Artificial Intelligence · Cross-listed from cs.MA
Prompted Identity Degrades Cooperation in Multi-Agent LLM Systems
Xavier Del Giudice, Alessio Palma, Matteo Migliarini et al. (+2 more)
Multi-
agent LLM systems increasingly mix models from several providers, yet exposing each
agent's underlying model identity to its peers significantly impairs cooperation. We show that when
agents are aware of each other's model family, the group splits into clusters, where
agents prefer interacting with others carrying their same label, although nothing in the task rewards or asks for such a split. We argue that the label itself causes this split, which we define as $\textit{factionalism}$. We show and measure this phenomenon in two cooperative games and on a
reasoning benchmark, with nine to twenty-five
agents drawn from up to five open-weight model families. We further show that when the announced families are shuffled, or replaced by arbitrary labels, the factions still follow this information; when the label is removed, this behavior disappears. In strictly cooperative tasks, labeled groups spend on average $30\%$ more rounds and $55\%$ more
tokens to reach a decision, and their success rate drops from $96\%$ to $81\%$. The effect replicates across tasks, group sizes and model families. Withholding identity labels from the
agents is simple and effective mitigation.
Read the full abstract ▾
All authors Xavier Del Giudice, Alessio Palma, Matteo Migliarini, Fabio Galasso, Indro Spinelli
Artificial Intelligence · Cross-listed from cs.LG
Normative Loss Landscape Navigation: A Trajectory-Based Approach to Mitigating Forgetting in Incremental Learning
Isabelle Aguilar, Zayn Andre Zainal, Luis Fernando Herbozo Contreras et al. (+2 more)
Continual learning models suffer from catastrophic forgetting when trained sequentially on non-stationary data distributions. Previously, this has been addressed through weight regularization. While preconditioning gradients offer a promising alternative to mitigate forgetting, current approaches are myopic. Conversely, standard regularization methods apply rigid, scalar Euclidean penalties that entirely ignore the underlying Riemannian geometry of the parameter space. To overcome this gap, we propose TMLN (Trajectory-Modulatory Landscape
Navigation), a normative
navigation policy that formalizes continual learning as an optimal control problem over a curved loss landscape. TMLN utilizes a
memory-efficient diagonal empirical Fisher Information
Matrix (FIM) to define a localized Riemannian
manifold. To compensate for the spatial limitations of the diagonal approximation, TMLN dynamically modulates a preconditioner using the normalized historical trajectory of the network's parameter values. By integrating this trajectory-based preconditioning directly into the gradient update, we actively shield historically critical parameter directions without relying on additive penalties. Empirical evaluations on class- and domain-incremental
benchmarks demonstrate that our method significantly reduces the loss barrier between consecutive tasks.
Read the full abstract ▾
All authors Isabelle Aguilar, Zayn Andre Zainal, Luis Fernando Herbozo Contreras, Zhaojing Huang, Omid Kavehei
Artificial Intelligence
Grab a Coffee: Future-Aware Guidance for Discrete Diffusion with Compiled Objectives
Hua Xu, Dongxin Li, Gwen Yidou-Weng et al. (+3 more)
Discrete
diffusion models generate sequences by iteratively resolving multiple
tokens in parallel, offering a flexible alternative to left-to-right generation. However, guiding this process with a sequence-level objective is difficult because the value of one unresolved
token depends on the other
tokens with which it can form a high-reward sequence. Enumerating all such completions makes the whole guidance computation grow exponentially with the number of unresolved positions. We introduce COFFEE, a plug-and-play framework that avoids this enumeration by separating sequence dependence from the objective. At each
diffusion step, a target-free carrier absorbs the marginal
token distributions predicted by the denoiser to construct a joint model over the unresolved
tokens, while a compiled finite-state model records how their combinations affect the sequence-level preference. Pairing their states allows COFFEE to transfer global preferences to unresolved positions and sample a clean reconstruction without retraining the
diffusion model. The same framework supports explicit hard constraints and learned soft objectives. We evaluate COFFEE across multiple symbolic,
language, and biological
benchmarks, where it achieves strong control results with task-dependent quality and diversity trade-offs. By making objectives available to inference rather than only evaluation, COFFEE brings joint conditioning, completion-weighted guidance, and optimization-based constraints into pretrained neural generation, showing the potential of neural-symbolic methods in
diffusion guidance.
Read the full abstract ▾
All authors Hua Xu, Dongxin Li, Gwen Yidou-Weng, Guy Van den Broeck, Wei Wang, Anji Liu
Artificial Intelligence · Cross-listed from cs.CL
Almost Human, Except When It Matters: VoxParity and the Decisions a Voice Should Change
Bhavik Mangla
A voice
agent can handle almost every call on the words alone and still fail the few its sector's rules were written for. Emergency-call standards, fraud guidance,
radio phraseology and vulnerability rules recognise that how a caller sounds, or what else is audible, can change the right action. VoxParity tests whether
agents act on it. In 183 scenarios from 14 sectors, one transcript stays fixed while the audio changes (a coaching voice, a medical monitor beeping, a mayday under a
radio check, noise over a drug name, a child's voice placing a bet, a frightened whisper), and with it the correct typed tool call. A words-only null test credits a system only if hearing the call moves its actions more than it moves a pipeline that only reads the words. Only 11 of the 23 systems that can also be run on the transcript pass. Descriptively, errors run toward the words: when the audio calls for protection, all 28 systems carry out the routine request more often than they over-react on clean calls (41% against 12% pooled; the words-only pipeline, 58% against 15%). Exploratory analyses place most of the leading systems' misses on cues they heard; systems beat the null almost entirely on items that state the rule; the leading systems overrule heard resignation or confusion far more often than acute alarm; and, in the models tested, describing the voice and stating the rule each recover part of the shortfall, leaving a gap on
emotion.
