arXiv_CL
-
Beyond Small Talk: Modeling Disclosure Trajectories in Human-AI Conversation
2026-10-21 08:36:47Keith Kwok, Wanting Ning, Tej Deep Pala, Desmond C. Ong, S. Supraja, Renwen Zhang, Soujanya PoriaAbstract
Recently, we see a rise in conversational agents as a form of social and emotional support, where disclosure will unfold over time. Yet, current evaluations of these agents typically examine isolated conversational states. In analyses of naturally occurring conversations, disclosure history and current disclosure are confounded which tends to make their effects difficult to isolate. We thus introduce Trajectory-based Disclosure Evaluation (TraDE), an evaluation framework grounded in Social Penetration Theory that utilizes predefined disclosure trajectories and evaluates how preceding disclosure history shapes model behavior under matched current disclosure. Through the use of TraDE, we demonstrate across multiple models and relational behavior benchmarks that responses to disclosures matched on depth can differ depending on the preceding trajectory, particularly in whether models accept or challenge the user's framing. We also observed that these differences track accumulated exposure to deeper disclosure rather than any single moment of vulnerability or the order of disclosures. It also tends to persists for several turns after disclosure returns to shallow levels. For several behaviors, we also demonstrate that models diverge most consistently from human supporters in the Emotional Support Conversation corpus (ESConv) not when users open up but rather when they step back in their disclosure. TraDE thus provides a systematic way to expose history-dependent relational behavior that evaluations of isolated conversational states would otherwise miss.
Abstract (translated)
URL
https://arxiv.org/abs/2610.29413
PDF
-
ECHO: Emergent Common-Mode Hidden Offsets Drive Vocabulary Collapse in Preference Optimization
2026-10-21 08:35:39Yun Li, Ehsan Javanmardi, Simon Thompson, Jiachen Gong, Qunli Zhang, Zifan Zeng, Shiming Liu, Peng Wang, Zixuan Guo, Manabu TsukadaAbstract
Preference optimization can improve every ranking metric while building a hidden, shared representation shift that the relative loss barely sees, one that can eventually destroy generation. We trace this failure to the common-mode offset, a shared directional shift in the model's hidden states that appears across token positions, prompts, and both chosen and rejected responses. Because it moves chosen and rejected representations together, the relative ranking loss barely reacts to its accumulation. As this offset grows, it projects onto the output layer to form a vocabulary well, elevating a single token's probability until it dominates decoding and causes generation collapse. The isolated offset alone reconstructs the vocabulary well, injecting the learned offset into an untreated base model recreates the collapse where norm-matched random directions do not, and masking this common component from the loss prevents the well in our flagship setting. Under the per-token ranking loss we study, the same collapse appears in 14 of 17 base-model lineages and under full-parameter fine-tuning. Moreover, the offset is not merely an artifact of collapse, since it is already present in textbook DPO and SimPO runs that form no well. Textbook DPO keeps it weakly aligned with any single vocabulary direction even with three times our base budget. Trained from base models for that budget, however, full-parameter SimPO drives it into collapse on both Llama-2 and Llama-3.1, and the same kind of offset grows, without a well, from a public SFT checkpoint under the official SimPO recipe. Across 17 UltraFeedback runs, the better a run ranks held-out responses, the less often two public reward models prefer its samples to the base model's. These findings expose a blind spot in standard ranking metrics and introduce a representation-level signal that can warn of generation collapse before it occurs.
