scoring replaces binary windows with a strictly proper threshold-weighted continuous ranked probability score whose weights are intent-conditioned timing kernels fitted to human floor-transfer-offset distributions, proving boundedness, consistency, and binary reduction. Across eleven systems the best model reaches 0.47 against a human topline of 0.86, is nearly invariant to speaker profiles, and TACT agrees with human judgments at Spearman 0.81 versus 0.46 for binary metrics.
Abstract (translated)
URL
https://arxiv.org/abs/2609.27372
PDF
-
Automated Extraction of Records of Processing Activities Using Hybrid RAG and Locally Deployed Large Language Models
2026-09-23 04:55:53To Duy Hinh, Nguyen Le Quoc Anh, Phan Van Tri, Khuong Nguyen-AnAbstract
Vietnam's Personal Data Protection Law (Law No. 91/2025/QH15) and Decree No. 356/2025/ND-CP, effective January 1, 2026, require organizations to establish and maintain Records of Processing Activities (RoPA). Manual RoPA preparation is labor-intensive, while cloud-hosted large language models (LLMs) may conflict with data-sovereignty requirements. We propose RoPA Manager, a system for automated RoPA information extraction using hybrid retrieval that combines lexical ranking over tsvector, dense-vector search, Reciprocal Rank Fusion (RRF), and locally deployed LLMs. We introduce a Vietnamese RoPA benchmark with 32 organizations, 77 processing activities, 12 field groups, and 4,338 reference values. Evaluation is reported at three distinct levels. The automated scorer, tested on perturbed data without invoking an LLM, achieved F1 = 0.9493 [0.9436, 0.9548]; this measures scorer robustness rather than end-to-end extraction accuracy. End-to-end extraction achieved token coverage of 50.04-55.25% against the reference labels. Two independent experts reviewed 1,558 reference values (35.9% of the benchmark), found no incorrect values, and achieved 99.68% agreement with PABAK = 0.9936. Value-level precision was not measured. Across 32 paired scenarios on a 24 GB GPU, locally deployed Qwen3.5-27B-GPTQ-Int4 showed no statistically significant difference from cloud-based DeepSeek-V4-Flash (difference 0.20 percentage points in favor of DeepSeek, 95% CI [-0.93, 1.32], p = 0.72), while Gemma-4-31B performed significantly worse (p < 0.01).
Abstract (translated)
URL
https://arxiv.org/abs/2609.27359
PDF
-
Guides That Cause Actions: An Offline Study of Guide-Action Mutual Reinforcement in Multimodal Web Agents
2026-09-23 04:49:07Chengguang Gan, Yunhao Liang, QingHao Zhang, Shiwen NiAbstract
Web agents are usually evaluated in live environments, where environment state and judge models drift between runs, so the same checkpoint rarely reproduces the same score, making controlled studies of training phenomena impractical. We present WebMRE, an offline benchmark of 541 tasks and 5,293 steps derived from successful WebArena trajectories, with fully audited test labels and a deterministic protocol that scores a checkpoint identically on every run without any environment. Each step pairs a human oriented guide sentence with a grounded action, enabling the first study of the mutual reinforcement effect between them in web agents. Averaged over three seeds the effect holds for both models in both decoding orders and grows with scale: jointly decoding a guide lifts element selection over an action only reference by 0.9 and 0.2 points for Qwen3.5-4B and by 1.7 and 2.2 points for Qwen3.5-9B. A mediation analysis shows that the guide is a causal channel rather than commentary: forcing the gold guide as a decoding prefix lifts action accuracy from .422 to .684, another step's guide collapses it to .055, and a paraphrase that renames the target still recovers half of the gain, so the channel carries instruction meaning and not only the label string. The same channel yields an offline reward that only a replayable protocol makes computable, though optimizing it from a strong checkpoint brings no gain yet. Our fine tuned models outperform GPT-5.5, Claude Opus 4.8, and Gemini 3.5 Flash, run zero shot, on every offline metric.
