Chat
-
False Floors: LLM Safety Routing Evaluations Break Under Distribution Shift
2026-10-01 12:09:28Amit Singh Bhatti, Vishal VaddinaAbstract
Safety routers send each request to one of several models and are judged against the best single model. A major routing benchmark picks that comparator on the evaluation data. In the benchmark's own setting this is harmless, but under distribution shift it is not. On HELM Safety the selection cost is 0.003-0.030 of harm under random splits and 0.045-0.113 under held-out categories, comparable to the whole deficit attributed to routing, with its direction holding under either published judge alone. It rises seven- to ninefold on AgentDojo when suites are held out. Across seven safety corpora chosen by rules fixed in advance, three meet a registered interval test and four beat a later permutation null, and three of the four interval misses are corpora where some models have zero observed harm. Prior work proves the direction of this bias. We size it on harm and accuracy, show that it is larger under the held-out splits we measure, and bound it by optimism plus a shift-dependent regret. Scored honestly under shift, routing buys little on these benchmarks. In most pool cells the nested router serves the honest baseline's model, and on the nearly saturated AgentDojo corpus a perfect pre-dispatch router is worth at most two points of harm. We also find a model's expressed recognition of a late injection steerable. On held-out reruns an attacker who knows which model it faces lowers GPT-5.4's judged recognition by 19.6 points, confirmed by an independent label. In an offline counterfactual composition into a controller, the same attack raises or lowers estimated harm depending on the fallback model. Safety routing should be evaluated under shift, against a baseline chosen without the test labels, and recognition-based defences should be scored on harm against an attacker who chooses what the model sees.
Abstract (translated)
URL
https://arxiv.org/abs/2610.01535
PDF
-
LLM-Assisted Discovery of Typed Semantic Links for Ontology Network Construction
2026-10-01 09:59:00Nouha Hayouni, Sheeba Samuel, Alsayed AlgergawyAbstract
Constructing typed, justified semantic links between ontologies is essential for enabling interoperability across heterogeneous and interdisciplinary knowledge domains. However, manually curating such links is difficult to scale. To address this challenge, we propose an end-to-end framework for ontology network construction that automates the discovery and generation of both intra-domain and inter-domain relationships. Our approach combines domain-adapted DistilBERT embeddings for dense contextual representation, clustering-based pre-filtering to reduce the candidate search space, and GPT-4o-driven relationship generation via iterative prompt engineering to produce semantically rich, interpretable links. Applied to ReproduceMeON - a network of 33 ontologies spanning machine learning, microscopy, computational science, and experimental workflow - the pipeline reduces approximately 800k raw concept pairs to 95k high-quality candidates. Human expert validation of 429 generated relationships by two independent annotators yields an overall precision of 80.19% (91.49% on high-certainty annotations) and an F1 of 0.890, with substantial inter-annotator agreement. Comparative experiments against five similarity-based baselines, including Sentence-BERT, show a substantial performance gap (best baseline F1 = 0.581), while an ablation study demonstrates that similarity-based methods alone fail to discriminate valid from invalid relationships (AUC approx 0.5) on the filtered candidate set. These findings highlight the necessity of LLM-based reasoning over concept roles and domain semantics for accurate relationship construction.
Abstract (translated)
URL
https://arxiv.org/abs/2610.01393
PDF
-
Gacha Decoding: Eliciting Diverse Generations Through Instruction Following
2026-10-01 09:48:48Scott Geng, Yufei Zhang, Joseph Lee, Jerry Li, Marjan Ghazvininejad, Pang Wei KohAbstract
We introduce Gacha Decoding, an inference-time method for eliciting diverse language model generations that scales with model capability. Across open-ended domains (in-the-wild chat, creative writing, planning for image generation, and protein design), Gacha Decoding significantly outperforms existing generation diversity approaches at equal quality (up to 2.4x Vendi over the next-best prior approach), reaching the same number of high-quality modes with over an order of magnitude fewer samples (11.0x) and discovering novel modes that no other approach surfaces. Our key insight is to treat diversity as an instruction-following problem: rather than relying on the LM's token entropy, we combine its instruction-following capability with randomness from an external RNG tool to scalably identify and realize distinct modes of the response space. This approach of "planning with dice" enables Gacha to invert the long-observed tension between diversity and model capability. As the underlying LM becomes a better instruction follower, diversity under Gacha Decoding consistently improves--even as its token entropy and diversity under prior approaches decline. Together, our results highlight that instruction following, rather than token entropy alone, can drive generation diversity.