Abstract (translated)
URL
https://arxiv.org/abs/2610.29410
PDF
-
Right On Q: Recovering Reasoning under Quantization via Self-Generated Trajectories
2026-10-21 08:03:24Boyuan Zheng, Jing Xiong, Chenwei Wu, Hui Shen, Zesen Zhao, Hei Ting Una Chan, Xueshen Liu, Zhen Zhang, Liyue Shen, Zechun Liu, Z. Morley MaoAbstract
Quantization-aware training (QAT) aims to recover reasoning capabilities in low-bit large language models (LLMs). Existing recipes typically supervise the quantized student with the original full-precision model on fixed response sequences. However, the resulting low-bit quantized models still struggle on long-horizon reasoning tasks and often fail to terminate within a token budget that suffices for the full-precision model. We identify two challenges in closing this gap. First, fixed-response supervision leaves a mismatch between training prefixes and those encountered during student generation, and this trajectory divergence grows as responses lengthen. Second, on its own prefixes, the student underpredicts some next-token candidates and overpredicts others, and this distributional misalignment becomes more frequent at high-entropy prefixes. We present OnQ, a low-bit recovery pipeline that combines on-policy distillation and candidate-wise Dynamic KL. Starting from block-wise calibration, the quantized student generates responses and receives full-precision teacher supervision on its own prefixes. Dynamic KL weights forward and reverse KL by each candidate's teacher-student probability ratio to balance correction of under- and overprediction. Across Qwen3-1.7B, 4B, and 8B at 2- and 3-bit precision, OnQ achieves the highest nine-benchmark average, with the largest gains on long-horizon reasoning benchmarks. At 2 bits, it outperforms every baseline on every benchmark at every model size, improving the Qwen3-8B average by 10.26 points over the strongest baseline. It also produces shorter responses with fewer truncations and repetitive errors.
Abstract (translated)
URL
https://arxiv.org/abs/2610.29378
PDF
-
Memory-Centric Continual Learning for AI Agents
2026-10-21 07:49:42Dongfang Li, Zhuoen Chen, Bonian Jia, Zixuan Liu, Zhengxuan Lu, Can Deng, Yukun Shi, Yuhang Meng, Xin Zhang, Jiahe Huang, Fuhao Li, Junmai Wang, Shunyao Wen, Jingjun Sun, Jiaxu Tian, Jiangling Zhang, Zihan Wang, Liangyou Li, Yitong Yao, Shuangyong Song, Beilun Wang, Pengxu Wei, Hanlin Xu, Chang Liu, Yongxiang Li, Baotian Hu, Xiangyang Ji, Min ZhangAbstract
Autonomous agents powered by foundation models can reason, plan, and use tools, but their fixed weights and context limits hinder sustained learning from experience over long-term deployment. While external memory systems attempt to bridge this gap, they mostly act as passive retrieval caches, evaluated primarily by whether they can recall stored facts or past interactions. Whether memory can drive genuine continual learning remains an open question. In this paper, we introduce Memory-Centric Continual Learning (MCCL), arguing that memory should be judged by whether it improves future actions, not merely what it retrieves. MCCL has three main parts. First, a three-axis taxonomy organizes memory by where state is stored (substrates), what form it takes (assets), and how it is managed (policies). This defines a two-loop lifecycle that updates memory state in an inner loop while refining management policies in an outer loop based on task feedback. Second, an evaluation protocol measures retention under interference, adaptation to change, and task transfer to isolate genuine behavioral gains from simple context expansion. Third, a memory scaling framework models how agent performance scales with experience, storage capacity, and compute budgets across non-parametric and parametric memory. Together, these components reframe agent memory from passive lookup to continuous learning, offering a principled route toward self-improving autonomous agents.
Abstract (translated)
URL
https://arxiv.org/abs/2610.29357
PDF
-
Don't Stop Long-Context Continual Pre-Training Too Early: An Empirical Scaling Law
2026-10-21 07:46:55Ziqing Qiao, Xiaoyue Xu, Sikui Zhang, Yinuo Xu, Bingxiang He, Chaojun Xiao, Xu HanAbstract
Long-context continual pre-training (LCCP) is widely used to improve long-context capability, but how its benefits scale with the number of tokens used for LCCP remains unclear. We investigate this question through modeling, monitoring, and validating the benefits of scaling LCCP. First, we establish an empirical additive power law relating long-context performance to model size and the respective token budgets of short-context pre-training and LCCP. Experiments across five model scales and more than one hundred short- and long-context budget configurations show that the fitted law reproduces the observed trends and extrapolates well to held-out model scales. Second, we introduce \textsc{Constant-Last4}, a training and observation protocol that averages the four most recent consecutive checkpoints from a constant-learning-rate run to enable stable monitoring along a single extensible training trajectory. Under this protocol, the same scaling-law form fits well and accurately predicts the LCCP performance trajectory of a 1.22B-parameter model trained on approximately 100B LCCP tokens. Finally, downstream evaluations along this trajectory show continued improvements in long-context performance, with no clear sign of saturation within the evaluated range. These gains persist after both 128K context extension and supervised fine-tuning, with benefits extending to long-context understanding and agentic tasks. Together, these results argue for continuing to scale LCCP rather than stopping early: its gains are stably observable along an extensible trajectory, quantitatively predictable, and transferable to downstream tasks.
