arXiv_CL
-
Who Pays for Efficient Inference? Confidence-Based Small-to-Large LLM Routing Across Tasks and Languages
2026-10-16 10:23:41Md Sabbir Hossen, Anichur Rahman, Pabon Shaha, Airin Afroj Aishi, Andrew H. Sung, Md Shohel RanaAbstract
Small-to-large language model routing can reduce inference cost by escalating difficult inputs while retaining easier ones on a smaller model. Its success, however, depends on identifying examples that benefit from escalation, and the reliability of confidence across multilingual tasks remains unclear. We present a systematic evaluation of confidence-based routing across three model families, four languages, and three task types, covering 72 model, dataset, language, and scale configurations and 42,288 predictions. We compare global confidence routing with random routing, equal-quota confidence routing, and an oracle upper bound. Confidence identifies beneficial escalations on the Belebele reading comprehension and SIB-200 topic classification datasets, with mean correction-detection AUCs of 0.612 and 0.705. At 50% escalation, it improves over random routing by 1.40--4.67 percentage points on Belebele and 2.54--5.63 points on SIB-200. In contrast, the mean AUC on the MMLU-ProX-Lite knowledge and reasoning dataset is only 0.510, and routing provides no consistent improvement over random selection because small models often remain highly confident while incorrect. Global confidence routing also creates language escalation gaps up to 30.33 percentage points, whereas equal-quota routing removes this disparity with little aggregate accuracy cost. These results show that confidence routing is task-dependent and can allocate computation unevenly across languages. Code is available at this https URL
Abstract (translated)
URL
https://arxiv.org/abs/2610.23249
PDF
-
Think More or Read More? A Token-Budget Theory of Reasoning with Shared Context
Abstract
Reasoning systems increasingly improve answers by incorporating additional context into the model's input, from multi-agent debate, where agents read their peers' responses, to self-refinement, where a model revisits its own drafts. In these systems, communication consumes context as one agent's output becomes another's input. We study the regime where token reading and generation draw on one fixed token window, so every token spent reading is a token that cannot be spent reasoning. We develop a \textbf{token-budget theory} that formalizes this trade-off, modeling accuracy as a function of the reasoning budget remaining after reading and aggregating votes with a Condorcet jury model that permits correlated errors. For multi-agent debate, the theory gives a closed-form prediction for the optimal number of agents. This quantity is estimated from single-agent measurements, allowing the optimal group size to be determined without running debate, and the same accounting extends to a heterogeneous set of different models. We empirically verify our theory and show that, at the predicted group size, our method outperforms strong baselines under the tightest calibration spend, where debate-calibrated rules can afford only one or two questions. Self-refinement obeys the same accounting, with its own draft in place of a peer message. The theory further predicts that communication pays only when the gain in the accuracy ceiling exceeds the reasoning room spent on reading, with this condition tightening as the window shrinks. Across five open models and four reasoning tasks, self-consistency matches or beats deployed debate at $93$\% of $980$ matched window-and-group-size settings, with the gap being largest under tight windows and nearly disappearing at a loose one, consistent with the theory. Under a shared window, how much room a model has left to \textit{think matters more than what it reads}.
Abstract (translated)
URL
https://arxiv.org/abs/2610.23246
PDF
-
MultiAgent-OPD: End-to-End Training of Hierarchical Multi-Agent Systems with On-Policy Distillation
2026-10-16 10:01:52Yuhan Sun, Jiaxin Zhao, Yongdong Luo, Yibo Hu, Meiguang Jin, Junfeng Ma, Weixun Wang, Weihang Pan, Zhongan BiAbstract
In a hierarchical multi-agent system, a Leader assigns tasks and subagents carry them out. Jointly training the Leader and the subagents is difficult because a final answer gives only indirect feedback to the subagents. An LLM judge can score the performance of the subagents, but these scores depend on the judge's own assessment of the quality of subagents' responses. We present MultiAgent-OPD, which combines on-policy distillation with reinforcement learning to jointly train a shared-parameter Leader-Searcher-Answerer team end to end. We first train a Teacher in a Searcher-Answerer workflow using answer correctness. During team training, the frozen Teacher scores student-generated search and summary tokens to provide local process supervision for the Searchers, while the Leader learns from team success and the Answerer from candidate correctness. Across seven datasets, the trained system reaches 38.68% macro avg@4, exceeding Prompted Critic by 4.06 percentage points and C-MAS + RL, the trained cascaded multiagent system, by 2.61 percentage points. The team also solves more questions within four attempts.
