Illusions

LLM-driven evolutionary search finds programs by launching seeds and iterating each one. Papers report a single budget setting, usually one seed run for a fixed number of iterations, and rank methods from that one point. We show this is not enough. We evaluate three evolutionary search strategies on five optimization tasks, commonly used by papers in the genre to report results. We run the analysis over a full grid of seeds and iterations. Our findings suggest that the best way to split a fixed budget between more seeds (width) and more iterations (depth) changes with the strategy, the task, and the total budget. Furthermore, we observe that the ranking of strategies also changes with the budget. On one task the strategy that looks worst at one seed is best at forty seeds. On another the best number of iterations is well below the value common in practice, so extra depth wastes budget that more seeds would turn into score. We provide a measurement protocol that reports the seeds-by-iterations frontier and practical guidance for using it.

Cyber-Physical Systems (CPS) operate through a continuous sense-compute-act loop within an open context environment, making it impossible to anticipate all the situations the system will face. To cope with this openness, stakeholders rely on assumptions, formalized into design models. However, these assumptions may no longer hold once the system is confronted with runtime reality, resulting in a discrepancy between expected and observed behaviour known in literature as the reality gap. Existing approaches mainly focus on reducing or overcoming it by making simulations more faithful to reality, with no unified methodology to structure and exploit invalidated assumptions that give rise to this gap as reusable design knowledge. We refer to the persistent reliance on invalidated assumptions -and the resulting false confidence in the design model's operational validity -as design illusions, and argue that they need to be made explicit, structured, and exploited as knowledge to support better design decisions. We propose a conceptual pipeline for illusions-awareness that identifies, classifies, characterizes, and leverages illusions to transform them into actionable design knowledge.

Language-model agents are increasingly used as synthetic people in surveys and social simulations, yet their apparent realism is often judged from population averages or distributional similarity. We tested what such evidence actually establishes by comparing Claude, Codex and Kimi under four prespecified prompt designs with matched farmer decisions from China and four African countries. Some configurations reproduced observed means and adoption rates. However, their person-level predictions were weak; their decisions clustered around typical values and policy-relevant extremes were largely missing. Most strikingly, a simple generator fitted only to the observed marginal dis- tribution, and given no information about any farmer, achieved greater distributional similarity than every language-model configuration. Prompt additions produced conditional gains rather than uni- versal improvement: results varied with model, outcome, population and validation target. We call this the average-farmer illusion: a synthetic population can look realistic while failing to repro- duce who does what or how behaviour varies. We provide a claim-matched validation framework and reusable modular prompts that turn prompt construction into an auditable experimental process. Population-level resemblance should therefore be treated as the start of validation, not as evidence of individual simulation.

Multimodal Sentiment Analysis (MSA) remains constrained by modality imbalance, yet the field continues to rely on optimization-based balancing methods that promise more than they deliver. We provide three contributions: 1) a unified evaluation framework testing gradient and loss-based balancing strategies under controlled settings; 2) a theoretical diagnosis explaining why these methods fail, as they conflate fitting speed with discriminative contribution; and 3) a research agenda toward held-out discriminative modality valuation. Experiments on CMU-MOSI and CMU-MOSEI reveal three shortcomings: no strategy reliably outperforms Late Concatenation; performance is sensitive to hyperparameters; and even ratio calibration fails to yield consistent gains. The core issue is fundamental: loss is not utility, and gradients are not importance. Modality imbalance remains unresolved, motivating utility estimation from held-out performance.

Video large language models (Video-LLMs) are increasingly used as the perceptual front end of world models, a role that assumes they can read motion: how fast something moves, which way it travels, how hard it is pushed. We show they cannot. A Video-LLM can watch two clips of the same person in the same room, name every object in both, and still fail to say which clip moves faster. We introduce MotionBlind, a contrastive benchmark of self-recorded video for physically grounded motion(speed, magnitude, and direction), the variables a world model must predict. Each instance is a pair of near-identical clips that differ only in motion. Each clip carries two complementary yes/no questions, giving four items per instance, and a model earns credit only if all four are correct. We report Instance Accuracy(IAcc), which has a 6.25% chance floor. Single-frame, appearance, and language-only shortcuts all collapse to it. MotionBlind complements the recent TimeBlind benchmark. We run a controlled study of six open and two frontier Video-LLMs, varying whether the video is present, whether frames are shown in the correct temporal order, and how frames are sampled (1 to 24 frames, four selection strategies). Open models sit near the 6.25% floor, and scale does not help. Removing the video drops every model to zero IAcc, and shuffling frames collapses IAcc to chance, so the task genuinely needs video in order. Neither more frames nor smarter frame selection closes the gap, because these change which frames are seen, not whether motion is read. Only Gemini3.1 Pro clears the benchmark overall, and even it fails on speed. A frontend that cannot tell two speeds of the same action apart is not yet a trustworthy source of supervision, reward, or evaluation for a world model.