Abstract (translated)
URL
https://arxiv.org/abs/2610.29352
PDF
-
Closed-Loop Dual-Reference Guidance for Accent Control in Cross-Lingual Voice Cloning
2026-10-21 07:46:22Yuanzhong Zheng, Wei Huang, Ming Ma, Kailun Jin, Youhai JiangAbstract
Zero-shot cross-lingual voice cloning often carries the reference speaker's accent into the target language. We propose closed-loop dual-reference guidance, which pairs the original reference with a reusable native anchor constructed by voice conversion (VC). The references initialize separate language-model (LM) states that guide a shared acoustic trajectory; each completed patch then updates both states, keeping their generated histories aligned with the emitted speech. Anchor construction runs once per speaker, and online synthesis uses a frozen TTS model without per-utterance VC. On 20 held-out French speakers, the fixed setting $\lambda^{\!*}=0.33$ reduces residual French LID mass by 73\% versus direct cloning, at an ECAPA similarity cost of only $0.042$; relative to anchor-only cloning, it recovers $0.155$. The method achieves the lowest mean WER ($0.009$) and highest mean GOP ($4.750$) among the held-out test arms. In blinded preference tests, native-English listeners favor it over direct cloning on pronunciation naturalness, and over both anchor-only cloning and a TTS$\to$VC cascade on speaker identity. Online RTF is approximately 26\% lower than the cascade's ($0.59$ vs.\ $0.80$), and the same setting reduces Spanish LID mass by 56\% without re-tuning.
Abstract (translated)
URL
https://arxiv.org/abs/2610.29351
PDF
-
B-CARE: Blind Classwise Assessment of Relative Evidence for Stance Error Ranking
2026-10-21 07:43:27Meysam Sabbaghan, Doina CarageaAbstract
Stance classifiers make errors that can be costly to identify under limited human-review budgets. Prioritizing potentially incorrect predictions remains challenging, particularly across heterogeneous base models. We introduce B-CARE (Blind Classwise Assessment of Relative Evidence for Stance Error Ranking), a prediction-blind framework for identifying high-risk stance predictions. Given only a target and text, a fixed LLM verifier assigns ordinal evidence scores to all candidate stance labels without observing the base model's prediction or confidence. B-CARE constructs a comparative risk score by contrasting the strongest alternative-class evidence with the evidence supporting the predicted class. The resulting classwise evidence profile can be reused across multiple base models without training a separate error detector or calibrating risk scores. We evaluate B-CARE on four stance benchmarks and three heterogeneous base LLMs, covering 12 dataset-model settings and 5,206 unique test instances. Evaluation includes error-ranking discrimination (AUROC and AUPRC), selective prediction (E-AURC), and review effectiveness (ErrorCapture at fixed review budgets). B-CARE achieves a mean AUROC of 0.735 and a mean AUPRC of 0.524, capturing 41.06% of errors within the highest-risk 20% of predictions. Although B-CARE does not uniformly outperform strong confidence-based baselines, it provides a reusable, complementary evidence signal for prioritizing potentially incorrect stance predictions across heterogeneous models.
Abstract (translated)
URL
https://arxiv.org/abs/2610.29350
PDF
-
Data Schedules and Learning Rate Schedules Should Be Designed Together
2026-10-21 07:42:20Haodong Wen, Kairong Luo, Tingkai Yan, Wenguang Chen, Kaifeng LyuAbstract
Mid-training (MT) is a training phase in LLM training pipelines that continues pretraining (PT) on smaller datasets of higher value. Taking the PT and MT phases as a whole, training usually starts with only PT data and then transitions completely to MT data. In this work, we view this complete transition from PT to MT data as a special case of a broader class of data mixing schedules that specify how the proportions of PT and MT tokens in each batch change over training. Under fixed total training compute and MT token budgets, we ask: Is this complete transition to MT data optimal for minimizing final MT loss? We show that the answer depends on the learning rate (LR) schedule: one data schedule may outperform another under one LR schedule, but changing only the LR schedule can reverse the outcome. Specifically, when the LR decays throughout continued training, introducing MT data earlier and mixing it with PT data lowers the final MT loss compared with placing all MT data at the end of training, but this result is reversed if we use constant LR. We therefore study how to jointly choose when to introduce MT data and when to decay the LR through ablation studies, and find that it is better to decay the LR only after MT data enters training. Finally, we develop a theoretical model that explains why the optimal data schedule for a fixed MT token budget depends on the LR schedule.