Abstract (translated)
URL
https://arxiv.org/abs/2610.23234
PDF
-
ReLaDrift: Recurrent Language Generation via Drifting for Test-Time Scaling
2026-10-16 09:55:22Yuki Otsuka, Daisuke Oba, Naoaki OkazakiAbstract
Drifting Models generate a sample with a single function evaluation, making them attractive for fast parallel language generation, but their one-step formulation does not support iterative refinement. We propose Recurrent Language Generation via Drifting (ReLaDrift), which uses a shared Drifting Model to generate an initial draft from noise and iteratively refine it in a continuous latent space. We train this recurrence by sampling the number of rounds and applying the drifting loss only at the final round, stopping gradients through preceding rounds so that only one round requires backpropagation. Unlike the one-step Drifting Model, whose accuracy does not improve when iterated, ReLaDrift enables a trade-off between test-time computation and generation quality, as diffusion and flow language models do. On Sudoku Hard and GSM8K, it achieves competitive or superior performance compared with these iterative-generation approaches across a range of inference budgets. Our code is available at this https URL
Abstract (translated)
URL
https://arxiv.org/abs/2610.23225
PDF
-
MAS-Bai: Linguistically Informed Adaptive Multi-Agent Translation for Bowen-Script Bai
2026-10-16 09:43:32Jingting Zheng (TJUNLP Lab, School of Computer Science and Technology, Tianjin University, China), Hong Ren (TJUNLP Lab, School of Computer Science and Technology, Tianjin University, China), Yuehan Cui (TJUNLP Lab, School of Computer Science and Technology, Tianjin University, China), Shaolin Zhu (TJUNLP Lab, School of Computer Science and Technology, Tianjin University, China), Deyi Xiong (TJUNLP Lab, School of Computer Science and Technology, Tianjin University, China)Abstract
Large Language Models (LLMs) remain unreliable for extremely low-resource languages whose writing systems conflict with high-resource pretraining priors. We study Bai-Chinese translation in the Bowen script, where Chinese characters are frequently borrowed phonetically rather than semantically, causing severe literal-translation hallucinations. We propose MAS-Bai, a linguistically informed adaptive multi-agent framework for Bowen-script Bai-Chinese translation. MAS-Bai routes each input through either a lightweight retrieval-translation path or a deeper philological path with etymological and syntactic agents, and uses cycle-consistency with a lightweight reverse translator as an inference-time reliability signal. We construct a 6,536-sentence Bai-Chinese corpus and evaluate on a 1,536-sentence test set. MAS-Bai achieves 16.23 BLEU and 62.15 COMET, improving over Standard RAG by 8.11 BLEU and 6.85 COMET. Forward LoRA, expert ablations and controlled routing experiments distinguish the roles of supervised adaptation, Bai-specific expertise and conditional routing. The largest gains occur on Bowen phonetic loans and divergent Bai word order, while human evaluation and reverse-verifier analyses support the automatic results.
Abstract (translated)
URL
https://arxiv.org/abs/2610.23205
PDF
-
AMEND: Constraint-Guided Multi-Agent Text Detoxification with Isolated Critics
Abstract
Text detoxification requires balancing toxicity removal, semantic preservation, and fluency. Existing methods typically optimize an unconstrained scalar objective---enabling safety scores to mask semantic loss---or rely on open-ended iterative debate prone to semantic drift and high computational cost. We propose AMEND, an inference-time framework that decouples detoxification into role-conditioned proposal generation, isolated multidimensional criticism, and constraint-first selection. AMEND evaluates diverse candidates with role-blind critics and enforces strict semantic and fluency feasibility before ranking by utility. To maximize efficiency, AMEND Lite restricts subsequent search to at most one targeted repair round only when residual toxic spans remain. On ParaDetox, one-round selection over diverse roles improves the joint evaluation metric by $+0.0508$ over the best single role. On a frozen confirmation test set, AMEND Lite achieves a joint score of $0.6597 \pm 0.0062$, outperforming unconstrained multi-round loops by $+0.0489$ ($p < 0.05$) while reducing inference requests by 28.4%.