Abstract (translated)
URL
https://arxiv.org/abs/2610.01382
PDF
-
ITC-MoE: Importance-guided Token-aware Compression for MoE Diffusion Language Models
2026-10-01 08:31:26Lianjun Liu, Shipeng Li, You Huang, Weiqi Yan, Mingte Qiu, Huazhong Liu, Xiaofeng Zhu, Yunshan ZhongAbstract
Mixture-of-Experts (MoE) Diffusion Language Models (DLMs) offer flexible parallel decoding and increased model capacity, but their large number of expert parameters incurs substantial computation and storage costs. Existing low-rank MoE compression methods largely rely on static factorization and fixed rank allocation, which overlook the distinctive properties of MoE DLMs. Specifically, we identify two properties: cross-mode non-uniform redundancy, where parameter redundancy and sensitivity to rank truncation vary across the input, output, and expert modes, and token-wise utilization variation, where hot and cold tokens exhibit distinct spectral characteristics and expert activation patterns. To address these challenges, we propose ITC-MoE, an Importance-guided Token-aware Compression framework for MoE DLMs. ITC-MoE consists of two complementary components. First, Importance-guided Adaptive Tucker Compression (IATC) incorporates activation and gradient importance into expert weight transformation, jointly factorizes expert weights across multiple modes, and adaptively allocates ranks under a fixed parameter budget. Second, Token-aware Compensation and Routing (TCR) applies lightweight low-rank compensation to compression-sensitive hot tokens and restricts the candidate expert set for cold tokens with concentrated routing patterns. By jointly adapting compression capacity and inference execution to both parameter redundancy and token-wise variation, ITC-MoE substantially reduces the computation and storage costs of MoE DLMs while preserving their generation quality. For example, on SDAR-30B-A3B-Chat-b32, ITC-MoE maintains an accuracy of 96.33% on MultiArith under a 30% compression budget, while achieving up to a 7.22x end-to-end speedup. The code is publicly available at this https URL.
Abstract (translated)
URL
https://arxiv.org/abs/2610.01296
PDF
-
Reputation, Strategy, and Emotion Effects on Generative AI Cooperation: A Comparison Across Reasoning and Non-Reasoning Models
2026-10-01 07:25:51Celso de Melo, Zishan Feng, James Hale, Kazunori Terada, Giorgio Coricelli, Jonathan GratchAbstract
As generative AI (Gen AI) systems take on increasingly autonomous roles in economically and socially consequential interactions, understanding their propensity to cooperate -- and the signals that shape this propensity -- has become essential. We examine cooperative behavior in frontier Gen AI models using the iterated prisoner's dilemma, manipulating counterpart reputation (positive, unknown, negative), strategy (extortion vs. generosity), and non-verbal emotional signaling (facial expressions conveying competitive or cooperative appraisals). In a first study with non-reasoning models (Claude 3.5, Gemini 2.0 Flash, GPT-4o), cooperation was systematically shaped by all three factors, paralleling patterns long documented in human behavioral research, though models varied substantially in how heavily each factor was weighted. A second study with reasoning models (Claude 4.6, Gemini 3, GPT-5.2) revealed a more concentrated reliance on strategy and reputation, a near-elimination of the Potemkin effect observed in non-reasoning models (evidenced by near-uniform cooperation in a diagnostic harmony game), and a more conditional role for emotion consistent with a hierarchical cue-weighting strategy rather than a simple loss of social sensitivity. Reasoning models also showed heterogeneous end-game behavior, ranging from sustained cooperation to systematic last-round defection effect, revealing model-specific exploitability profiles with direct practical relevance for deployment in negotiation and other multi-round interactions. Together, these findings characterize Gen AI models as increasingly sophisticated, though heterogeneous, social actors, and underscore the practical value of developing standardized cooperation benchmarks to inform the responsible deployment of Gen AI in interactive, socially consequential settings.