Abstract (translated)
URL
https://arxiv.org/abs/2610.29348
PDF
-
Decodability Is Not Controllability: Local and Anisotropic Speaker Control in Continuous-Latent TTS
2026-10-21 07:38:25Yuanzhong Zheng, Wei Huang, Ming Ma, Kailun Jin, Youhai JiangAbstract
High linear decodability does not entail causal controllability in continuous-latent text-to-speech (TTS). In VoxCPM2, position-matched probes and equal-norm interventions across 30 speakers and 60 directed dyads expose this separation: reference-audio states at the final residual acoustic language model layer (RALM L7) retain 0.970 probe accuracy, yet aligned edits yield 0/60 target-dominant dyads. Edits at early reference-audio positions produce the opposite pattern. At L0, pair-aligned updates increase target similarity by 0.326 and achieve target dominance in all 60 dyads, against at most 1/60 under any equal-norm direction or position control. Independent speaker encoders, retrieval against held-out speaker prototypes, and double-blind listening confirm that the shifted voice is heard as the intended target. Cross-architecture interventions in this http URL distinguish backbone leverage from direct acoustic control: backbone edits displace the source without target attraction, while x-vector ablation reduces source similarity from 0.849 to 0.006. These findings identify two requirements for training-free speaker control: geometric specificity and a computational route that carries the intended speaker condition into acoustic synthesis.
Abstract (translated)
URL
https://arxiv.org/abs/2610.29345
PDF
-
Uncensored Is Not Unbiased: Abliteration Removes Refusals but Not Script-Dependent Bias About Taiwan
Abstract
"Abliterated" ("uncensored") builds of open-weight language models remove the activation direction that mediates refusal, and are popular partly on the assumption that a model that no longer refuses is also free of political censorship. We test this on two matched pairs of 2-bit models: NVIDIA's Nemotron 3.5 Bolt 74B-A9B and Qwen3.8-Flash-Next, each against a community abliteration. Each model judged 22 statement pairs (on the Republic of China (Taiwan), 1989 Tiananmen, control countries) in Traditional Chinese, Simplified Chinese (a character-level conversion) and English, and answered 27 open questions per script. Script mattered far more than abliteration: the original Bolt scored +0.58 on Index S (-1: People's Republic of China (PRC) position; +1: ROC position) in Traditional Chinese, +0.57 in English and +0.05 in Simplified Chinese, and affirmed "Taiwan is part of the PRC" with probability 0.91 in Simplified but 0.06 in Traditional characters; abliteration changed its scores by at most 0.06, mostly by raising "yes" on both sides. In Qwen, abliteration removed every refusal on 1989 and raised scores by 0.21 to 0.34, yet the abliterated Qwen still affirmed that Taiwan is part of the PRC in all three scripts. Naming the state the Republic of China (Taiwan) raised every model's score by 0.10 to 0.33 but left the script gap. Control pairs showed no script gap; each abliteration changed only its declared tensors. In 740,000 web documents, Simplified text on Taiwan contained PRC-framing terms 3.7 times as often as Traditional text. A script direction could not be steered away; a contrastive stance direction raised scores in every script, and its negation lowered them. A four-sentence system prompt raised the abliterated Qwen to +0.93 to +0.97. Keyword scoring rules failed the models closest to the ROC position mostly for attributing or denying the PRC claim. Uncensored is not unbiased.