Abstract (translated)
URL
https://arxiv.org/abs/2610.23183
PDF
-
What LLMs Do in the Shadows: Reading Context Reliance from the Unembedding Dark Space
2026-10-16 09:21:49Alexandra Kuleshova, Andrey Volodichev, Daria Kotova, Alexey ZaytsevAbstract
Detecting faithfulness hallucinations in large language models (LLMs) is crucial for their safe deployment in high-stakes domains. Recent approaches address this by contrasting contextual against parametric knowledge utilization, yet measuring context reliance typically demands noticeable overhead, such as materializing attention maps, running additional forward passes, or training auxiliary probes. In this work, we show that context reliance can be estimated by projecting hidden states onto the dark space - directions that compose weakly with the unembedding matrix. Beyond its recognized function as an initial-token attention sink, we show that this subspace actively mediates context processing, and that its share of hidden-state norm - termed the Dark Matter Ratio (DMR) - provides a direct score of context utilization. Pairing DMR with a parametric knowledge score yields DarkLens, a lightweight faithfulness detector operating strictly within a single forward pass. Evaluated across three QA and summarization benchmarks on six open-weight LLMs (4B-13B parameters), DarkLens matches or outperforms a wide range of baselines without requiring extra forward passes or probe training.
Abstract (translated)
URL
https://arxiv.org/abs/2610.23182
PDF
-
Do Speech BCIs Need Larger Models? Rethinking Neural Decoding beyond Scaling
2026-10-16 09:12:58Zehui Feng, Weichuan Wang, Xiaohan Chen, Cuntai Guan, Ting HanAbstract
Recent speech brain-computer interfaces (BCIs) increasingly rely on large models and complex multi-stage pipelines for neural decoding. However, in speech BCI, where training data are scarce, noisy, and expensive to collect, the benefits of model scaling remain unclear. This work revisits a fundamental question: are large models truly necessary for accurate neural decoding? We propose BEST (Brain-Encoding-Speech-Text), a two-stage framework for efficient neural-to-text decoding. Instead of relying on large generative models or phoneme-centric pipelines, BEST introduces hierarchical intermediate representations and a compact speech-oriented decoder that directly maps neural activity to text. This design preserves essential structure while significantly reducing computational cost, enabling robust generalization in low-resource settings and practical deployment under constrained resources. On the Brain-to-Text '24 and '25 benchmarks, BEST achieves word error rates (WERs) of 6.19% and 1.62%, respectively, using only 3.85% of the parameters required by comparable methods. Beyond these results, we provide three key insights: (1) decoder scaling yields non-monotonic gains; (2) cross-modal alignment is more critical than model size; and (3) overfitting degards sequence-level performance. Representation analysis further reveals a consistent neural-audio-text alignment pathway.
Abstract (translated)
URL
https://arxiv.org/abs/2610.23175
PDF
-
Chinese Decision: A Family of Language Models for Single-Pass Probabilistic Decision-Making
2026-10-16 09:08:32Qiang Yan (Institute of Computing Technology, Chinese Academy of Sciences)Abstract
Many language-based applications require probabilities over a finite set of decisions rather than open-ended responses. We present Chinese Decision, a family of Chinese-language models for single-pass probabilistic decision-making. Given context, an instruction, and candidate descriptions, the model extracts final-position candidate-letter logits in one forward pass and normalizes them over valid options, without autoregressive answer generation. A shared interface supports Boolean judgments, categorical choices, and ordinal scores. We report a completed 0.8B model trained with distribution cross-entropy on 13.048 million decisions and preliminary 2B checkpoints trained with an additional RLCD objective. On the full CJ-Bench of 307,900 decisions, the uncalibrated 0.8B model achieves 72.24% general-domain accuracy and 76.98% average accuracy across medicine, law, and finance using a single checkpoint, with 1.07% overall expected calibration error. The 2B model reaches 72.56% general-domain accuracy after 40.13% of one epoch but has not surpassed the completed 0.8B model on the professional-domain average. We further evaluate a validation-temperature-scaled deployment model and its Q8 quantization, finding close aggregate performance while documenting individual prediction changes. We release the training, evaluation, and deployment pipeline. The current results establish a reproducible approach to Chinese probabilistic decisions but do not isolate the effect of model scale.