Abstract (translated)
URL
https://arxiv.org/abs/2610.01222
PDF
-
Toward Elastic Speech Inference: Training-Free Wake-Word Detection from Pretrained ASR
2026-10-01 06:58:36Hwayeon Kim, Youngwon Choi, Hyeonyu KimAbstract
Recent ASR development has placed growing emphasis on generalization across diverse domains and acoustic conditions. Existing approaches typically adapt pretrained ASR models to front-end functions such as wake-up word (WuW) detection through additional training or task-specific modules. In this work, we explore the use of a shared pretrained ASR backbone for WuW detection without gradient-based fine-tuning and examine whether a compact encoder can be extracted using the PCA-based structured pruning approach of SliceGPT. Experiments with Parakeet-TDT-0.6B-v3 and Moonshine-base show that WuW detection performance remains relatively stable when the encoder channel dimension is reduced by 50%. These results suggest that task-relevant compact encoders can be derived from pretrained ASR models without fine-tuning.
Abstract (translated)
URL
https://arxiv.org/abs/2610.01182
PDF
-
CortexBridge: Cortical Alignment of EEG Montages for Foundation Models
2026-10-01 06:08:56Jiazhen Hong, Xiaotian Zhou, Zihao Ding, Kailong Wang, Yu WuAbstract
Electroencephalography (EEG) foundation models are often pretrained with a fixed channel vocabulary or a limited set of montages, making transfer difficult when electrode layouts change. We propose CortexBridge, a lightweight adapter that combines EEG features with electrode and atlas coordinates to map arbitrary montages into a shared cortical latent space. Evaluated with three frozen foundation models on five brain-computer interface (BCI) datasets from the Mother of All BCI Benchmarks (MOABB), CortexBridge improves performance in 13 of 15 evaluations. The gains in balanced accuracy average 0.80% for EEGPT, 0.70% for LaBraM, and 3.26% for CBraMod, with a maximum gain of 13.02% on 12-class steady-state visual evoked potential (SSVEP) classification. Visualizations of the learned atlas representations reveal task-dependent spatial patterns, with SSVEP showing a more concentrated representation in the Yeo Visual network than auditory P300. These results establish cortical alignment as a learnable and anatomically grounded routing mechanism from heterogeneous EEG montages to pretrained foundation models.
Abstract (translated)
URL
https://arxiv.org/abs/2610.01124
PDF
-
Jev-IDS: System One Models for Network Intrusion Detection
2026-10-01 05:21:46Paulo Severo, Silvio E. Quincozes, Amanda DiasAbstract
Machine-learning Network Intrusion Detection Systems (IDS) depend on substantial labeled datasets and task-specific training, whereas Large Language Models (LLMs) detection can analyze flow records directly but incurs higher inference cost and latency, with less constrained outputs. This paper presents JEV-IDS, an open experimental general NIDS based on the Jev System One Model (SOM) to detect zero day intrusions Under label scarcity. JEV-IDS serializes one flow per request and asks JEV two questions: a binary attack probability and a finite-choice traffic category. Our results show that, at k=1, JEV was 4.8 times faster and 3.8 times cheaper than GPT-5.6 Luna, with 1.5 times higher novel-attack recall; it also produced 15 times fewer false alarms than a low-data Random Forest. Across 5,400 decisions on a 300-flow NSL-KDD pilot split, JEV achieved F1-Score 0.859, precision 0.941, recall 0.790, and novel-attack recall 0.838. Increasing k to 2 reduced its F1-Score to 0.839.
Abstract (translated)
URL
https://arxiv.org/abs/2610.01079
PDF
-
Empty Commitments: When Agents Promise What Their Runtime Cannot Deliver
2026-10-01 04:37:37Jiaqi Tang, Lan Wei, Bingyu Shen, Boyang LiAbstract
A chatbot that says "I will remind you tomorrow" will not run again until the user writes. We call such a promise an empty commitment: a promise of an action after the current turn that nothing in the agent's tools or runtime can carry out. Unlike a broken promise, its emptiness follows from the agent's configuration alone; no later trajectory is needed. We define empty commitments on top of commitment semantics, with three failure types, an anchoring condition for promises that a tool could make real, and a response-level outcome taxonomy. We then describe a measurement protocol: follow-up requests run in five setups that add one persistence affordance at a time, with the environment either left implicit or stated.