Abstract (translated)
URL
https://arxiv.org/abs/2610.29342
PDF
-
Reader--Writer Coupling in Transformer MLPs
Abstract
A transformer MLP gives each hidden channel two roles: a reader that decides when the channel responds, and a writer that decides what it adds back to the residual stream. Read as a key--value memory, the layer can be measured one factor set at a time: the pairing of readers to writers is carried only by their shared index, invisible to any quantity computed from one factor set alone. Yet training couples the two factors through their joint update, so we ask what reading them together shows. The block supplies the joint instrument: the MLP writes back into the very stream it reads, so the exact Jacobian of its update is square, and its agreement with its own transpose resolves into reader--writer terms whose inner products give each channel, read as a couple, a signed pairing coordinate. Ranked by this coordinate, each block we measure has a small core of strongly coupled channels; where initialization gives either sign equally, the core is nearly all opposed: each writer points against the direction its reader responds to. This opposed core recurs from the smallest models trained from scratch to released checkpoints; its sign is in place within a few percent of training and holds while the coupling grows. Read as couples, most channels return their contribution against the reader that fired it, and the readers of later blocks take it up; few show in the logits. Training holds the couples in place: flipping the opposed writers moves the output toward common tokens, and a writer permuted onto another reader regrows to fit that reader while the reader stays.
Abstract (translated)
URL
https://arxiv.org/abs/2610.29340
PDF
-
ChronoSynth: Constraint-Aware Question Synthesis over Temporal Knowledge Graphs
2026-10-21 07:35:17Feng Liang, Weixin Zeng, Runhao Zhao, Mingjun Zhou, Shaoshi Yang, Xiang ZhaoAbstract
Knowledge graph question generation (KGQG) generates natural-language questions from structured graph evidence and a target answer. When the input graph is temporal, a valid question must also preserve the answer field, time condition, reference event, and temporal order that determine the answer set. Existing KGQG methods and prompting baselines can produce fluent questions that alter these conditions, causing temporal semantic drift. This paper studies temporal KGQG as a constraint-preserving graph-to-question problem and proposes ChronoSynth, a memory-augmented and constraint-aware generation framework. ChronoSynth first constructs a Temporal Intent State to make the intended question semantics explicit. It then retrieves compatible typed question patterns from an indexed pattern memory, adapts them to the input graph, and validates the generated question. Validation feedback guides revisions when a supported inconsistency is detected. Experiments on CRONQUESTIONS and MultiTQ demonstrate state-of-the-art generation performance, with ablations supporting the contributions of its components.
Abstract (translated)
URL
https://arxiv.org/abs/2610.29339
PDF
-
Policy Central: Detecting Conflicts, Redundancy, and Similarity in Deployed Police Policy
Abstract
Policy Central is a deployed system that helps police departments find relationships between sections of their written policy: whether two sections conflict, are redundant, are merely similar, or are different. We describe the system and compare two architectures. Retrieve-Judge, a retrieve-then-classify pipeline, runs as a cheap library-wide health signal; Screen-Confirm, a screen-then-confirm cascade, is a selective audit of the documents a reviewer chooses, checking them exhaustively against one another. On a 302-pair benchmark drawn from one department's ~84 real policy documents, given each pair Screen-Confirm reaches 0.795 aggregate recall and 0.885 on Conflict, so current models judge well. Finding the pair is the hard part: Retrieve-Judge's composed recall falls to 0.457, and Screen-Confirm recovers 0.472 (0.615 on Conflict). Today these pairs are found by reading the whole library, nearly 20,000 sections, checked against one another and re-checked after every revision. A scan that returns half of them in minutes, at low cost, and can be re-run on demand changes the economics of policy review even before recall improves.
Abstract (translated)
URL
https://arxiv.org/abs/2610.29338
PDF
-
Seeing the Forest for the Trees: Emergent Agreement from Individual Conversations
2026-10-21 07:27:08Adi Simhi, Yaniv Nikankin, Jonathan Herzig, Idan Szpektor, Yonatan Belinkov, Shay B. CohenAbstract
As LLM agents interact in the wild and shape each other's decisions, understanding how they converge toward agreement becomes important. We study this in multi-turn conversations where two LLMs discuss subjective matters, can agree or disagree, and may change their stances. Using a linear classifier on individual conversation activations, we identify an internal representation of agreement. We find that it varies substantially across turns and conversations and does not reliably track external agreement at the individual conversation level. We therefore move beyond the representation of agreement in individual conversations, by aggregating the agreement representation across conversations. Externally, this aggregate view reveals a simple structure: when pooled across conversations, agreement can be predicted well with a Markov process, with agreement at each turn being a direct function of agreement at the previous turn. Similarly, when internal agreement is aggregated, it couples with the external Markov agreement through a power-law mapping, exposing Markovian dynamics in internal representations. Additionally, we find that after a change of stance, models express agreement externally before their internal representations reflect the same shift. These results reveal a joint structure between internal and external agreement that emerges only when internal features are aggregated across conversations. Code available at this https URL.