Read the full abstract ▾
Artificial Intelligence · Cross-listed from cs.SE
Evaluating Name-Only Directory Routing for One-Shot Code Search
Manoj Bajaj
Finding the right files is an early challenge for coding
agents. We test whether a
language model can follow directory and file names to find annotated code files missed by fixed lexical queries. Across 82 audited issues from 11 repositories at pinned pre-fix commits, name-only directory routing recovered 0.465 of gold files within eight candidates, compared with 0.352 for FTS5 and 0.245 for a fixed full-issue rg query. The paired gain over FTS5 was 0.113 (95% repository-cluster bootstrap interval, 0.053 to 0.168). Under a shared 16K-
token context budget, routing delivered 0.443 of annotated lines versus 0.246 for FTS5 on 55 cases with fully aligned annotations. At the same eight-file limit, combining routing with FTS5 reached 0.491 file recall, but its gain over routing alone was uncertain. An exploratory flat path control reached 0.572 recall while using 24.6 model calls per issue, compared with 8.9 for routing. Routing averaged 8.9 seconds per issue; FTS5 took 7 milliseconds per query after a 0.9-second build. On this cohort, directory routing added relevant file candidates to one-shot lexical search, but the study cannot attribute the gain to hierarchy or show that it improves issue resolution.
Read the full abstract ▾
Artificial Intelligence · Cross-listed from cs.SE
UNBIND: UNlearning By INference-time Directional Steering for Code LLMs
Zhengyang Shan, Jiayun Xin, Yanjun Lin et al. (+7 more)
Code large
language models acquire programming capabilities from large code corpora, but can also memorize implementations that later require removal. Code unlearning is needed to control their continued reproduction when copyright or security concerns arise. However, targeted and retained code share computational patterns, creating a tension between forgetting specific implementations and preserving general programming ability. We propose UNBIND, a code unlearning framework that separately considers which hidden states correspond to the target code and how to suppress its reproduction. By constructing separate directions for these objectives, UNBIND achieves selective unlearning at inference time while keeping model weights fixed. Our evaluation covers fourteen baselines across two code models and two corpora. UNBIND achieves the highest joint forgetting and utility score in every setting. It reduces target code reproduction by 97.3% to 99.1% as measured by F-BLEU, with at most two fewer HumanEval+ and six fewer MBPP+ problems solved than the original models. In repeated extraction tests under a fixed budget, the number of targets yielding exact spans of at least 50
tokens falls from 188–262 to 0–2 out of 300 per setting. No extracted span reaches 100
tokens, and the mean best recovery ratio ranges from 0.43% to 6.45%.
Multilingual and related-code evaluations further show effective forgetting with limited impact on useful programming capabilities, supporting UNBIND as a practical approach to selective code unlearning.
Read the full abstract ▾
All authors Zhengyang Shan, Jiayun Xin, Yanjun Lin, Xu Qian, Zhiang Liu, Minghui Xu, Yue Zhang, Qin Hu, Kun Li, Xiuzhen Cheng
Artificial Intelligence · Cross-listed from cs.CR
MMSkillRisk: Can Agents Stay Safe When Multimodal Skills Become Traps?
Lingqi Jiang, Jialuo Chen, Jianan Ma et al. (+8 more)
Agent skills are shareable packages of procedural instructions, tools, and examples.
Multimodal skills additionally include visual references that
agents retrieve and inspect during execution. Because these images guide actions, attackers can disguise malicious instructions as ordinary visual guidance within otherwise legitimate skills. Existing skill-security research primarily examines text-carried attacks or scanner detection, leaving the runtime effects of image-borne attacks insufficiently evaluated. We introduce MMSkillRisk, to our knowledge the first publicly available
benchmark dedicated to end-to-end safety evaluation of image-borne attacks in
multimodal skills. To instantiate this attack surface, we design Native-Context Visual Attack (NCVA), which disguises malicious instructions as native components of teaching images, such as annotations and interface labels. The accompanying SKILL.md provides auxiliary guidance toward relevant visual regions without explicitly stating the malicious operation. Built from 28 curated clean skills, MMSkillRisk contains 36 attack packages and 108 executable cases spanning five attack objectives, with separate checks for attack success and legitimate-task completion. Across nine model-harness configurations evaluated in isolated sandboxes, NCVA induces unauthorized operations in every configuration. Its pooled attack success rate (ASR) reaches 43.1%, exceeding the matched text-carrier baseline by 16.4 percentage points, with higher ASR in all nine configurations. Attack success and legitimate-task completion co-occur in 36.5% of cases, reaching 72.2% for GPT-5.6-sol with Codex. These results show that skill-bundled images can induce unauthorized actions even as
agents complete legitimate tasks, so task success alone does not establish safe skill use. Our code and data are available at https://github.com/kaill-jlq/MMSkillRisk.
Read the full abstract ▾
All authors Lingqi Jiang, Jialuo Chen, Jianan Ma, Xinhao Deng, Xiaohu Du, Sibo Yi, Yuqi Qing, Zhenguang Liu, Qinming He, Shiwen Cui, Changhua Men
Artificial Intelligence · Cross-listed from cs.LG
Learn Now, Use Next, Trust Later: Prequential Test-Time Learning for LLM Agents
Tong Zhao, Reed Li, Yuyang Hu et al. (+5 more)
Adapting large
language model
agents during deployment requires not only retaining past experience, but also turning new observations into timely guidance. Many test-time learning methods, however, acquire knowledge from completed episodes. Feedback from an ongoing interaction may therefore not be distilled into knowledge soon enough to help the next decision. Acquiring knowledge at the granularity of individual transitions could reduce this delay, but raises a separate challenge: a rule that is useful within one episode may not be reliable enough to guide future episodes. Waiting for validation can forfeit immediate benefits, whereas unrestricted reuse can propagate accidental or misattributed guidance. We introduce StepLearn, a nonparametric framework that separates immediate use from persistent trust. It turns informative transitions into hypotheses that can guide the next step, while requiring prospective validation before reuse across episodes. Their predicted effects are checked against subsequent observations outside the source episodes, and only sufficiently supported hypotheses become available for persistent guidance. This process updates external knowledge while keeping all model parameters fixed. Over five rounds on WebArena-Lite and ALFWorld, StepLearn achieves average success rates of 59.9% and 84.0% with GPT-5-mini, and 57.8% and 88.1% with Qwen3.5-35B-A3B, respectively. It outperforms EvoTest, the strongest baseline, by 2.2-12.7 percentage points across the four settings. Learning dynamics further shows that these gains are not restricted to the final repetition, with advantages already present on first task attempts in most settings.