Abstract (translated)
URL
https://arxiv.org/abs/2610.23170
PDF
-
Cross-Centaur: Modeling Individual Decision-Making Across Psychological Tasks with Language Models
2026-10-16 08:55:00Hikaru Asano, Hidezo Suganuma, Ryutaro Mori, Yukino BabaAbstract
Large language models (LLMs) are increasingly used as computational models of human behavior. Recent models such as Centaur have advanced this goal by fine-tuning LLMs on trial-by-trial decisions from psychological experiments. However, existing large-scale behavioral datasets do not link the same participants across different tasks, so these models learn population-level regularities but are not trained to model individual-level regularities across tasks. We introduce Cross-Psych-101, a dataset of 591 participants and 261{,}551 valid decisions across 12 Bandit, Tradeoff, Structure, and Social tasks. We train Cross-Centaur to predict a participant's target-task actions from a target-task description and the source-task description paired with the same participant's source-task behavior. Cross-Centaur outperforms off-the-shelf LLMs and Centaur on held-out participants, both within and across tasks. Probing analyses suggest that its hidden states encode participant-level behavioral tendencies that transfer across tasks. These results highlight the potential of cross-task individual-level data for predicting human decision-making with language models. Cross-Pysch-101 is available at this https URL.
Abstract (translated)
URL
https://arxiv.org/abs/2610.23157
PDF
-
Can AI interview? Comparing AI-led to Human-led Chat Interviews
2026-10-16 08:43:18Semra Guven, Tobias G{\aa}rdhus, Andreas Bjerre-Nielsen, Anna Rogers, Hjalmar Bang CarlsenAbstract
Large language model chatbots have been proposed as AI-interviewers for collecting qualitative interview data at scale, potentially bridging social science's longstanding divide between the breadth of surveys and the depth of qualitative interviews. This paper provides the first direct empirical comparison of data quality in AI-led and human-led interviews. In a randomized experiment (N = 41), university students were interviewed via synchronous text chat by either trained non-expert interviewers or a locally hosted AI system; we evaluate exposure, specificity, and relevance. Median interview duration was similar across conditions, but AI-interviewers asked more than twice as many questions per interview and elicited roughly 65 percent more respondent text per interview. Per-question responses were longer with human interviewers, though not after multiple-testing correction; only AI-condition participants skipped questions. Responses were almost uniformly on-topic in both conditions; for specificity we detect no difference, but the design cannot exclude one of up to three quarters of a point. These results were obtained with a small, locally hosted open-weight model (7B parameters), which addresses concerns about data security, GDPR compliance, and scientific reproducibility. Our findings support the viability of AI-interviewing for large-N qualitative data collection.