Abstract (translated)
URL
https://arxiv.org/abs/2609.27353
PDF
-
EnSIMem: Entity-Structured Indexing for Long-Term Agent Memory
2026-09-23 03:09:15Xuanyu Meng, Xing Fan, Xinyi Fan, Chenlei Guo, Yixuan Xie, Jiawei HanAbstract
An agent that interacts with users over long periods must recall facts, preferences, events, and changes from a continuously growing interaction history. Existing memory systems often compress interactions into generic summaries or retrieve anonymous text chunks, making it difficult for an agent to identify the correct entity, property, and supporting evidence. We present EnSIMem, an entity-structured long-term memory architecture for an agent. During offline construction, the system organizes interactions into theme-coherent episodes and builds dialogue-grounded index entries of the form [entity][entity type][property:value]. Each entry preserves its source turns, temporal information, and available multimodal fields. During online interaction, the agent's request is decomposed into evidence requirements whose properties are aligned with the memory index. Entity-property lookup and adaptive retrieval then collect the evidence needed for point, temporal, compositional, and aggregation reasoning. The agent generates its response from the preserved source evidence rather than from lossy memory summaries. On long-term agent-memory benchmarks, EnSIMem achieves high answer accuracy while maintaining compact contexts and favorable online efficiency. These results show that entity-structured indexing and episode-level provenance provide a reliable foundation for long-term memory in agents. The code of our model is available at this https URL.
Abstract (translated)
URL
https://arxiv.org/abs/2609.27279
PDF
-
Listening and Mirroring: The Effects of Verbal Attunement and Behavioral Mimicry on Social and Empathic Perceptions of Embodied AI Agents in VR
2026-09-23 02:30:43Nathalia Gomez, Haig Shamlian, Omar Khan, Tiffany D. DoAbstract
As embodied agents take on increasingly social and relational roles in VR, visual realism and embodiment alone may be insufficient; users must also perceive these agents as emotionally attuned, supportive, and humanlike. Prior work suggests that verbal attunement and nonverbal mimicry can each improve users' social evaluations of embodied agents. However, behavioral mimicry has largely been studied outside of real-time, conversational AI interactions, leaving limited understanding of how users respond when an agent simultaneously generates contextually responsive dialogue and adapts its nonverbal behavior during an immersive conversation. To address this gap, we developed an embodied AI counselor that combines conversational AI with real-time facial-expression and posture mimicry, while producing either verbally attuned or neutral responses. We evaluated the system in a 2 X 2 within-subjects study with 20 participants, manipulating verbal attunement and behavioral mimicry. Results showed that verbal attunement was the most reliable driver of perceived empathy. Behavioral mimicry showed a marginal relationship with perceived humanness, while greater mimicry exposure showed preliminary, exploratory positive associations with empathy, positivity, and humanness, particularly among female participants. Together, these findings show that multimodal synchrony is not a simple additive strategy for designing empathic conversational agents in VR and underscore the need to consider how verbal and nonverbal behaviors are combined during real-time interaction.
Abstract (translated)
URL
https://arxiv.org/abs/2609.27246
PDF
-
Provably Complete Generalized Planning with LLMs
2026-09-22 22:02:16Katharina Stein, Chaahat Jain, J\"org Hoffmann, Alexander KollerAbstract
Generalized planning aims to compute a plan that solves all instances of a planning domain. Recent work has used LLMs to automatically generate and debug such generalized plans in the form of Python programs and achieved perfect test data coverage for several domains. However, whether these generalized plans are actually complete, i.e. solve all instances of the domain, could only be determined by manual evaluation. Here, we present an approach for automatically generating generalized plans in Lean together with proofs of their completeness relative to a specification of the domain constraints provided as input. We introduce a semantic-preserving PDDL-to-Lean conversion, and use an LLM to generate both the generalized plan and the formal proof that it solves every instance satisfying the domain constraints. The correctness of the completeness proof is determined by Lean's kernel. We evaluate our approach on 13 commonly used benchmark domains, using GPT-5.6-Sol as the LLM. For 12 of the domains we obtain generalized plans together with valid completeness proofs. This is a major advancement of the state of the art in automatic generalized-plan completeness proofs.