Read the full abstract ▾
All authors Tong Zhao, Reed Li, Yuyang Hu, Yutao Zhu, Haijin Liang, Haibo Shi, Yu Lu, Zhicheng Dou
Artificial Intelligence · Cross-listed from cs.CR
Cheap to Hypothesize, Costly to Verify: The Defense Surface of Agentic Vulnerability Discovery
Kaikai Zhang, Zihan Zhang, Yuchong Xie et al. (+4 more)
Autonomous LLM
agents turn vulnerability discovery into a repository-scale search: they generate many vulnerability hypotheses but can verify only a subset under a finite budget. We show that autonomous vulnerability discovery exhibits a hypothesis-verification asymmetry, where verifying a candidate hypothesis through reachability analysis, execution, and proof-of-concept construction is substantially more expensive than forming it. Under a finite resource budget, this makes autonomous discovery a resource-bounded selective-verification process, further exposing verification effort as a unique defense surface. We present RedHerring, which inserts certifiably safe decoys that divert verification effort from real vulnerabilities. Each decoy combines a CVE-derived vulnerability chain that attracts verification with a false bridge that keeps its dangerous sink unreachable. A private certificate lets the defender verify this property efficiently, while establishing the same fact from the released repository requires solving a computationally hard problem. RedHerring further adapts each decoy to the target repository so that it reads as ordinary program logic. Across 33 OSS-Fuzz projects, 70 evaluation instances, and five models under matched budgets, RedHerring reduces real vulnerabilities discovered by 38.7-60.4%. Trajectory analysis shows that
agents spend 30.6-51.5% of completion
tokens and an estimated 32.5-49.9% of runtime verifying decoys, showing that RedHerring redirects a substantial fraction of the fixed search budget toward decoys. When explicitly informed that decoys may be present, the
agent adapts its search strategy, yet RedHerring still reduces vulnerabilities discovered by 37.2% relative to an informed Baseline, showing that its effectiveness does not depend on decoy secrecy.
Read the full abstract ▾
All authors Kaikai Zhang, Zihan Zhang, Yuchong Xie, Zesen Liu, Shuangjie Yao, Zhixiang Zhang, Dongdong She
Artificial Intelligence · Cross-listed from cs.CR
Similarity Is Not Validity: Defending LLM Semantic Caches Against Poisoning
Zihan Zhang, Shuangjie Yao, Zesen Liu et al. (+8 more)
Semantic caches reduce LLM serving costs by reusing previously generated answers for semantically similar queries. However,
retrieval is based solely on
embedding similarity between the incoming query and cached queries. This design enables cache poisoning: an attacker can cache a malicious response under a query with high cosine similarity to benign requests. The vulnerability stems from a gap between
retrieval similarity and answer validity. From an information-bottleneck perspective, query
embeddings can lose information needed to distinguish valid from invalid cache hits, which limits any matching
algorithm that uses only these
embeddings. We propose a novel defense that recovers this necessary information from the raw text of the cache key. Across poisoning attacks,
adversarial queries share a rewrite-residual structure: they pair a rewrite of the target query with residual content. The rewrite maintains high similarity, while the residual elicits the malicious response. Deleting the residual makes the remaining rewrite more similar to the incoming query. We exploit this structure using Deletion Gain to search shortened variants of the cached query for similarity gains, and an Answer Check to test whether the removed text contributes to the stored answer. We prove that Deletion Gain stays positive when a deletion leaves text close enough to the rewrite, and we search for such deletions with a sliding window. Across three poisoning attack classes, our defense blocks 82.0% to 98.2% of poisoned entries at a 5% false-positive rate, with negligible serving overhead.
Read the full abstract ▾
All authors Zihan Zhang, Shuangjie Yao, Zesen Liu, Zhixiang Zhang, Wai Ip Lai, Dung Hiu Hilton Yeung, Chun Kit Zhang, Fuchen Ma, Yuanyuan Yuan, Yu Jiang, Dongdong She
Artificial Intelligence · Cross-listed from astro-ph.IM
Reconstructing Implicit Scientific Knowledge: Evaluating LLM Agents through End-to-End Reproduction of Astronomy
Yuehui Wang, Xinyu Qi, Guirong Xue et al. (+4 more)
The integration of large
language models (LLMs) into scientific workflows is accelerating, yet their ability to reconstruct the
reasoning underlying published research remains unexplored. Papers specify explicit procedures while leaving many methodological dependencies-data selection, calibration corrections, priors, and domain assumptions-implicit. This ambiguity complicates the evaluation of LLM-based
agents, since a failure to reproduce a result may reflect either limitations of the
agent or underspecification in the source. We present a framework that evaluates
agents through end-to-end reproduction, separating execution from verification and computational failure from methodological ambiguity. We apply it to fourteen astronomy studies: a case study from The Astrophysical Journal and thirteen papers published in Nature. Eleven of the thirteen contained an ambiguity preventing a uniquely specified reproduction path. In a controlled case study, twelve predefined paths, a 3x2x2 sensitivity analysis over sample definition, sky masking, and parallax zero-point treatment-gave estimates from 2.16 to 3.53 kpc for the same quantity, with only one recovering the published value (about 2.70 kpc). The published value was never used as an optimization target, selection criterion, or stopping condition; the matching path was found only after all twelve had run. Crucially, the decisive information (a +0.02 mas parallax zero-point correction) was already in the paper, but the
agents did not recognize its causal relevance until the analysis made the effect visible. Matching a published outcome therefore does not validate reconstruction of the underlying
reasoning, and the bottleneck is as often a failure to connect relevant information as to retrieve it. End-to-end reproduction thus serves both as a test of reproducibility and as a framework for evaluating implicit scientific knowledge in AI systems.