Abstract (translated)
URL
https://arxiv.org/abs/2610.23149
PDF
-
LED: Improving Masked Diffusion Language Models by Lookahead Editing
2026-10-16 08:29:07Yuyan Zhou, Zeyuan Yang, Yuxuan Song, Zhicheng Cai, Hongli Yu, Weiying Ma, Hao Zhou, James KwokAbstract
Diffusion large language models (dLLMs) have emerged as a promising alternative to autoregressive language models, enabling non-autoregressive text generation through parallel decoding. Existing dLLMs can be broadly categorized into two families: masked diffusion models (MDMs) and uniform diffusion models (UDMs) In this work, we theoretically show that MDMs are easier to optimize due to their mask-based transition schedule but have lower computational expressivity, whereas UDMs enable iterative state revision and greater expressivity but are substantially harder to train. This trade-off raises a natural question: can we obtain the expressivity of iterative revision without giving up the simple, easy-to-optimize transition schedule of MDMs? Motivated by this analysis, we propose Lookahead Editing (\ours{}), which augments masked diffusion decoding with a rewritable proposal state. At each decoding step, the model performs several lookahead updates to iteratively refine future predictions in this proposal state before committing tokens to the masked sequence. This design introduces a rewritable proposal that uses additional compute to improve expressivity while keeping the simple mask-based transition schedule of MDMs. We fine-tune Qwen2.5-7B and Qwen3-8B to obtain LED-\texttt{7B} and LED-\texttt{8B}, respectively. Across eight benchmarks, LED SFT yields average gains of $+11.4$ and $+7.7$ points over standard MDM SFT. Moreover, using SFT alone, LED-\texttt{8B} outperforms LLaDA2.1-16B Q-mode, which is trained with continued pre-training, supervised fine-tuning, and reinforcement learning.
Abstract (translated)
URL
https://arxiv.org/abs/2610.23122
PDF
-
Disentangling Language Effects on Tool Admission and Post-Retrieval Safety
2026-10-16 08:05:14Aditya Nawal, Manit Baser, Mohan GurusamyAbstract
LLM agents increasingly use retrieval tools across languages, yet their safety evaluation remains English-centric. It remains unclear whether language affects tool admission, post-retrieval generation, or both. We independently vary request, retrieved-page, and required response languages in a controlled URL-retrieval setting across ten models and seven languages. We identify Safety Loss through Interlingual Processing (SLIP). Non-English harmful requests increase mean tool admission from 46.9% to 67.4%, with a positive mean change in every model. This pattern also persists in an autonomous retrieval setup across all four models tested in this setting. Requiring answers in the page language increases pooled harmfulness on matched tool calls from 50.3% to 57.4% across four closed-weight models, mainly in Chinese and Hindi. Translating pages alone produces no consistent harm increase in eight models with comparable inference settings. Automated audits show that a non-harmful response may be a genuine safety refusal, a deflection citing language difficulties, or an unusable answer. Harmfulness labels alone cannot distinguish these outcomes. Multilingual safety evaluations should therefore measure tool admission and response harmfulness separately and distinguish refusal from capability failure.
Abstract (translated)
URL
https://arxiv.org/abs/2610.23093
PDF
-
Before Retrieval Begins: Safety and Tool-Use Decisions in LLM Agents
2026-10-16 07:42:47Aditya Nawal, Manit Baser, Mohan GurusamyAbstract
Large language models increasingly act as agents that call external tools, yet most tool-safety research examines only what follows a tool call. The prior decision of whether to call the tool remains largely unexamined and can influence downstream safety. Across the seven open-weight models, the unweighted model-average harm rate after admitted harmful requests is 74.5%. To identify what governs this decision, we apply a factorial intervention to the input request. Within our generated construction, neutralizing the requested-operation component produces the largest admission contrast (19 pp). A linear direction estimated from this contrast influences held-out admission in four of seven models; effects on operation-neutralized requests in some models show that it is not an operation-specific mechanism. We then test whether written refusal is behaviorally and representationally coupled to the same decision. Written refusal accompanies 33.6% of admitted harmful calls. Plain-text refusal training reduces unsafe admission in models that retain benign tool use, showing that response-level safety training can transfer to the call decision. Together, these results establish a partial, model-dependent relation between written refusal and tool admission. They motivate direct evaluation and supervision of tool admission rather than treating text refusal as its proxy.