Existing vision-language model (VLM) benchmarks emphasize perception and reasoning accuracy (how well VLMs describe and reason about what they see in an image), with limited attention to behavioral reliability under uncertainty (how they behave when visual evidence is missing or misleading). We introduce SciFigBench, a diagnostic VLM benchmark for scientific figure understanding that jointly evaluates perception, reasoning, and behavioral reliability under uncertainty. It contains 250 figures with high-quality human annotations across three evaluation aspects, totaling 600+ hours of annotation effort. We further extend these figures via image transformations, reasoning questions, resistance probes, caption-bias probes, and confirmed selective-blur targets, producing over 34,000 evaluation setups for stress testing. We further propose the Admittance-Resistance-Inductance (A-R-I) framework to evaluate whether models acknowledge insufficient evidence, resist misleading context, and infer cautiously from partial information. Our results reveal substantial behavioral differences among models. GPT-5.2 achieves the highest description quality (MQM 91.6) with strong reasoning accuracy (78.4%), yet hallucinates unreadable content in 96% of cases, whereas Gemini 3.1 Pro, a comparably capable model (MQM 90.2, reasoning 81.0%), admits uncertainty in 71% of such cases and achieves the strongest resistance score (0.91). These findings show that high perception and reasoning accuracy alone do not guarantee behavioral reliability, a dimension critical for deployment in scientific workflows.

Safety alignment in large language models (LLMs) is largely developed in English, assuming these safeguards generalize across multilingual settings. However, this assumption remains underexplored and exposes a vulnerability in low-resource languages. We investigate cross-lingual safety transfer in four African languages, Twi, Hausa, Amharic, and Swahili, using LoDNA, a new safety dataset that pairs literal translations with culturally localized prompts. To move beyond generation-based evaluation, we propose a latent geometric framework that probes hidden-state refusal representations in LLMs. Our experimental results show that cross-lingual safety transfer is severely limited; harmful prompts retain less than 10% of the English refusal signal across most language-model pairs. Literal and localized prompts are semantically aligned (cosine 0.95-0.996) but drift across layers, suggesting models encode the concepts without routing them to safety mechanisms. These findings demonstrate that current multilingual safety alignment is superficial, providing strong evidence against the assumption of a universal, language-agnostic harm manifold within the specific low-resource languages studied. Warning: This paper contains example data that may be offensive or harmful.

Recently, AI-driven video generation has attracted considerable attention. This surge increases the demand for reliable video quality assessment (VQA) metrics to evaluate AI-generated content (AIGC) videos and guide model optimization. Existing studies assess video quality through visual harmony, video-text consistency, and domain-specific alignment, yet lack quantitative metrics for measuring fidelity to physical laws. To address this limitation, we present a novel benchmark that evaluates the quality of AIGC videos based on their compliance with physical principles by quantitatively measuring geometric consistency across frames extracted from generated sequences. This serves as a proxy for estimating the extent to which generated videos conform to real-world physical rules. Specifically, GeoCon-Bench captures global motion through translation estimation, fits homography or fundamental matrix models using background correspondences, and reports complementary metrics, including inlier ratio and geometric error. We also release a dataset containing 20 scenes across six motion categories. Experiments on state-of-the-art AIGC models demonstrate the reliability of GeoCon-Bench as a video quality assessment metric.

The Tower of Hanoi is a simple planning puzzle that in prior work has proven challenging for large reasoning models (LRMs). Current models solve the standard formulation of the puzzle, but still struggle with the flat-to-flat variant (where initial and goal states are not restricted to have all rings on a single peg). This paper presents an in-depth study of how both small, in-house Transformers and large, third-party LRMs solve this task. To understand the failures mechanistically, we first train small Transformers from scratch on precomputed solution traces. Using a variety of interpretability techniques, we show that these Transformers develop an emergent world model: a linearly decodable, geometrically faithful representation of the puzzle's state space (the Sierpinski triangle), that is causally involved in solving the puzzles. Second, we return to the large LLMs and apply our techniques to two frontier reasoning models, Qwen3.6-27B and DeepSeek-R1-Distill-Qwen-32B, that attempt to solve the task through extended chain-of-thought. Surprisingly, we find that both models encode the Sierpinski world model near-perfectly at the end of the prompt, and yet fail at the majority of tasks when there are more than 3 rings. We locate the source of this failure in the decaying representation of the world model. We probe for the representation at different stages during planning, and establish causality by showing that performance can be improved by injecting the prompt-time representation at inference. The failure of the models is thus one of maintenance of the required representations, not their absence, and performance is at least partially recoverable. These results thus reframe the reported collapse in performance from prior work: current Large Reasoning Models build a world model, and then lose it.