Abstract (translated)
URL
https://arxiv.org/abs/2610.01045
PDF
-
Groundability, Not Scale Alone: When Weak Reviewers Can Audit Strong Coding Agents
2026-10-01 04:10:59Junyu Guo, Shangding Gu, Ming Jin, Javad LavaeiAbstract
Coding agents can return plausible patches that omit required behavior. These failures are hard to review because long traces and confident summaries often hide what was missed. We ask when a nominally weaker reviewer can reliably decide whether a patch solves its issue. We study 411 execution-labeled traces from three agents and 101 controlled cases. On 154 GPT-5.4 traces, structured but unchecked evidence raises both defect catch and over-rejection. We then provide official execution evidence as an upper-bound diagnostic. After choosing and freezing one of two formats per reviewer, five of six reviewers improve both rates on 122 held-out traces; two classify every trace correctly. Reviewer size is not a consistent predictor of quality. Because official tests are unavailable in deployment, we also evaluate a frozen cascade with patch-caused static errors and generated tests that first fail on the unpatched repository. On 121 scored held-out GPT-5.4 traces and 59 Gemini traces, its coverage is 0.89 and 0.86, risk is 0.33 and 0.26, catch is 0.76 and 0.80, and over-rejection is 0.66 and 0.67. Most false rejections occur when unresolved cases reach the reviewer. Official execution evidence shows the potential of weak review when decisive checks are available. Producing equally reliable checks without official tests remains the main bottleneck.
Abstract (translated)
URL
https://arxiv.org/abs/2610.01023
PDF
-
Towards Automatic Video Annotation with ASH: Zero-Shot Open-Vocabulary Multi-Object Tracking and Segmentation
2026-10-01 04:09:48Arash Rocky, Q. M. Jonathan WuAbstract
Memory-attention-based Video Instance Segmentation (VIS) methods have demonstrated strong zero-shot tracking capability, yet their substantial memory requirements confine them to short video clips and their single-prompt inference design makes multi-category open-vocabulary tracking computationally prohibitive. This work introduces two contributions toward fully automated tracking annotation of arbitrary video. The Generalized Presence Token (GPT) reformulates SAM3's inference pipeline to process N text prompts simultaneously via virtual prompt batching, reducing image encoding cost from O(N) to O(1) with no modifications to any learned component. The Annotation and Segmentation Handler (ASH) extends any memory-attention VIS tracker to sequences of arbitrary length through overlapping temporal chunks with IoU-based inter-chunk identity matching, requiring no dataset-specific training. Instantiated on SAM3, the resulting pipeline -- SAM3-ASH -- achieves state-of-the-art HOTA on MOTS20 under fully zero-shot conditions and remains competitive with trained specialists across seven additional benchmarks, while peak GPU memory consumption stays below 25 GB, establishing a practical baseline for scalable, training-free automated video annotation.
Abstract (translated)
URL
https://arxiv.org/abs/2610.01022
PDF
-
GUI-HARVEST: Self-Improving GUI Agents through Evidence-Driven Harness Evolution
2026-10-01 02:31:20Geyi Yang, Zikun Qu, Xiang Li, Zhiyong Wang, Min Zhang, Shipei Zeng, Zhongxiang DaiAbstract
The executable harness surrounding a GUI model determines how observations are assembled, actions are executed, and verification, recovery, and termination are controlled. Compared with harness optimization for non-GUI agents, automatically optimizing this harness poses three coupled challenges: reconciling model intent with observed visual effects, diagnosing failures under variable execution outcomes, and identifying recurrent failure patterns across tasks and translating them into reusable runtime changes. We introduce GUI-HARVEST, an automatic harness optimizer that enables self-improving GUI agents with frozen backbone models. First, to ground diagnosis in observed action effects, it aligns model outputs and executed actions with before-and-after screenshots, tying findings to specific interface transitions. Second, to account for execution variability, it treats repeated runs of the same task as a joint evidence unit, using within-task comparisons to locate outcome-relevant behavioral differences. Third, it consolidates verified findings across tasks into recurring failure patterns, maps them to bounded source-code edits with predictions recorded before evaluation, and checks the predicted behavioral effects alongside task performance through repeated execution. Experiments on OSWorld-Verified show consistent held-out gains across six general-purpose open, GUI-specialized open, and proprietary backbone models; Qwen3-VL-32B-Instruct gains 12.33 points on the full suite. Frozen-harness transfer improves GPT-5 by 13.87 percentage points on WindowsAgentArena at 50 steps without further optimization. With the same backbone and initial harness, GUI-HARVEST outperforms Self-Harness and Meta-Harness, suggesting that GUI-specific diagnosis and validation help harness improvements generalize to unseen tasks. The code is available at this https URL.