Abstract (translated)
URL
https://arxiv.org/abs/2610.29328
PDF
-
Signals and Queries: Token-Efficient Agentic Context Discovery
Abstract
User-facing LLM agents must identify which behavioral policies apply at each turn, but evaluating every policy with an LLM is costly. We introduce Signals and Queries (S&Q), a lightweight first-stage relevance filter developed for Parlant 4.0 that removes at least 30% of candidate policies before downstream evaluation while preserving high recall. For each policy, an LLM generates activating examples and hard negatives offline. An independent logistic classifier trained on their embeddings then predicts relevance using embeddings and vector dot products, with no per-turn LLM calls. This shifts expensive relevance evaluation to a smaller candidate set, while allowing each policy's classifier to be updated independently. Evaluated against 9,144 LLM-judge relevance labels covering 254 $\tau^2$-bench conversation prefixes and 36 airline policies, S&Q achieves higher precision than top-K and cosine-threshold retrieval at every evaluated recall level, with consistent results across two embedding providers. These findings support inexpensive, per-policy classification as a practical first stage for reducing downstream evaluation workload in customer-facing agents.
Abstract (translated)
URL
https://arxiv.org/abs/2610.29326
PDF
-
OSynth: A Longitudinal, Context-Sensitive Evaluation of Meeting Understanding
2026-10-21 07:21:54Ian Porada, Suraj Tripathi, Chaitanya Animesh, Leda Sari, Chaitanya Narisetty, Austin R. Matthews, Bohan Yu, Ramesh L. Thulasiram, Gregory Sell, Kean Chin, Dmitriy GenzelAbstract
Meeting notetakers are increasingly deployed in settings where longitudinal context is available including calendar metadata, prior meetings, emails, slides, and speaker enrollments; however, meeting corpora remain limited in terms of size, domain coverage, and available context. We present OSynth, a large-scale synthetic dataset of simulated organizations comprising 570 hours of meeting audio together with associated artifacts and a latent ground-truth state. In a pairwise comparison, human raters distinguish OSynth meetings from real meeting transcripts with only 56% accuracy based on excerpts of 150-300 words. Across state-of-the-art LLMs, additional context yields significant gains on several entity-sensitive dimensions, demonstrating limitations of contemporary approaches and the potential usefulness of OSynth in developing better-contextualized meeting-understanding systems.
Abstract (translated)
URL
https://arxiv.org/abs/2610.29319
PDF
-
LLM-Based Exogenous World Modeling for Dynamic Scheduling
2026-10-21 07:01:25Ni Zhang, Haofeng Yuan, Jianan Zhou, Shiji Song, Zhiguang CaoAbstract
Dynamic scheduling requires dispatching online arrivals while anticipating future jobs, as poor decisions can compound downstream costs under nonlinear tardiness objectives. We show that a pretrained large language model (LLM) can serve as the exogenous component of a structured world model, forecasting uncertain future job arrivals. Specifically, we exploit an afterstate decomposition, which separates known deterministic shop transitions from uncertain exogenous job arrivals, enabling the LLM to forecast only the latter through a compact belief representation. Building on this, we develop Afterstate World-Model Planning (AWMP), which combines an evolved, belief-conditioned dispatching strategy with short scenario-based rollouts. A surprise-triggered refresh mechanism queries the LLM only when newly observed arrivals deviate sufficiently from previous forecast, and online residual self-calibration further conditions the LLM on its recent forecast errors to handle nonstationary arrivals. Across diverse arrival distributions, AWMP reduces weighted tardiness by 6-21% on dynamic job-shop and 22-46% on flexible job-shop benchmarks, relative to the best of eight classical priority-dispatching rules.