The "thinking-with-images" paradigm equips multimodal LLMs with active visual operations such as crop-and-zoom. However, models using these operations often achieve only marginal or negative gains over direct inference at substantially higher token cost. They may also repeatedly crop irrelevant regions and fail on questions that direct inference answers correctly. We ask whether the returned visual evidence causally affects the answer. To answer this question, we formulate visual tool-use as a causal graph that separates observation-mediated paths from action-induced shortcuts. We then audit it through interventions at the three levels: policy (comparing tool-use with direct inference), trajectory (corrupting all observations during rollout), and step (counterfactually replacing one individual observation under a fixed prefix). Our step-level estimand, Visual Evidence Gain, isolates the contribution of each returned observation. Across six representative models and five fine-grained perception benchmarks, we uncover policy miscalibration with two failure modes. In Calling Without Looking, returned observations have no causal effect on the answer. In Looking Without Planning, observations are informative but the call schedule is incoherent. A trajectory-level diagnostic decomposes the policy-level accuracy gain into per-group contributions and shows that the gain is concentrated in a Calibrated minority. We term this discrepancy the illusion of visual tool-use: despite aggregate accuracy gains, visual tool-use is not causally effective across a broad range of rollouts. The code is available at https://github.com/OpenCausaLab/CauAudit.

While neural networks are typically evaluated by their training and test performance, these metrics do not reveal how robust a learned representation is. Recent studies have shown that solutions occupying larger volumes in parameter space, as quantified by Boltzmann entropy, often exhibit superior generalizability compared to those reached by conventional optimization, a phenomenon known as the high entropy advantage. Here we ask whether this advantage persists beyond generalization. Specifically, we investigate models' robustness, the ability to retain the learned knowledge when the model is subsequently trained to acquire new information. Using grokking in modular arithmetic as a controlled setting, we design a noise injection experiment to evaluate the robustness difference between AdamW-trained transformers and high-entropy model sampled from Wang-Landau Molecular Dynamics with identical saturated performance. By forcing both models to fully remember new data with random labels, we find that AdamW-trained models suffer from catastrophic forgetting, with original task test accuracy dropping from 100% to below 75%, whereas the high-entropy models maintain approximately 95% test accuracy. We term this hidden fragility behind apparent generalization the "grokked illusion." Through singular value decomposition of the neural network weights, we discover that high-entropy neural networks possess significantly higher effective rank in attention and MLP layers both before and after noise injection, indicating richer feature representations can serve as a buffer against catastrophic forgetting. Our findings demonstrate that perfect generalization does not imply equal robustness, offering a new perspective on what makes a trained model robust to interference.

Large Vision Language Models have integrated reasoning capabilities, elevating cognitive performance to new levels. However, existing evaluations either focus solely on perception or rely on specific domains such as maths or coding. Evaluation for reasoning capabilities that align with an open-world environment is still required, especially one that considers perception and reasoning jointly. To bridge this gap, we propose to evaluate LVLMs by exploiting visual illusions as a diagnostic tool. Visual illusions are phenomena in which the human visual system misinterprets objective signals, resulting in an understanding that deviates from reality. We constructed IllusionReasoning, a benchmark of illusion images collected from the real world, incorporating diverse annotated question-answer pairs. Based on IllusionReasoning, we show that the reasoning capabilities of a wide range of LVLMs are not as advanced as claimed. Our work provides new insights into LVLMs and offers future direction for optimisation.