Abstract (translated)
URL
https://arxiv.org/abs/2610.00948
PDF
-
Finding the Right Fit: Model-Harness Interactions across Agent Tasks
2026-10-01 01:49:59Yixuan Li, Yiyun Zhou, Yao Long Teng, Fuchao Yang, Yanchen Deng, Zhiyi Lyu, Xuyu Dong, Feng Chen, Bo AnAbstract
Choosing an agent system means choosing both a language model and the harness through which it acts. We ask whether a strong model, harness, or pairing stays strong when the setting changes. We evaluate 66 configurations: four configurable harnesses (OpenHands, DeepSeek Harness, PI, and openJiuwen) paired with five models on TUA-Bench, ALE-CLI, and Terminal-Bench 4, plus the native Codex-GPT and Claude Code-Claude pairings. Model rankings reverse across harnesses. On Terminal-Bench 4, Claude leads GPT by 7.94 points in OpenHands but trails it by 30.16 points in PI. For four of the five models, the best harness changes from one benchmark to another, yet some pairings hold: openJiuwen gives Kimi its highest score on all three benchmarks, by 5.61 to 11.11 points. A model's own vendor harness is not reliably its best, and higher cost does not reliably buy a higher score. On Terminal-Bench 4, GPT scores higher under PI than under DSH at less than a quarter of the cost per task. Matched trajectories suggest why fit varies. Models start almost all repairs themselves, so much depends on whether the harness hands failures back in a form the model can use. GPT does best with PI's lean scaffold, while Kimi, which often issues malformed tool calls, does best in openJiuwen. We argue that the model, the harness, and the task should be evaluated together, and we release the harness adapters, evaluation code, and all 6,204 scored trajectories at this https URL and this https URL.
Abstract (translated)
URL
https://arxiv.org/abs/2610.00917
PDF
-
FORALL-LEAN-AGENT for Auditable Reasoning in Formal Mathematics and Software Verification
Abstract
Coding agents increasingly automate Lean proof development, but successful compilation alone does not establish that a candidate proves the intended statement under acceptable assumptions. We present FORALL-LEAN-AGENT, a frontend-agnostic framework for auditable reasoning in formal mathematics and software verification. The framework combines isolated workspaces, Lean tools, and fresh review with statement comparison, axiom audits, and independent proof checking where supported. Verification evidence and reviewer decisions are bound to the same candidate artifact, making acceptance traceable. We evaluate the framework on VeriSoftBench, PutnamBench, and both problems in the Lean Eval softwareverification track. On the 100-task VeriSoftBench subset, integration with FORALLLEAN-AGENT raises benchmark-rule success from 93 to 100 for GPT-5.6 Sol at low effort while reducing cost from $69 to $62. The PutnamBench evaluation accepts all 672 problems at an average of $4.72 each. These results show that agent harness design can improve correctness and efficiency while providing evidence beyond aggregate solve counts.
Abstract (translated)
URL
https://arxiv.org/abs/2610.00885
PDF
-
Kepler: Auditable World Models for ARC-AGI-3
Abstract
ARC-AGI-3 evaluates agents in interactive environments whose rules and objectives must be inferred from observation. We present Kepler, an open-source harness that represents hypotheses as executable world models and validates them through retrospective transition checks and conditional prediction checks. Under one frozen Claude Opus 5 configuration, Kepler obtained a server-verified 100.00 RHAE on all 25 public games, with no per-game model selection or score-conditioned reruns. On 181 of 183 completed levels, the final Opus attempt used no more actions than the corresponding median-human baseline. The retained board runs used 8,256 environment actions, of which 7,292 occurred in scored levels. Retained local provider-session records yield 858.0 million tokens, 97.37% cache reads, and a \$777.72 cost at September 1, 2026 API list-equivalent rates. We also report three evaluation failures: source-code leakage that produced an invalid perfect run, agents reconstructing a removed harness in a control condition, and autonomous repair masking a broken planner. A single-game observation case study showed that animation frames contained task-relevant information absent from settled text grids. Across the final Claude Opus 5 and GPT-5.6 Sol boards, 48 of 50 game-model cells reached 100. These results indicate that public-set score alone has limited discriminative value and motivate first-attempt, cost-conditioned, and verification-aware reporting.