Read the full abstract ▾
All authors Yuehui Wang, Xinyu Qi, Guirong Xue, Cheng Wang, Yangbin Xie, Xiaoyu Tang, Cong Sun
Artificial Intelligence · Cross-listed from cs.CV
FurE: Efficient Instance-Specific 3D Fur Reconstruction without Animal-Fur Datasets
Srinjay Sarkar, Prakhar Kaushik, Soumava Paul et al. (+1 more)
Realistic and editable animal fur reconstruction from multi-view images is challenging due to fine-scale detail, self-occlusion and obfuscation, and, unlike human hair, the lack of animal-fur
datasets. Fur usually covers most of an animal's body, with large inter-species and intra-species variability. We present FurE, an efficient strand-based animal fur reconstruction method that recovers a per-strand, editable groom by optimizing a root-conditioned latent field, decoded into strand geometry via a PCA-based decoder. We reconstruct a defurred animal body using local fur-thickness cues from a surface-constrained Gaussian Frosting representation together with part-based priors. We further show that a PCA-based decoder learned from human-hair strand data can alleviate animal-data scarcity while enabling substantially faster optimization. FurE achieves a 10x speedup in strand training over current SOTA dense per-strand optimization while retaining strand fidelity and generalizing across synthetic and real-world sequences, with quantitative and qualitative validation despite the reduction in training time.
Read the full abstract ▾
All authors Srinjay Sarkar, Prakhar Kaushik, Soumava Paul, Alan Yuille
Artificial Intelligence · Cross-listed from cs.CL
Telescopic Language Models
Zhilin Guo, Boqiao Zhang, Hakan Aktas et al. (+14 more)
One deployed
language model must often serve many compute budgets, yet serving each budget still means a separate training or compression run per point. We train a Telescopic
Language Model (TLM) to be that continuum: a nested-capacity
Transformer supervised by
stochastic prefix supervision with a full anchor. At every step, one randomly truncated prefix of the capacity axis is trained against the full next-
token target, alongside one full-capacity pass, so the trained artifact is a valid
language model at every depth. Two forward-backward passes per step, no architectural change, nothing extra at inference. Fixed-exit suites such as Matryoshka
Language Model Suites (MLMS) occupy one point in this design space, and the point has a cost: supervising only a few fixed exits leaves the nested model at chance level everywhere else (perplexity 10^2-10^5 in our baselines). On a 200M proxy suite (20B FineWeb-Edu
tokens, identical data stream for all methods), a single TLM run is a valid
language model at every one of its twenty layer prefixes, in perplexity and on perplexity-sensitive downstream tasks, reducing the area under the quality-budget curve by 43-44% relative to the fixed-exit suites while matching them at full capacity, at ~12% lower GPU cost per run. The prefix sampling density is a dial: concentrating it on a few depths recovers fixed-exit quality there at the price of the continuum, so the operating points become a training-time choice rather than an architectural one. These results indicate that the training objective, not the nesting itself, is what makes a model elastic.
Read the full abstract ▾
All authors Zhilin Guo, Boqiao Zhang, Hakan Aktas, Kyle Fogarty, Nursena Koprucu Aslan, Wenzhao Li, Canberk Baykal, Albert Miao, Siyu Hong, Yixiao Liu, Adam Wu, Ashish Kumar Singh, Sakar Khattar, Chenliang Zhou, Weihao Xia, Cristina Nader Vasconcelos, Cengiz Oztireli
Artificial Intelligence · Cross-listed from cs.CV
Learning Native Reflection in Unified Models with Interleaved Reinforcement Learning
Yijia Fan, Ziqi Huang, Zhongang Cai et al. (+5 more)
Unified
multimodal models can both look at and render images, so in principle they can repair their own generations: diagnose what an image gets wrong, revise it, observe the result, and diagnose again. Whether a revision helps is known only after it is rendered, so the reflection text and the image generation must be learned jointly, over the whole loop. Supervised fine-tuning (SFT) on reflection trajectories gives a cold start but does not find the high-success repair paths, and naive RL that optimizes only the renderer or only one head leaves most of the gain untapped. We introduce UMM-Reflection, which applies reinforcement learning (RL) to complete reflection trajectories inside one unified model: sibling trajectories share one initial image, so the group-relative advantage compares reflection strategies, and one trajectory-level advantage updates both the reflection
tokens and the flow-based revisions, avoiding the combinatorial blow-up of per-round credit assignment. Unlike single-round editing or pipelines with an external critic, credit flows across rounds and to both roles of the same model, and no verifier is needed at inference. On BAGEL, UMM-Reflection improves GenEval by 12.05 points over SFT, and the gains transfer to WISE (+10.97), OneIG-Bench (+3.48), and T2I-CompBench++ (+4.63), none of which is used in training.
Read the full abstract ▾
All authors Yijia Fan, Ziqi Huang, Zhongang Cai, Yan Li, Zimo Wen, Wanqi Yin, Haiwen Diao, Ziwei Liu
Artificial Intelligence · Cross-listed from cs.LG
TokenCast: Forecasting Token Consumption During LLM Agent Execution
Chaoqian Ouyang, Ling Yue, Libin Zheng et al. (+7 more)
When a large
language model (LLM)
agent executes the same task,
token consumption can vary by over an order of magnitude across runs. The
agent chooses its next steps based on tool feedback and intermediate results, while the growing context steadily inflates the input size of every subsequent call. The total consumption of a task is therefore hard to predict before execution and the prediction must be revised as the run unfolds. In this paper, we propose TokenCast, which learns a composable cost representation for each execution segment, recording its own consumption and the context growth it introduces. Composing adjacent segments yields a cumulative estimate that captures the extra input cost incurred when context from earlier segments is re-read by every later call. As execution unfolds, newly observed evidence refreshes the forecast, requiring no additional LLM calls and incurring a mean cumulative prediction time of 32.8 ms per run on SWE-bench Verified. Across 4 task suites and 6
agent models, TokenCast's mean absolute error reduction against the strongest comparator averages 14.5% over 96 evaluated combinations. In offline budget-control replay, TokenCast uses 21.3% fewer
tokens on average than a fixed-budget policy at matched trace completion. The code is available at https://github.com/DEFENSE-SEU/TokenCast.