Large Language Models (LLMs) are increasingly integrated into software development workflows, yet their ability to autonomously generate secure authentication code remains uncertain. This paper evaluates the security architecture of authentication systems generated by five prominent AI coding assistants through a bi-modal assessment framework combining static code analysis and dynamic penetration testing, mapped to NIST SP 800-63B guidelines. The study examines model behavior across four prompting strategies Basic, Secure, NIST-Based, and Reprompting to reflect varying levels of developer guidance. Empirical results demonstrate that code generated from functional or generically secure prompts consistently omits critical protections, particularly concerning brute-force resistance, session management, and robust password handling. While providing explicit, single-shot NIST context significantly improves compliance, the findings reveal that this remains structurally inadequate. Instead, iterative Reprompting: forcing models into a contextual self-auditing loop is strictly required to achieve a comprehensive, defense-in-depth security architecture. Ultimately, this study proves that current AI coding assistants do not produce secure-by-default applications, dictating that enterprise deployments must transition from single-shot prompt engineering to continuous, standards-driven verification pipelines.

As large language models (LLMs) grow more capable, they are increasingly deployed in context-rich settings where task inputs are often accompanied by long, partially irrelevant context. In a controlled setting, we find that state-of-the-art models often appear robust to task-irrelevant context at the aggregate level: prepending it to benchmark questions causes little change in overall accuracy. This aggregate stability, however, masks significant per-example instability. Even semantically meaningless pseudo-words, formed by randomly combining characters, can markedly shift model predictions on a small fraction of examples, degrading performance on some while improving it on others. This two-sided effect holds consistently across a wide range of models and datasets, yet the affected examples are largely model-specific. We further show that this instability is modulated by context type, context length, test-time compute, and model development stage. Together, our findings reveal context-induced tail risks concealed by aggregate accuracy, motivating per-example reliability evaluation of language models.

Post-training quantization is widely used to deploy large language models in resource-constrained settings, yet its evaluation relies almost exclusively on accuracy and perplexity. We show that these metrics fail to capture behavioral changes induced by quantization. We introduce correctness agreement, a decision-level metric that measures overlap in correct predictions between a base model and its quantized variants, independent of absolute accuracy. Across multiple models and quantization schemes from 8-bit to 2-bit, we find that behavioral divergence emerges under moderate quantization even when task performance appears preserved. To explain this effect, we analyze quantization as a structural operator on attention weights and quantify layer-wise distortions using statistical and distributional measures. Our results reveal non-linear breakpoints at low bit-widths and show that query and key projections are consistently more sensitive than value and output projections. These findings expose an illusion of equivalence between base and quantized models and motivate behavioral evaluation beyond conventional performance metrics.

Mixed-precision quantization (MPQ) has become a key technique for deploying large language models under stringent memory and compute constraints. We first identify a phenomenon that we term the Perplexity Illusion: layers ranked as important by perplexity-based sensitivity show little rank correlation with those that are most influential for complex reasoning performance, with Kendall $τ\approx 0$ in our analysis. We further reveal an Alignment-Diversity Tradeoff: using only target-task calibration data can degrade post-quantization performance, whereas incorporating general-domain data stabilizes sensitivity estimation and improves robustness across tasks. Based on these observations, we propose TASA (Task-Aware Sensitivity Analysis), a two-level framework that jointly optimizes calibration-data composition and mixed-precision bit allocation. Specifically, TASA searches for a calibration-data mixture using a training-free gradient-trace alignment criterion, and then aggregates perplexity and reasoning-oriented sensitivity signals to guide both inter-layer and intra-layer bit allocation. Experiments on LLaMA-3-8B and Qwen2.5-7B reveal a precision inversion: appropriately allocated 3.5-bit models can match or surpass less task-aware 4-bit baselines. At an average precision of 3.5 bits, TASA matches or outperforms several competitive 4-bit uniform baselines in aggregate accuracy, and improves over the strongest W3 baseline on GSM8K by more than 20 absolute points on LLaMA-3-8B. These results show that calibration-data composition substantially affects task-sensitive quantization, a factor underexplored in prior work.

Safety alignment of text-to-image (T2I) diffusion models aims to suppress harmful generations while preserving utility on benign prompts. Recent methods often appear to deliver high safety with high utility, but this conclusion rests largely on coarse global utility metrics (e.g., FID, CLIPScore) that are insensitive to fine-grained semantic correctness, creating an illusion of high utility. We show that when utility is measured with structured evaluation, this illusion breaks: on TIFA (Text-to-Image Faithfulness evaluation with Question Answering), safety-aligned models suffer substantial drops in semantic fidelity, including failures in object counts, attributes, and relationships. To diagnose the source of this gap, we analyze the text-encoder prompt embedding space and uncover semantic collapse, a contraction of embedding spread coupled with distortion of inter-prompt similarity structure, which strongly correlates with structured utility loss. Guided by this insight, we propose StructureAware Geometric Regularization (SAGE), a safety alignment objective that explicitly preserves embedding spread and inter-prompt relational structure during adaptation. Our method restores structured utility (TIFA +5.0% over prior state-of-the-art) while maintaining strong safety performance and competitive coarse-grained utility scores. Our source code and trained models are available at https://adeelyousaf.github.io/SAGE_ECCV26_Project_Page/.