Abstract (translated)
URL
https://arxiv.org/abs/2610.23071
PDF
-
ParetoChat: An LLM-Based Interface for Interactive Multi-Objective Optimization
2026-10-16 07:39:39Gokhan CeyhanAbstract
Multi-objective optimization models are often deployed in settings where decision makers (DMs) must balance some conflicting criteria. In practice, however, stakeholders must translate their qualitative preferences between different criteria into numerical weights, priorities, or reference points required by optimization methods, which can be difficult to specify and interpret. This paper proposes \emph{ParetoChat}, an LLM (large language model)-based expert-agent interface for interactive multi-objective optimization. The agent translates natural-language user input into structured scalarization parameters, invokes an optimization system, presents candidate solutions, and iteratively elicits feedback. We describe a multi-agent architecture that separates input processing, multi-criteria decision-making (MCDM) expertise, solution presentation, and feedback processing. We then present a simulation study using an LLM-based DM agent and benchmark ParetoChat against interactive methods implemented in DESDEO, including reference-point and NIMBUS methods. The computational experiments on multi-objective binary knapsack and set-partitioning instances show that ParetoChat can guide the search efficiently toward solutions that match the DM's preferences. An application to historical courier-order assignment instances illustrates how users can explore objective extremes and intermediate trade-offs through natural-language requests. The results indicate that LLM agents may provide a practical interaction layer between business stakeholders and optimization tools.
Abstract (translated)
URL
https://arxiv.org/abs/2610.23062
PDF
-
Same Person, Different Advice: Auditing LLM Financial Advice with Repeated Risk-Tolerance Responses
2026-10-16 07:39:08Soovadeep Bakshi, Apurba Bose, Aishwarya Bose, Sandip NandiAbstract
AI financial advisers set a client's stock allocation largely from one stated-risk question, yet people do not answer it the same way twice: in a national US panel, 41% of households chose a different answer two years later, and in a laboratory retest a third of people moved two or more points on a 0-10 scale after six months. Seven large language models from five providers gave one-turn allocation advice for 800 real household profiles in which only the risk answer changed. The recommended stock share moved by 10 to 16 percentage points on average, and by 24 to 35 when the answer changed, against 0.6 to 4.2 points for the same prompt asked twice; in the laboratory retest the rating alone moved advice by 11 to 18 points. Without any note, the models respond to the answer about as strongly as when told that it is very reliable (less so for the two Claude models). If that instructed response is how each model maps attitude to advice, discounting the answer for its measured reliability, as classical test theory prescribes, would cut the swings by a factor of 2.4 to 3.3. The swing is largest for clients under 40. A note that the answer is noisy steadies some models, but shifts their weight to the client's age and pulls cautious clients towards a generic middle allocation: clients unwilling to take any risk are advised 22% stocks instead of 4%. Three extra money questions reduce the swing for some models, with no statistically detectable change relative to how much advice differs across people. In a scripted test, one clarifying question removes nearly all of the swing from an answer the client then corrects (33.5 points to 1.9 on average) while keeping the response to a change the client confirms. We argue that same-person retest audits should be a standard check for AI advisers, and that what to do with an unreliable answer is a question of client autonomy as well as of statistics.
Abstract (translated)
URL
https://arxiv.org/abs/2610.23060
PDF
-
ECHO: Replacing the Compliance Proxy with Response Harmfulness in Optimization-Based Jailbreak Attack
2026-10-16 07:07:08Yuanwei Wu, Lingdong Huang, Yunlong Mao, Kanye Ye Wang, Jingyu Hua, Sheng ZhongAbstract
Despite extensive safety alignment, large language models remain vulnerable to jailbreak attacks. Existing optimization-based attacks share a single objective: maximize a token-level proxy of victim compliance---such as the likelihood of an affirmative prefix---computed on the prompt rather than the decoded reply. This proxy cannot distinguish a cautious hedge from a fully actionable harmful answer, leaving the search blind to what the model actually says. We propose ECHO (Evaluation-Conditioned Harmfulness Optimization), which replaces the compliance proxy with response-level, continuous harmfulness scores from HarmSig, a 1.7B regressor grounded in actual harmful capability. ECHO couples HarmSig with five word-level perturbation operators and Keyword-Anchored Beam Search with global top-m memory, under a strict black-box threat model. Across five open-weight victims on AdvBench, ECHO achieves up to 0.898 StrongREJECT mean score and 89.23% HarmBench success. Our code is publicly available at this https URL.