Abstract (translated)
URL
https://arxiv.org/abs/2610.00834
PDF
-
Interpreting Reasoning of Large Language Models via Partial Information Decomposition
2026-09-30 18:44:13Barproda Halder, Qiuyi Zhang, Sanghamitra DuttaAbstract
Large reasoning models (LRMs) have achieved substantial improvements in solving complex mathematical problems, but often produce lengthy, repetitive, or erroneous reasoning trajectories. In this work, we introduce a new interpretability framework, SLIDER, to evaluate the quality of the reasoning process. SLIDER leverages an emerging body of work from information theory called Partial Information Decomposition to disentangle the information about the final answer between two consecutive reasoning steps into non-negative components: unique information (in preceding steps or current step), redundant information, and synergistic information. Building on this decomposition, we propose the *Step-wise Repetitive Reasoning Index (Step-RRI)*, a theoretically grounded measure that assesses whether the answer-relevant information in the current step $S_i$ is predominantly redundant with the past steps $S_{<i}$, relative to its unique and synergistic contributions. To evaluate the effectiveness of Step-RRI in detecting repetitiveness, we apply SLIDER to the redundancy class of the PRMBench dataset where Step-RRI improves step-level redundancy identification accuracy by over $10$ points compared to embedding-similarity and information-gain baselines. Next, we define *Trajectory-RRI*, an aggregate measure of repetitiveness for an individual reasoning trajectory. To demonstrate its practical relevance, we show that average Trajectory-RRI strongly correlates with actual reasoning length across QwQ-32B, DeepSeek-R1-Distill-Qwen-32B, and GPT-4.1, motivating its use as a signal for improving reasoning efficiency. Finally, we introduce *Trajectory-RRI-guided data selection for fine-tuning*, demonstrating that selecting training data based on Trajectory-RRI can improve a fine-tuned model's reasoning efficiency while largely preserving its task performance.
Abstract (translated)
URL
https://arxiv.org/abs/2610.00571
PDF
-
Rules Amortize, Pairings Don't: Linguistic Structure Determines What Latent Task Representations Can Replace In-Context Learning
2026-09-30 18:15:17Gunmay JhingranAbstract
In-context learning (ICL) can be amortized into latent objects (task vectors, function vectors, context vectors) that recover few-shot behavior at zero-shot inference cost, but recent theory shows a static vector acts as a single synthetic demonstration and must fail on high-rank mappings such as word-level bijections. We ask a linguistic version of this question: which linguistic operations can be amortized out of the prompt? We train a 2.6M-parameter network that reads the geometry of a few-shot support set (centroid, principal subspace, spectrum, computed once and cached) and produces an input-conditioned additive update to the query's residual stream at a mid-depth layer of a frozen GPT-2-large/XL. Across eight inflectional directions and one lexical relation, under a canonical split that bars inverted-pair leakage between directions, three regimes emerge. On forward inflection, where 10-shot ICL is strong (0.67-0.89) and extracted task vectors collapse (<=0.06), the transform matches ICL at strictly zero-shot per-query cost. On lemmatization directions, which frozen GPT-2 can execute but 10 demonstrations systematically fail to convey (ICL 0.13-0.48 at 1.5B), the transform is not capped by ICL at all: it reaches 0.78-0.92, up to +72 points over ICL (past to present: 0.85 vs. 0.13). On arbitrary pairings (antonymy) every amortizer plateaus near half of ICL at every scale, capacity, and seed tested. Controls show the support manifold acts as a causally necessary task fingerprint: wrong-task manifolds collapse accuracy to <=0.06, query-only variants cannot disambiguate tasks sharing an input space, and leave-one-task-out transfer is zero. Productive rules amortize into latent task representations, sometimes better than prompting can convey them; memorized pairings do not.