Abstract (translated)
URL
https://arxiv.org/abs/2609.27105
PDF
-
EduBehaviors: Assertion-based Schemas for Auditable Coding of Educational Dialogues
2026-09-22 20:38:54Julian Bernado, Ana Trindade Ribeiro, Xander Beberman, Susanna LoebAbstract
Large language models have allowed the rapid deployment of pedagogical annotations corresponding to constructs of interest, allowing a natural language interface for generating classifications on a conversational dataset. However due to the opaque nature of LLM reasoning, we have no verifiable, mechanistic insight into why a model chose a label for an utterance. We introduce the EduBehaviors framework, an interpretable, scalable approach to annotating educational data that uses LLMs to measure repeated observable behaviors relevant to many constructs of interest and then learns a classifier for the construct based on these observable behaviors. We evaluate the framework on the TalkMoves dataset, predicting the Teacher TalkMoves labels. Our best configuration results in a macro-F1 of 0.673 and 0.688 Cohen's kappa, proving competitive with direct prompting approaches. In addition, we release EduBehaviors Toolkit, two tools allowing researchers to operationalize the EduBehaviors framework in their own data.
Abstract (translated)
URL
https://arxiv.org/abs/2609.27043
PDF
-
Recognized but Not Produced: A Generation Benchmark for Culturally Specific Kinship Terms
2026-09-22 18:30:09Sahil Pardasani, Madhusudan SinghAbstract
Current literature evaluates large language models (LLMs) on multilingual kinship understanding using multiple choice benchmarks, treating it as a recognition problem. We instead prompt five open weight LLMs to generate kinship terms in three non Western languages (Hindi, Tamil, and Korean) across two communicative tasks and pair this with a matched option-supported selection baseline. On identical relation language cells, GPT OSS120B selects the correct term in 90.67% of 75 valid cells but produces an accepted term in 36.00% of the corresponding attempts; Llama 3.370B shows the same pattern (77.92% versus 24.24%). Since the four-option condition displays the candidate terms and does not require script production, the difference is interpreted as an evaluation format gap rather than direct proof that lexical knowledge is intact. On explicitly specified L3 prompts, accuracy varies sharply, from GLM-5.1 at 72.29% to Llama-3.370B at 24.24%. The paternal-lineage advantage is language specific; it is large in Hindi but weak or reversed in Korean, while Tamil shared-term pairs provide a control for measurement variation. These results show that culturally specific kinship generation remains difficult even when the relationship is explicitly stated and motivate generation-based evaluation alongside multiple-choice testing.
Abstract (translated)
URL
https://arxiv.org/abs/2609.26942
PDF
-
COMED: The Missing Middle Between Routing and Collaboration in Multi-LLM Inference
2026-09-22 18:10:07Norah Alballa, Wenxuan Zhang, Salma Kharrat, Fares Fourati, Zafar Ayyub Qazi, Mohamed Elhoseiny, Marco CaniniAbstract
No single Large Language Model (LLM) is uniformly reliable across queries, motivating multi-model inference systems that either route among models or combine their outputs. However, routing stops after selecting an initial model, while dense collaboration invokes peers on every query. We show that collaboration is non-monotonic: peers can recover failures that no model solves alone, but can also corrupt initially correct answers. We introduce COMED (Controlled Model Escalation for Multi-LLM Deliberation), a post-anchor controller for selective cross-model collaboration. COMED uses anchor self-consistency, router margin, and a lightweight peer probe to accept confident answers, verify ambiguous cases, and escalate only when collaboration is likely beneficial. We formalize this trade-off with a rescue-harm decomposition showing that selective collaboration improves when rescued errors outweigh collaboration-induced harms. Across medical, scientific, and general reasoning benchmarks, COMED improves fixed and routed anchors in all 16 open-weight settings, with gains up to +10.7 percentage points on MedQA while invoking fewer models and using fewer decoded tokens than dense collaboration. On HLE with frontier models, COMED improves GPT-5.5 from 23.1% to 28.1%, outperforming dense collaboration and achieving the best results.