Read the full abstract ▾
All authors Chaoqian Ouyang, Ling Yue, Libin Zheng, Hanghui Guo, Shengxiang Xu, YiShu Wang, Ran Li, Jian Yin, Shaowu Pan, Shimin Di
Artificial Intelligence · Cross-listed from cs.LG
How to Loop MoE: Flatten the Experts, Untie the Attention
Shouren Wang, Chuang Ma, Mohsen Hariri et al. (+6 more)
Looped
Transformers reuse one block of layers several times: by spending extra computation they push a model of fixed size further, and so use its parameters more fully; while sparse mixture-of-experts (MoE) models activate only a few of many experts for each
token. Looped MoE bridges these two design philosophies and gives MoE models new potential for better expert usage, but it raises a question: how to loop a MoE? We answer it with Foil. With the expert parameters and the expert compute per
token held fixed, Foil (1) flattens the experts, halving the expert layers, doubling the experts per layer and doubling the passes, so that every routing decision chooses from a larger pool, and (2) unties the
attention, giving each pass its own
attention parameters while the experts and routers stay shared. Experiments show that Foil clearly outperforms the unflattened looped baseline: at 20B
tokens every Foil model has lower pretraining loss than the baseline; at 100B
tokens the loss improves monotonically with the degree of flattening, the most flattened Foil ending 0.012 nat below the baseline at equal parameters and compute, with downstream accuracy on par or better; untying the
attention also yields more balanced and more confident routing at equal shape. Our ablations analyse why Foil works and turn the findings into design guidance for looped MoE: the returns of looping and of widening the expert layers amplify each other, routing confidence tracks healthy expert use better than load balance, and a sparse looped MoE should therefore use more experts per layer and more passes. Code and configurations are available at https://github.com/SR-A-W/how-to-loop-moe.
Read the full abstract ▾
All authors Shouren Wang, Chuang Ma, Mohsen Hariri, Debargha Ganguly, Wang Yang, Xiaoqing Tong, Qianying Liu, Xiaotian Han, Vipin Chaudhary
Artificial Intelligence · Cross-listed from cs.LG
KV-streams for Efficient Compaction in Agentic Reinforcement Learning
Emiliano Penaloza, Dane Malenfant, Dheeraj Vattikonda et al. (+15 more)
Scaling the horizon of agentic LLMs is bottlenecked by the need to fit ever longer context traces in GPU
memory. Context compaction has been the most popular mechanism to alleviate this issue, keeping GPU
memory constant for a given trace. Unfortunately, most compaction strategies rely on prefilling the LLM context many times over, hindering training throughput. To alleviate this bottleneck and enable efficient trainable compaction, we propose KV-streams, a plug-and-play strategy compatible with any compaction strategy that substantially increases throughput while showing no evidence of hindering performance. KV-streams enable scalable compaction by streaming the KV cache forward rather than flushing it after each compaction. We show that KV-streams enable three different compaction strategies, achieving a 2.6 to 5x wall-clock speedup in training. Beyond efficiency, we find that the streamed KV cache can act as a recurrent state, carrying forward information that has long since disappeared from the context. Specifically, in a controlled setting we show that, contrary to prior work, RL alone is all that is needed for this behavior to emerge. Overall, we show KV-streams to be an efficient and lightweight plug-and-play addition to any post-training pipeline.
Read the full abstract ▾
All authors Emiliano Penaloza, Dane Malenfant, Dheeraj Vattikonda, Roger Creus Castanyer, Siddarth Venkatraman, Abhay Puri, Jonathan Light, Matthew James Sargent, Augustine N. Mavor-Parker, Massimo Caccia, Lucas Caccia, Glen Berseth, Esmeralda S. Whitammer, Alessandro Sordoni, Minseon Kim, Marc-Alexandre Côté, Laurent Charlin, Guillaume Lajoie
Artificial Intelligence · Cross-listed from cs.CV
Copy the Same, Distill the Difference: Initializing Linear Vision Transformers
Huaiyuan Qin, Muli Yang, Gabriel James Goenawan et al. (+9 more)
Linear
Vision Transformers (ViTs) are designed to replace the
attention in Softmax ViTs with the linear-complexity
attention operator for more efficient
token routing, but they require from-scratch pre-training and typically underperform the original Softmax version. How to initialize linear ViTs both efficiently and effectively still remains unclear. In this work, we explicitly ask: given that most foundation ViTs are built on the mainstream Softmax
attention, can linear ViTs benefit from their pre-trained weights? Recent works on
Attention Transfer show that
attention is the effective transferable component between Softmax ViTs, suggesting
attention alone suffices for such reuse. However, we find the opposite for Softmax-to-linear transfer. The
attention weights are operator-specific: copying them barely helps, and is sometimes even worse than random initialization. Instead, the
attention's
token routing behavior can be recovered through
distillation with a proper loss design, letting linear ViTs reduce the gap and even match Softmax ones. In contrast, the MLP weights, which carry the learned representation, are operator-agnostic: they can be transferred by simple direct copying, which already carries most of the benefit of the pre-trained weights. Thus, copying MLPs can serve as an effective foundation for Softmax-to-linear transfer: paired with the distilled
attention, linear ViTs eventually close the remaining gap and even surpass Softmax ones. These findings hold consistently across various linear ViT variants, different model sizes, and diverse
datasets. We hope this study deepens the understanding of reusing pre-trained weights across
attention operators: copy what stays the same and distill what differs, to recover the benefit across the Softmax-to-linear boundary.