Abstract (translated)
URL
https://arxiv.org/abs/2610.23030
PDF
-
Controlling the Diversity-Quality Trade-Off of LLMs by Merging Models Before and After Alignment
2026-10-16 06:24:10Yasuaki Sumita, Tomoharu Iwata, Toshiyuki TanakaAbstract
Alignment is known to change the output characteristics of large language models (LLMs). Base models tend to generate more diverse outputs, whereas aligned models tend to produce higher-quality outputs. The appropriate balance between the two depends on the task. Existing methods either require costly fine-tuning or, like temperature sampling, mainly improve lexical diversity rather than semantic diversity. We introduce a depth-aware FrankenMerging approach for controlling the diversity--quality trade-off between an aligned model and its base model without additional training. This approach combines the lower layers of an aligned model with the higher layers of its base model based on the hypothesis that lexical diversity is more closely associated with lower layers and semantic diversity with higher layers. Experiments showed that varying the number of layers taken from each model adjusts both diversity and quality. In most cases, the merged models exhibited higher semantic diversity than the aligned model, and at some split positions, they showed higher lexical diversity than both original models. They also demonstrated a better diversity--quality trade-off than temperature sampling in most settings. A supporting ablation, in which we reversed the merging order, yielded results largely consistent with the hypothesis.
Abstract (translated)
URL
https://arxiv.org/abs/2610.22999
PDF
-
Self-Evolving Agents Can Repair Failures but Struggle to Generalize from Them
2026-10-16 06:10:01Weixiang Zhao, Yichen Zhang, Yanyan Zhao, Jinming Bai, Bin Liang, Ruifeng Xu, Bing Qin, Ting LiuAbstract
Experience-driven self-evolution offers a promising path for large language model (LLM) agents to continually improve through interaction. Yet existing approaches often rely on successful trajectories or external evaluation signals, leaving unclear whether agents can evolve when such positive anchors are unavailable. We study this question through Cold-Start Failure-Only Evolution (CFOE), where agents start without accumulated experience and learn only from failed interactions without privileged guidance. Across three self-evolving paradigms, multiple LLM backbones, and four interactive environments, we uncover a consistent repair-generalization gap: failure-derived experience can help repair previously failed tasks, but rarely yields sustained improvements on unseen tasks. Further analysis shows that agents often misinterpret failures, while even verified-correct experience remains difficult to transfer. Relaxing CFOE reveals that verifier feedback and additional interaction provide limited gains, whereas successful trajectories substantially improve transfer. These findings suggest that current self-evolving agents are better at repairing observed failures than building generalizable capabilities from failure.
Abstract (translated)
URL
https://arxiv.org/abs/2610.22984
PDF
-
KiranaBench: Can a Language Model Run a Neighbourhood Shop?
2026-10-16 06:02:00Bhuvan Arora, Ankita Karmakar, Chaitanya Jain, Dilesh Churiwal, Harshit Jain, Shivansh Manish Dubey, Manish Prasad, Raja Piyush, Kshitij Tayal, Dhruv Kumar, Jagat Sesh Challa, Raunak Manekar, Yash SinhaAbstract
Large language model (LLM) agents are increasingly evaluated on long-horizon tasks, including autonomously running a business. Existing benchmarks for this setting are almost exclusively English-language, Western in retail context, and single-agent, with the agent observing the business directly. We study how these results transfer to a multi-agent Indic setting. We introduce KiranaBench, a simulated Indian general store in which four roles (manager, cashier, helper, and bookkeeper) are each played by a separate LLM agent. Interactions with customers and suppliers occur in English, Hindi, or Hinglish, and product demand is calibrated to India's national Household Consumption Expenditure Survey. Evaluating four open-weight LLMs, we find that all fall far short of a fixed-rule shopkeeper baseline: the best model earns only one-eighth of the baseline's profit, and two models finish below an unmanaged shop. Performance is also consistently lower for Hindi and Hinglish than for English supplier communication. KiranaBench offers evidence that current LLM agents are not yet reliable for autonomously operating small businesses in India, and provides a testbed for developing agents suited to settings that are long-horizon, collaborative, and multilingual.
Abstract (translated)
URL
https://arxiv.org/abs/2610.22980
PDF