Abstract (translated)
URL
https://arxiv.org/abs/2609.26913
PDF
-
TwinCheck: Evidence-Grounded Negative-Twin Verification for Stateful Tool Agents
2026-09-22 18:09:40Jiaxuan Dai, Tianyi HuangAbstract
A single locally plausible tool call can derail an otherwise successful agent trajectory. Suspicion alone does not justify intervention, because the replacement itself can introduce the very failure verification is meant to prevent. We introduce TwinCheck, an inference-time verification policy that considers replacement only when the trace satisfies an evidence condition tied to a trace-local failure hypothesis. It constructs a trace-grounded counterfactual alternative, a negative twin, and replaces the agent's proposal only if the twin passes structural checks and the pairwise verifier prefers it in both candidate orders. For paired evaluation, exact replay holds the agent's parsed responses and actions fixed until the first accepted replacement, separating intervention effects from resampling. In the primary analysis of 159 multi-turn BFCL V4 tasks with complete exact-replay pairs, the complete policy raises task success for GPT-5.6 Sol from 45.3% to 58.5% (95% task-bootstrap CI [8.2, 18.8]), with no observed success-to-failure regressions. Together, these findings recast execution-boundary repair as a constrained comparison, making the counterfactual action itself the object of verification.
Abstract (translated)
URL
https://arxiv.org/abs/2609.26911
PDF
-
CORE-STACK+: Meta-Learning for Deep Stacked Generalization
2026-09-22 18:03:36Noor Islam S. MohammadAbstract
Stacking heterogeneous vision backbones (CNNs, ViTs, and hybrids) is the de facto recipe for accuracy, calibration, and robustness, yet two coupled pathologies limit its returns. Prediction-space multicollinearity ill-conditions the meta-learner's Gram matrix, inflating weight variance and producing brittle solutions on a thin manifold. Calibration collapse compounds constituent miscalibration through naive linear stacking, so adding more models can hurt expected calibration error (ECE). Existing remedies, ridge regularization, greedy selection, model soups, and SWAG address at most one of these issues, and none jointly target conditioning and calibration in heterogeneous prediction pools. We introduce CORE-STACK+, a preconditioning pipeline with four components: (i) a kernelized redundancy filter that removes non-linear inter-model dependencies invisible to Pearson correlation, using Centered Kernel Alignment (CKA) [23]; (ii) a $<15$K-parameter differentiable meta-feature gate that learns per-sample attention over ensemble statistics; (iii) a spectrum-adaptive Ridge penalty $lambda^{star}=lmax(Chat)/SNR(Chat)$ derived from a Marchenko-Pastur signal-noise decomposition, eliminating nested cross-validation; and (iv) a Laplace-approximate Bayesian blender replacing inverse-RMSE heuristics. We prove a PAC-Bayes excess-risk bound that, for the first time, jointly accounts for prediction-space redundancy and meta-learner capacity. Across six benchmarks, CORE-STACK+ delivers $+1.8\%$ top-1 on ImageNet-1K, $-4.2$ mCE on ImageNet-C, $+0.9$ mIoU on ADE20K, and $+1.3$ AP on COCO, while reducing retained models by 35-57% and inference FLOPs by up to $41%$. ECE improves $2.1\times$ over deep ensembles without post hoc temperature scaling.