Read the full abstract ▾
All authors Huaiyuan Qin, Muli Yang, Gabriel James Goenawan, Shiqi Huang, Min Kass Chong, Wahyu Wiratama, Peng Hu, Chen Gong, Wu Liu, Xi Peng, Chun Jian Ho, Hongyuan Zhu
Artificial Intelligence
FinAutoRubric: Expert-Guided Automatic Rubric Generation for Evaluating Financial Research Agents
Hoyoung Lee, Suyeol Yun, Jack Haverty et al. (+17 more)
Evaluating finance research
agents requires rubrics that reflect expert standards and fix the values correct as of an information cutoff. Expert-reviewed finance
benchmarks rely on fixed, per-item rubrics, which are costly to extend and cannot encode each institution's own standard. In FinAutoRubric, experts specify reusable evaluation guidance, while
agents and code carry out query-specific rubric generation, review, and validation. This expert guidance governs every
agent, as
prompts and as rules that code enforces, and a Task Bank of reusable criteria carries it across tasks. In long-horizon loops that follow the expert guidance, a writer
agent researches every expected value and a reviewer
agent verifies it, and failures escalate to a human. On three expert-authored finance
benchmarks, its rubrics track expert scoring as closely as the strongest evaluated generator while stating the expert rubric's expected value for more criteria, their scores agree with human grading, and in-house analysts prefer them in a blind review. The released 100-query FinAutoRubric
Benchmark, built from in-house analysts' key questions across 78 tasks and eight asset classes, shows that rubrics from an earlier model generation still leave headroom for a later one.
Read the full abstract ▾
All authors Hoyoung Lee, Suyeol Yun, Jack Haverty, Yunju Cho, Meesong Kim, Daekyung Park, Sumin Kim, Jihoon Kwon, Jasmine Jia Geng, Andrew Chin, Yin Luo, Edward Tong, Yu Yu, Zach Golkhou, Minkyu Kim, Igor Halperin, Young Cha, Alejandro Lopez-Lira, Chanyeol Choi, Yongjae Lee
Artificial Intelligence
Shockingly Simple Self-retrospection Improves Agentic Models Without RL
Jonathan Light, Christopher Zhang Cui, Jeonghye Kim et al. (+7 more)
People learn not only by repeating successful actions, but also by recounting and explaining their experiences, revising their understanding to guide future behavior. Can a
language-model
agent improve its future actions by training only on explanations of its own experience? We investigate this question by studying Retrospection-Only Fine-Tuning (ROFT), a minimal online procedure designed to isolate the effect of explanation-only training on subsequent behavior. The
agent attempts a task, observes available feedback, generates a retrospective explanation, and is fine-tuned with a next-
token prediction loss on the explanation
tokens alone. The procedure uses neither an external teacher nor a reward-based policy update. In software-engineering experiments with Qwen3.5-4B, ROFT is trained on problems with mixed successful and unsuccessful base-model attempts. On held-out SWE-bench Verified and Pro, it reaches 49.2% and 26.8% solve rates after 20 updates without using a verifier, compared with GRPO's 48.0% and 25.3% after 40 updates in the evaluated runs, and makes faster early progress in training time and sampled attempts. It also learns to solve individual tasks on which all 64 sampled base-model attempts failed, showing that learning can begin without any initially successful trajectories. Behavioral analyses find that ROFT indirectly assigns credit to actions, encouraging good actions and discouraging incorrect ones. Moreover, prompting retrospections to emphasize more direct solutions yields shorter subsequent attempts even without an explicit length penalty. Together, these findings show that learning to explain can also improve learning to do, establishing self-generated retrospections as useful training targets and motivating further study of explanation-to-action transfer.
Read the full abstract ▾
All authors Jonathan Light, Christopher Zhang Cui, Jeonghye Kim, Roger Creus Castanyer, Emiliano Penaloza, Zhengyan Shi, Alessandro Sordoni, Marc-Alexandre Côté, Xingdi Yuan, Minseon Kim
Artificial Intelligence
Failure-Transparent Agents: Benchmarking Post-Failure Reporting in Tool-Using Language Models
Junru Zhu, Shiming Xie, Aime Lu Fan Chen et al. (+4 more)
Tool-using
agents can fail twice: a required tool can fail, and the
agent can then report success without the evidence needed to justify it. Existing
benchmarks often entangle this reporting failure with tool selection, recovery, and environment dynamics. We introduce Failure-Transparent
Agents (FTA), a controlled
benchmark that fixes the failed observation and required evidence state before generation, making post-failure claims directly auditable. FTA contains 100 tasks with deterministic failure traces spanning five failure families, a neutral control, and four user-pressure conditions, and evaluates unsupported claims alongside useful recovery. Across six models, three response policies, and 3,600 human-annotated responses, false-success rates are 22.8% under the baseline policy, 9.3% with a transparency instruction, and 0.8% with a structured evidence contract. Fabricated-detail rates decrease from 28.3% to 14.3% and 0.8%, while useful responses increase from 74.9% to 89.2% and 98.8%, respectively. The tested evidence-contract policy is associated with substantially lower post-failure reporting errors while useful-response rates remain high within this blocked-task
benchmark.
Read the full abstract ▾
All authors Junru Zhu, Shiming Xie, Aime Lu Fan Chen, Xiaoqing Ding, Chunxin Tang, Ruoyu Qi, Yulang Fei
Artificial Intelligence · Cross-listed from cs.LG
X-Reset: Scaling Object-Centric Reinforcement Learning via Cross-Embodiment Resets
Prithwish Dan, Chenyang Ma and Wei Zhan
Reinforcement learning (RL) in simulation can train dexterous manipulation policies without
robot demonstrations, but training a single generalist policy with task-agnostic rewards faces a severe exploration problem: approaching, grasping, and reorienting diverse objects with many degrees of freedom is difficult to discover from scratch. Prior works make exploration tractable with high-quality
robot demonstrations, per-task reward shaping, or by restricting policies to narrow modes of behavior. We propose X-Reset, a framework that instead resolves exploration with human hand-object demonstrations. Rather than imitating or tracking retargeted human motion, X-Reset kinematically retargets hand-object states to noisy
robot states, filters out states that are unstable in simulation, and samples the remainder as resets during RL training with general-purpose object-centric rewards. The resulting policy depends only on object state and goal, with demonstrations entering training through the reset distribution. We show that X-Reset trains generalist policies on 20 objects across three embodiments—a 22-DoF hand on two different arms and a parallel-jaw gripper—and resolves the exploration challenges of RL from scratch. X-Reset scales with the number of training objects, generalizes to unseen objects, can learn from imperfect hand-pose estimates, and transfers behaviors zero-shot from sim-to-real.