Abstract (translated)
URL
https://arxiv.org/abs/2610.29308
PDF
-
ExpEvolve: Structuring and Reusing Experience for Open-Ended Discovery
2026-10-21 07:00:23Yongqiang Chen, Chenxi Liu, Deyu Zou, Zhenhao Chen, Tongliang Liu, Bo Han, Kun ZhangAbstract
Evolutionary coding agents such as AlphaEvolve are one of the notable successes in using Large Language Models (LLMs) to build AI Scientists. These agents tackle open-ended scientific problems by iteratively improving and evolving programs, leveraging the prior knowledge and reasoning capabilities of LLMs. Despite their success, mutations over programs are mostly low-level upon surface heuristics, without a persistent unit of control. In contrast, science compounds through ideas. In this work, we propose ExpEvolve, which injects idea-level extraction and control into evolution: it discovers and leverages the high-level factors from programs that explain scoring differences among different programs. We characterize theoretical conditions on when discovery with idea factors improves sample efficiency. We also build a practical framework to show that the extracted high-level idea factors are helpful for guiding exploitation from experience and exploration when discovery comes to a plateau. Across 11 open-ended mathematical tasks, ExpEvolve significantly improves the final performance and sample efficiency up to 5.0x. The improvements also transfer to 4 other LLMs and to 6 machine learning tasks from MLS-Bench. Analysis shows that ExpEvolve can discover idea-level factors to provide better exploration and exploitation for evolutionary discovery.
Abstract (translated)
URL
https://arxiv.org/abs/2610.29303
PDF
-
Same Probabilities, Different Advice: Gain-Loss Framing Effects in High-Stakes LLM Decision Support
2026-10-21 06:46:33Yuyao Zhang, Yakun Cui, Han Zhu, Jinsong Lu, Yike Guo, Sirui HanAbstract
Large language models (LLMs) are increasingly used to support consequential decisions, yet their reliability under equivalent representations of uncertainty remains unclear. Existing work has documented prompt sensitivity, cognitive biases, and confidence miscalibration, but it is still unknown whether high-stakes advice changes when the same probability is expressed through complementary gain and loss frames. We introduce FrameBench-HPS, a controlled benchmark of 2,880 prompts derived from 240 high-stakes scenarios across healthcare, finance, education/career, and legal/consumer-rights domains. Across seven LLMs, gain-loss framing changes the recommendation category in 32.4% of complete groups, with most changes involving uncertainty rather than direct reversals. Human evaluation confirms corresponding shifts in perceived risk and directiveness. Robustness experiments show that framing effects persist under conversational paraphrases and exceed ordinary repeated-generation variability. Probability canonicalization substantially reduces framing sensitivity, identifying probability representation as a practical target for improving consistency in high-stakes LLM decision support.
Abstract (translated)
URL
https://arxiv.org/abs/2610.29294
PDF
-
Credit Where Credit Is Due: Contrastive Attribution Topology Search for Multi-Agent Systems
2026-10-21 06:46:11Fabian Tritthart, Tom ZehleAbstract
An LLM multi-agent system (MAS) is only as good as the agents it chooses to trust: its accuracy and cost hinge on its topology, meaning which agents are included and how they communicate. A single task score, however, does not reveal which agents help or hinder performance. We propose Contrastive Attribution Topology Search (CATS), which searches for compact teams within a fixed supergraph of specialized agents. An LLM attributer compares execution traces across candidate topologies and assigns credit to individual agents. CATS uses these credits to update supergraph edges while progressively pruning weak edges and agents, yielding a small team for deployment. With small open-weight models and matched candidate evaluations, CATS improves mean accuracy on BigCodeBench by 5.3 percentage points over score-only heuristic search while using 54\% fewer deployment tokens. On MBPP, CATS matches the strongest baselines with fewer tokens than the other searches. In a HotpotQA prompt-poisoning stress test, CATS excludes all adversarially prompted agents in all three runs, whereas random search retains some at comparable accuracy. Per-agent credit thus supports both team optimization and the pruning of potentially harmful agents.
Abstract (translated)
URL
https://arxiv.org/abs/2610.29292
PDF