Abstract (translated)
URL
https://arxiv.org/abs/2609.26905
PDF
-
Harness as a Language: A Minimalist Agent Framework With Maximal Expressivity
2026-09-22 18:00:08Zhening Li, Joshua Liu, Mateja Vukelic, Nicole Shen, Supriya Lall, Amitayush Thakur, Alex Zhang, Omar Khattab, Jonathan Light, Armando Solar-LezamaAbstract
Modern language-model agents are built around the \textit{agent loop}, where the LLM is placed in an environment exposing a set of tools, and the LLM has full control over the workflow by alternating between tool calls and observing their output. However, certain workflows currently require additional engineering beyond the agent loop itself, such as memory systems and self-improving systems. We built an LLM agent framework, JAZ, to explore the extent to which a minimal harness that is little more than the agent loop itself can accomplish tasks these specialized systems are built for. JAZ exposes a single LLM-based primitive invoke and provides a set of built-in hooks that allow the programmer to apply constraints and monitoring. Generalizing existing code-mode agent loops, \texttt{invoke} is the simplest loop that satisfies two defining properties: (1) the LLM can write arbitrary executable code that can include recursive \texttt{invoke}; (2) everything visible to the LLM --- all inputs to \texttt{invoke} as well as its interaction history with the code environment --- are variables in the code environment. We motivate our design from first principles, viewing \texttt{invoke} as a language primitive representing a function whose implementation is provided at runtime by an LLM every time it is called. To validate the design of our core \texttt{invoke} primitive, we evaluate \texttt{invoke} --- with only prompting, no manually designed tools, harness, or external systems (e.g., memory or the file system) --- on workflows traditionally implemented through specialized external harnesses. On long-horizon workflows requiring recall beyond the context window, JAZ invoke outperforms Letta (MemGPT) by 8\% at half its cost on the recall-heavy portion of StuLife. On continual self-improvement, JAZ invoke outperforms ACE by 4\% at a lower cost on AppWorld.
Abstract (translated)
URL
https://arxiv.org/abs/2609.26891
PDF
-
HARMONY: Hierarchical Agentic Reasoning for MONocular Image-to-Scene Synthesis
2026-09-22 17:59:47Shufan Sun, Chen Wang, Enxin Song, Jiatao Gu, Lingjie LiuAbstract
Compositional 3D scene reconstruction has recently been explored from two directions: agentic reasoning that provides semantic understanding of spatial relationships but lacks precise alignment with input images; and visual geometry foundation models that predict dense point maps from input images but the reconstruction quality is limited. Therefore, recovering a complete 3D scene from a single monocular image with accurate inter-object relationships and high-fidelity reconstruction quality remains challenging. In this paper, we present HARMONY, a hierarchical chain-of-thought framework that leverages both agentic reasoning and visual geometry foundation. Given an image of an indoor scene, starting from an empty 3D floorplan, HARMONY first calibrates the camera against the reference image to establish a semantically-grounded spatial frame, then uses agentic VLM reasoning to recover the 3D room layout and an initial placement order. It then places the objects in a hierarchical order, from wall-mounted elements, free-standing furniture, to dependent decorations on top of furniture. We also use depth-first traversal for furniture so each placement conditions on previously resolved structure and a reflective feedback loop to avoid error accumulation. After each object placement by VLM, we use the point cloud estimations to perform geometry-based refinement so that the rendered image aligns better with the input. HARMONY can produce 3D scenes that are semantically consistent and perceptually aligned with the reference image, extending single-image compositional reconstruction to complex indoor scene images. Experiments on synthetic and real-world images demonstrate that HARMONY outperforms the evaluated reconstruction baselines, while qualitative comparisons with GPT-6 Astra suggest more faithful object arrangements and better preservation of scene details.
Abstract (translated)
URL
https://arxiv.org/abs/2609.26793
PDF
-
Pinocchio: Fast Uncertainty Estimates for Black-Box Language Models
2026-09-22 17:58:07Kevin David Hayes, Arka Pal, Haosong Zhang, Tom Goldstein, Micah GoldblumAbstract
In high-stakes decision-making applications of large language models (LLMs), practitioners require not only accurate LLMs but also uncertainty estimates for their predictions. Existing approaches to uncertainty estimation for LLMs require access to log-probabilities output by the model or require fine-tuning access. However, many industrial LLM products use closed-source API models, and many such API models like GPT do not return log-probabilities and may not allow fine-tuning. We introduce Pinocchio, an external calibrator that estimates the correctness of responses from black-box API models. Trained jointly on responses from seven LLMs, it achieves 0.862 AUROC predicting the correctness of held-out responses from those same models, and shows zero-shot transfer to thirteen unseen models across eight organizations. Our model needs only a single forward pass to generate an uncertainty estimate and requires no access to the target model's logits, weights, or internal states. A lightweight text only 0.8B checkpoint matches our largest model's AUROC. We release code for adding uncertainty estimation to existing repos in only two additional lines of code.