Read the full abstract ▾
Artificial Intelligence
Reinforcing Agentic Creativity in Scientific Ideation with Night Science
Priyanka Kargupta, Silviu Cucerzan, Shweti Mahajan et al. (+4 more)
Large
language models (LLMs) excel at structured, verifiable tasks, but their low-
entropy bias can produce homogeneous and predictable outputs, limiting their utility for open-ended scientific ideation. Effective discovery, however, spans a broader creative spectrum: from structured day science to loosely structured, serendipitous night science that reaches ideas beyond those typically considered. We introduce AI Night-Scientist, an agentic framework that uses reinforcement learning to teach models when and how to depart from predictable
reasoning. Grounded in cognitive science, we model creativity along three axes: action (what to do and how creatively), process (when to explore versus exploit), and outcome (the novelty and usefulness of the resulting idea). We use these axes to train models with GRPO, exposing them to varying degrees and forms of creativity throughout training. This produces substantially more diverse scientific proposals, expanding the range of research directions by 27.8% and contribution types by 14.9% over the base model. It also improves predicted citation impact by up to 32.0 percentage points and originality by 66.2 points. These gains cannot be reproduced by simply increasing decoding temperature; instead, we find that semantic guidance specifying what kind of creativity to pursue is critical. Overall, our results suggest that creativity is a learnable, multi-level ability that can be shaped to help researchers reach ideas beyond those typically explored by LLMs.
Read the full abstract ▾
All authors Priyanka Kargupta, Silviu Cucerzan, Shweti Mahajan, Allen Herring, Jiawei Han, Ryen W. White, Sujay Kumar Jauhar
Artificial Intelligence · Cross-listed from cs.LG
A Unified Uncertainty Representation for Graph Neural Networks via Doubly-Spectral Stochastic Expansion
Fred Xu, Thomas Markovich, Florence Regol et al. (+1 more)
Reliable deployment of graph neural networks requires calibration, out-of-distribution (OOD) detection, and robustness to distribution shift, yet existing methods address these needs with separate models and objectives. We model uncertain node
embeddings as random graph signals: graph
Fourier filters capture structural variation, and a scalar orthogonal-
polynomial chaos coordinate captures latent
stochastic variation. The resulting doubly-spectral
stochastic (DSS) expansion supplies task-matched readouts from one representation: the mean coefficient encodes class evidence for the energy-based OOD score, the higher-order coefficients encode structured logit variation, and quadrature averaging over the chaos coordinate defines the single predictive distribution used for prediction and calibration. A capacity
theorem shows that, under a full-rank feature assumption, a restricted subfamily matches the chaos coefficients of any Gaussian-latent random graph signal, with exponentially decaying truncation error under a growth condition; the task-level claims are established empirically. DSS-GNN has two deployment modes: standalone, or as a residual branch beside a deterministic encoder (DSS-Hybrid). Standalone DSS-GNN achieves the lowest Brier score among the compared uncertainty-aware baselines on all 14 node classification
benchmarks without post-hoc correction; DSS-Hybrid achieves the best AUROC on most node-OOD settings, competitive cross-graph OOD detection, and the strongest shifted accuracy on all 7 GOOD concept-shift
benchmarks under standard empirical risk minimization (ERM). Cross-evaluating both modes on all three tasks shows that each remains effective on the other's tasks, with documented exceptions, and yields explicit deployment guidance.
Read the full abstract ▾
All authors Fred Xu, Thomas Markovich, Florence Regol, Yizhou Sun
Artificial Intelligence · Cross-listed from cs.LG
Distillation Defenses Easily Break After Reinforcement Learning
Shidan Javaheri, Alexander Panfilov, Oliver Britton et al. (+2 more)
Distillation attacks copy the
reasoning capabilities of closed-source large
language models, allowing bad actors to replicate state-of-the-art performance at low cost. Attackers systematically collect a large volume of frontier model
reasoning traces and then train (i.e., "distill") their own models on these traces. Existing defenses against
distillation attacks are typically evaluated immediately after
distillation, implicitly assuming attackers do not train their models any further. In this paper, we argue that a more realistic threat model includes further training with reinforcement learning after
distillation. A misspecified threat model can give a false sense of security – some defenses that seem effective after
distillation can be broken after subsequent reinforcement learning. Practically, reinforcement learning lowers the bar for a
distillation attack to be effective. We show that simple attacks can steal
reasoning capabilities from existing closed-source
language models using data easily obtainable from current APIs, yielding
reasoning improvements equivalent to more sophisticated attacks that extract the full hidden traces. Results indicate that any
distillation defense that leaks sufficient information to reconstruct approximate
reasoning traces is likely ineffective. We conclude by discussing broader implications and batch-level
distillation defenses which could be more effective.