Knowledge Editing (KE) has emerged as a frontier for updating specific facts in LLMs without costly retraining, but its reliability and underlying mechanisms remain poorly understood. In this work, we examine KE from an adversarial elicitation perspective, revealing that edited knowledge is often not fully erased and continues to surface, with consistent failures observed across diverse model architectures. To explain this behavior, we conduct a mechanistic analysis of popular KE methods. We show that low-rank updates do not overwrite existing knowledge but instead redistribute it within the model's representation space. Furthermore, we find that these methods act as targeted suppression mechanisms that reduce the likelihood of expressing original facts, rather than removing them from the model. Analysis of the loss landscape reveals that edited knowledge lies in narrow, anisotropic regions that are highly sensitive to perturbations, making them highly vulnerable to indirect prompting and adversarial attacks. By exposing these profound architectural vulnerabilities, our work proves that KE algorithms are inherently bypassable and motivates a fundamental reevaluation of how we deploy post-hoc updates in several LLM applications.

Benchmarks for LLM-generated GPU kernels (KernelBench, TritonBench, GEAK) score correctness through fixed-shape, small-sample allclose-style checks. The number of inputs varies between benchmarks. The shape, dtype, and tolerance are fixed for each kernel. We test that oracle empirically. We construct a controlled corpus of 24 Triton and CPU stand-in kernels (15 correct controls and 9 LLM-style buggy variants seeded with documented transcription errors) and re-evaluate it under op-schema-aware seeded fuzzing with a high-precision (fp64) CPU reference and per-(op, dtype) absolute tolerances. The seeded oracle flags 9 of 9 buggy kernels and passes 15 of 15 correct controls, at zero precision cost on controls. We extend the corpus to 26 ops (adding a flash-attention pair) and re-run the same protocol on five GPU classes (RTX 3060, A10, L40S, A100 SXM4, H100 NVL). The verdicts are identical across all five GPUs: 10 of 10 illusions caught and 16 of 16 controls clean. The corpus result is about LLM-style transcription bugs that the allclose-on-one-shape oracle certifies as correct, not about the bug rate of any specific deployed LLM. Every flagged failure replays byte-for-byte from a stored seed.

Reject inference methods are widely used to mitigate survival bias in credit scoring, yet their effectiveness remains poorly understood. We systematically evaluate several such methods and uncover a structural failure mode: in a natural retraining cycle, models whose accuracy improves while recall collapses create an illusion of improvement that leads practitioners to believe the system is getting better when, in fact, its rejection quality -- the ability to correctly screen out defaulters -- is deteriorating. We then propose a controlled exploration strategy that breaks the feedback loop without statistical assumptions: the lender deliberately approves a fraction of rejected applicants and observes their true outcomes. We show that accuracy and rejection quality give opposite recommendations on whether to explore: accuracy favors no exploration, while rejection quality improves with it, confirming that standard evaluation metrics are misleading under selection bias. Even minimal exploration rates (2--5%) prove sufficient in our experiments to diagnose the severity of the feedback loop at near-zero cost. Our findings are consistent across two machine learning methods and three real-world datasets, and suggest that standard evaluation protocols are inadequate for assessing models trained under survival bias.

Multimodal Foundation Models are increasingly used as reasoning agents, making reliability, knowing when a model may hallucinate, critical. A common intuition, which we call the Attention-Confidence Assumption, holds that reliability follows from "structural" visual perception: tight attention on relevant regions should signal a trustworthy answer, while scattered attention signals confusion. We challenge this through the VLM Reliability Probe (VRP), a systematic cross-family study of reliability signals in contemporary Vision-Language Models (VLMs). We introduce structural-attention metrics, cluster counts (C_k) and spatial entropy (H_s), to quantify the visual encoder's gaze, and track its evolution (Delta H_s) across layers. This reveals a "Symbolic Detachment": models often "Early Lock" visual features only to diffuse attention later, severing early perception from final generation. Contrary to the grounding hypothesis, we find a "Cluster Failure": spatial attention has near-zero correlation (R approx 0.001) with accuracy. Instead, reliability is a phenomenon of generation dynamics and internal-state distributions. Self-Consistency, the agreement rate across sampled reasoning paths, is the dominant predictor of truth (R = 0.429). Scaling causal interventions exposes a sharp architectural divergence: LLaVA locks its prediction in a fragile late-stage bottleneck, whereas PaliGemma and Qwen2-VL distribute reliability globally, staying resilient even when ~50% or more of their most predictive layer is destroyed. For current VLMs, reliability signals are detached from visual grounding maps and are best inferred from generation-time dynamics and hidden-state probes.