Abstract (translated)
URL
https://arxiv.org/abs/2609.24881
PDF
-
Agensh: Scaling Organizational Intelligence to 1,024 Agents
2026-09-22 17:56:25Zhihao Zhan, Ting Song, Li Dong, Shaohan Huang, Jianxun Lian, Yan Xia, Furu WeiAbstract
A multi-agent system can reduce latency on complex tasks by executing work concurrently. Several pioneering harness frameworks support multi-agent systems. However, the scalability of current multi-agent harnesses is often constrained by a central orchestrator's capacity to allocate tasks and coordinate workers. To address this limitation, we introduce Agensh, a scalable self-organized multi-agent harness without a central orchestrator: concurrent workers execute a multi-agent cooperation loop, continuously gathering context, claiming and self-assigning sub-tasks, taking action and sharing findings, verifying results, and merging progress in an asynchronous manner. The loop is supported by the agentic organization infrastructure comprising three components: a shared workspace holds proposed, ongoing, and completed work; a message interface lets workers communicate; and shared context retains reusable findings and work intentions. To test the scalability of Agensh, we evaluate it on the five hardest ProgramBench tasks with GPT-5.6-sol (high). Scaling from 1 to 128 agents raises the mean final test-pass rate from 19.31% to 28.78%, an approximately 49% relative improvement. Larger organizations reach comparable test-pass rates earlier. On pandoc, scaling from 1 to 1,024 agents raises the final test-pass rate from 33.89% to 55.06%. Worker trajectories further show that different forms of self-organized cooperation gradually emerges and standardizes as the organization grows. These results reveal the number of agents as a new scaling dimension for multi-agent organizations to expand the frontier of general intelligence, offering a practical solution for complex tasks under hard latency constraints or time budgets.
Abstract (translated)
URL
https://arxiv.org/abs/2609.26781
PDF
-
SpeakerMem-R1: Speaker-Centered Dual-Track Memory for Multi-Party Dialogue
2026-09-22 17:56:12Haobo Zheng, Tan Tang, Yan Chen, Weijie Wang, Yingcai WuAbstract
Long-term conversational memory in multi-party settings requires more than retrieving relevant content from long-term conversations: it must distinguish who said what, whom each statement concerns, how individuals perceive one another, what information is shared by the group, and how states change over time. Recent studies on multi-party dialogue benchmarks show that existing general-purpose LLM memory systems tend to lose person and group relations or struggle to integrate clues distributed across members, groups, and time. Together, these issues reveal two core bottlenecks: message attribution and relational understanding in multi-party dialogue, and state reconstruction from interleaved histories. To address both, we propose $\textbf{SpeakerMem-R1}$: its dual-track memory stores speaker-labeled verbatim messages and derived states organized into person-level and group-level views, then combines evidence from both tracks by entity, event, and time at query time. To reduce attribution and update errors during structured memory construction while enabling local deployment, we train Writer-R1 with SpeakerLevenshtein and speaker-conditioned GRPO. On GroupMemBench, SocialMemBench, and EverMemBench, SpeakerMem-R1 achieves binary accuracies of 47.9%, 69.2%, and 61.9%, respectively. On the publicly reported EverMemBench leaderboard from EverMind-AI, we achieves 62.33%, the best reported result among the latest state-of-the-art frameworks. It also achieves 70.85% on all 1,986 LoCoMo questions, which we use as a two-person long-term conversation boundary test. In a controlled evaluation of 305 questions, RL raises the SFT Writer's mean accuracy from 57.38% to 68.20%. We report both binary accuracy and token-F1, and ablations show that the verbatim and structured tracks, as well as person-level and group-level views, are complementary under the standardized evaluation interface.
Abstract (translated)
URL
https://arxiv.org/abs/2609.26780
PDF