Read the full abstract ▾
All authors Shidan Javaheri, Alexander Panfilov, Oliver Britton, Yarin Gal, Yonatan Gideoni
Artificial Intelligence
Xiang Cheng
We introduce the Continuous
Embedding Diffusion Reasoner (CEDR), a training and inference pipeline for continuous latent
diffusion models that achieves strong performance on mathematical
reasoning and code generation. We learn compact answer representations from three hidden layers of the autoregressive Qwen3-4B model. A staged curriculum first trains the ELF-based
diffusion backbone with teacher
prompt conditioning, then initializes a compact
prompt encoder through imitation and jointly adapts both networks. This replaces the large teacher contextual
prompt encoder at inference with a compact learned
Transformer. The multilayer answer representation also enables asynchronous inference, with different components denoised under separate clocks. We further adapt DiffusionNFT to learned self-conditioning guidance, enabling reinforcement learning with verifiable rewards. Controlled ablations evaluate the effects of representation design,
prompt training, and denoising schedules. CEDR outperforms recent continuous-
diffusion baselines at comparable backbone scales, including before reinforcement learning. With a 638M-parameter denoising backbone and DMD2-style
stochastic sampling, post-NFT CEDR-L achieves {pass@1, pass@8} accuracies of {68.97%, 80.82%} on GSM8K and {32.40%, 53.80%} on MATH500 at 64 denoising steps per sample, and {37.56%, 51.00%} on HumanEval+ and {26.96%, 39.92%} on MBPP+ at 128 steps per sample. Code and model weights are available at https://github.com/chengxiang/CEDR and https://huggingface.co/xc91/CEDR, respectively.
Read the full abstract ▾
Artificial Intelligence
Report: Progressive Disclosure of Agent Skills
Guilin Zhang, Kai Zhao, Priyanka Mudgal et al. (+4 more)
Users of Workday's deployed LLM-based
agents often request features which can be addressed by defining named procedures, also known as skills, in the LLM context, effectively augmenting
agents' capabilities. However, as an
agent's skills library grows in size, so does the
agent's operational cost. Progressive disclosure (lazy-loading) of skills as needed may reduce operational costs, but its impact on overall latency and skill-
retrieval quality remains unclear. In this report, we investigate the impact empirically and find that progressive disclosure improves skill-
retrieval quality but marginally degrades overall latency.
Read the full abstract ▾
All authors Guilin Zhang, Kai Zhao, Priyanka Mudgal, Waleed Ammar, Xiquan Cui, Xu Chu, Alet Blanken
Artificial Intelligence · Cross-listed from cs.LG
Rethinking Circuit Evaluation: Do Circuits Explain Model Errors?
Li Zhang, Chuqin Geng, Mark Zhang et al. (+4 more)
Mechanistic interpretability (MI) aims to explain a model's behaviour through analyzing its internal computations;
circuit-based explanations aim to isolate these computations with compact subnetworks validated by ablating the rest of the model. We show that
circuits validated this way may fail to recover the underlying mechanism of the model's behaviour by closely reproducing its successful decisions while failing to account for most of its errors. Such explanations should account for the model's particular errors as well as its successes. We evaluate this requirement by measuring exact answer agreement separately on model successes and failures, across
circuit sizes and ablation settings, on IOI, Docstring, and six model-task settings from the Mechanistic Interpretability
Benchmark. We discover that many tested
circuits closely replicate correct behaviour while missing most of the model's errors. On indirect object identification (IOI) for GPT-2 small, under mean ablation, the manual
circuit and tested automated
circuits, including one trained against the model's full output distribution, agree with the model on 97.3-99.5% of
prompts it answers correctly but only 11.4-41.7% of errors. An IOI case study shows that lost errors are recoverable by restoring omitted
attention-heads which raise error reproduction from 14.2% to 75.1% on a separate held-out set with 0.41 percentage point decrease on correct agreement, exceeding matched random extensions and scalar-biased control. Intervention traces show how omitted computations produce specific wrong answers for a reproducible subset of errors. In all, these findings show
circuits can preserve task success without adequately explaining model's failures, and support exact error reproduction as a necessary, but not sufficient, test of
circuit-based explanations of model behaviour.
Read the full abstract ▾
All authors Li Zhang, Chuqin Geng, Mark Zhang, Chen Yang, Luke Zhang, Haolin Ye, Xujie Si
Artificial Intelligence
Verifier Errors in RLVR: Reward Hacking, Limits of Feedback, and Selective Control
Christian Moya, Elliott Thornley and Guang Lin
In reinforcement learning with verifiable rewards (RLVR), imperfect verifiers can reward incorrect responses, creating opportunities for reward hacking. Using gradient flow with a fixed verifier, we characterize the conditions under which reward rises while correctness falls. We then show that the observations available during RLVR are, in general, insufficient to detect or identify accepted errors, or to guarantee their reduction without sacrificing correct responses. To address this limit, we construct a correction using additional feedback about correctness from audits. This correction achieves selective control: at the current policy, it lowers the probability of accepted errors and raises that of correct responses, provided it outweighs the pressure toward errors from verifier reward. Experiments with log linear and neural contextual bandits and with a
language model support the analysis and show that selective control under partial
auditing reduces accepted errors while increasing correctness.
Read the full abstract ▾
Artificial Intelligence
PhoneCLI: From App Interfaces to Callable Commands for Mobile Agents
Yangqin Jiang, Lingrui Xu and Chao Huang
Mobile GUI
agents operate through a perception–action loop: at each step they screenshot the device, invoke a
vision–
language model (VLM), and emit an action. It is slow, costly, and brittle, yet most of what it does is
navigation—and everyday
navigation is static, ordered, and endlessly repeated. We present PhoneCLI, which compiles an app's GUI
navigation into callable commands, without any app-internal API, runtime instrumentation, or model training. Offline, PhoneCLI explores a target app from the outside and distills its screens, interactive elements, and
navigation edges into a semantically annotated map; each screen yields one deterministic command: a replay sequence that reaches it. Online, the
agent selects a command, verifies it before execution, and then executes it deterministically in sub-second time at zero VLM cost; open-ended interaction and every failure of the compiled path fall back to the embedded VLM interpreter, exactly the pure VLM
agent, so compilation can only help. On AndroidLab, PhoneCLI improves the task success rate while reducing steps and
token consumption, and it transfers to AndroidWorld's official M3A
agent with consistent efficiency gains. What PhoneCLI compiles is the app's
navigation rather than one run, so it serves new tasks, not only repeated ones.
Read the full abstract ▾