Abstract (translated)
URL
https://arxiv.org/abs/2610.00526
PDF
-
EvoDuet: Bilevel Co-Evolution of Web Searching and Task Solving for Scientific Discovery
2026-09-30 17:58:28Young-Jun Lee, Jinheon Baek, Soyeong Jeong, Minki Kang, Seungyeon Jwa, Jonghyun Choi, Seungho Han, Dongyeop KangAbstract
Evolutionary search with large language models (LLMs) can stall when progress requires external knowledge the model lacks. Supplying relevant documents helps, but simply adding web search tool can keep returning the same pages as solutions change. We introduce EvoDuet, a bi-level optimization method that co-evolves solutions and search queries with fixed model parameters. At each iteration, a retrieval gate lets the LLM assess its knowledge gap and choose to retrieve new documents, reuse stored ones, or proceed without them. An inner loop refines queries and ranks documents by the solution scores they are predicted to yield; an outer loop generates candidates in parallel from these documents and records the evaluated outcomes for later searches. Across 21 optimization tasks with one candidate per iteration, EvoDuet raises OpenEvolve's normalized discovery gain from 74.1% to 78.0% with GPT-5.6-Luna and from 61.3% to 82.3% with Gemini-3.8-Flash, whereas Qwen3.5-9B does not benefit. Our best runs surpass the previously reported best scores on eight tasks, including Swap Reduction on Q20 and Rosetta, and match them on three more. EvoDuet also improves with other scaffolds (e.g., Top-K, EvoX) on Sums/Diffs and Denoising, demonstrating its applicability across evolutionary search scaffolds.
Abstract (translated)
URL
https://arxiv.org/abs/2609.40340
PDF
-
Comparison of techniques for fine-tuning open-weight models for entity extraction from radiology reports
2026-09-30 17:32:17Aawez Mansuri, Kush Mehta, Mohammadreza Chavoshi, Jahanzaib Malik, Theodorus Dapamede, Frank Li, Rohan Isaac, Beatrice Brown-Mulry, Chiratidzo Rudado Sanyika, YoungSeok Jeon, Judy W. Gichoya, Ali Emami, Hari TrivediAbstract
Converting free-text radiology reports into structured labels supports cohort building, quality assurance, and monitoring of clinical imaging models, but the strongest label extractors are hosted proprietary models whose use raises privacy, cost, and reproducibility concerns. We asked whether a fine-tuned open-weight model (Gemma-3-12B) can match GPT-4o at multi-label intracranial hemorrhage (ICH) acuity extraction from non-contrast head-CT reports, and which ingredients matter. Using a 2x2 design, we crossed two adaptation strategies (a discriminative classification head, CH; generative instruction fine-tuning, IFT) with two training-data sources (distillation of real GPT-4o-labeled reports; synthetic reports generated by GPT-4o from real exemplars), across five training sizes, benchmarked on 100 expert-adjudicated reports against GPT-4o and the un-tuned open-weight base. The distilled instruction-tuned model (DIFT) matched GPT-4o (macro-F1 0.845 vs 0.850; p = 1.000) and exceeded the base model by 0.178. The decisive factor was the training-data source, not the fine-tuning method: both synthetic-data models failed to exceed the un-tuned open-weight base at any training size and underperformed the distilled models across all acuity classes. Fine-tuning and inference fit within the memory envelope of a single 24 GB consumer GPU. For narrow, high-value clinical label-extraction tasks, distilling real reports, rather than generating synthetic ones, is what closes the gap to a hosted model, enabling a private, low-cost, version-stable on-premises alternative.
Abstract (translated)
URL
https://arxiv.org/abs/2609.40236
PDF
-
SCB: SpeechConversationBench for Evaluating Multi-Turn Reasoning in Speech-to-Speech Models
2026-09-30 17:13:30Kanpat Vesessook, Saksorn RuangtanusakAbstract
Speech-to-speech systems must solve tasks whose requirements emerge across conversational turns. We introduce SpeechConversationBench (SCB), a focused evaluation of spoken mathematical reasoning using 103 sharded GSM8K problems. The framework compares the original problem delivered in one turn (full), its concatenated information shards delivered together (concat), and incremental spoken disclosure across turns (sharded). We report final-answer accuracy for four commercial speech systems and LEGO, a proprietary speech pipeline developed internally by the SCBX Innovation Lab team with explicit conversational context management. Relative to concat, sharded accuracy decreases by 5.0-25.3 percentage points across the four commercial systems. LEGO achieves 77.5 percent accuracy in all three conditions, compared with 76.6 percent sharded accuracy for GPT-4o Realtime. The two single-turn baselines distinguish sensitivity to problem reformulation from the additional challenges introduced by incremental spoken interaction.
Abstract (translated)
URL
https://arxiv.org/abs/2609.40198
PDF