Machine interpreting (MI), the live, real-time application of speech translation, has achieved remarkable progress on standard benchmarks, with some systems approaching human parity on textual fidelity. Yet the user experience remains far inferior to interpreter-mediated communication, revealing what we term the accuracy illusion: systems that appear accurate on paper but fail in practice to support smooth, goal-oriented interaction. This paper defines MI as a distinct subfield of speech translation, with its own characteristics and the need for evaluation methods grounded in communicative effectiveness rather than isolated fidelity metrics. Drawing on insights from interpreting studies, we identify critical dimensions of professional interpreting practice that are overlooked by current systems, and consolidate them into three interdependent design priorities for future MI: agency (context-sensitive initiative and repair), grounding (multimodal and discourse-level situational awareness), and experience (adaptive improvement through real interaction). Together, these priorities chart a path toward closing the usability gap and enabling systems that can sustain authentic multilingual communication in real time.

Prevailing wisdom posits that Multi-Agent Systems (MAS) are superior to Single-Agent Systems (SAS), citing advantages like context protection, parallel processing and distributed decision-making. However, empirical support for this claim relies primarily on comparisons with SAS baselines using benchmarks that prioritize isolated reasoning tasks, which do not adequately assess these advantages. Focusing on automatically generated MAS that are designed for enhanced generalizability over manually-designed counterparts, we perform a rigorous, systematic evaluation against SAS, specifically Chain-of-Thought with Self-Consistency (CoT-SC). Across traditional reasoning datasets and tasks with interactive multi-step workflows (e.g., BrowseComp-Plus), we demonstrate that automatic MAS consistently underperform CoT-SC despite being up to 10x more expensive. To isolate these failures from limitations inherent to task structure, we introduce a diagnostic synthetic dataset tailored for MAS featuring explicit task decomposition, context separation and parallelization potential. We show that expert-architected MAS consistently outperforms automatically generated architectures in both raw performance and cost-efficiency on this dataset, demonstrating that existing evaluation frameworks mask critical architectural gaps and inefficiencies of complex MAS by failing to account for the marginal utility of increased computational cost. Critically, systematic deconstruction of the generated MAS architectures reveals that current automated design paradigms produce architectural bloat that prioritizes superficial complexity which does not translate into functional utility, exposing a fundamental misalignment with multi-agent principles.

Multi-agent LLM systems for medical question answering often treat consensus as a reliability signal: if multiple agents agree on an answer, it is presumed trustworthy. However, answer-level consensus does not entail reasoning-level alignment. We introduce CARA (Cross-Agent Reasoning Alignment), a family of automated metrics that measure whether agents who agree on an answer also agree on the reasoning. Applying CARA to a standard debate system on two medical QA benchmarks, MedQA-USMLE and MedThink-Bench, we identify the consistency illusion: a failure mode where debate reduces detectable contradictions between agents while simultaneously decreasing the semantic similarity of their reasoning chains; agents appear to agree more but reason less consistently. To improve this misalignment, we propose the Grounded Debate Protocol (GDP), a prompt-level intervention that requires agents to commit to named medical facts and take explicit stances on other agents' claims. GDP produces large, consistent alignment improvements, with Cohen's d ranging from +1.43 to +1.99, across two datasets and two backbone models, without adding LLM calls or modifying system architecture. Our results motivate cross-agent reasoning alignment as a quantity to audit alongside accuracy in safety-critical domains.

Recent work shows that LLM agents struggle to correct errors in their own reasoning traces yet show markedly higher correction rates when identical claims appear under external sources. We ask whether this asymmetry reflects a capability deficit or a role-label artifact: does an agent's willingness to correct a wrong claim depend causally on the chat-template role that carries it, rather than on the claim's content? Our setup keeps the erroneous claim byte-identical across all conditions (SHA-256 verified) and varies only its wrapping role: the agent's own \role{<thought>}, a \role{user} message, a \role{tool} response, or a \role{system <memory>} block. Across 13 model-domain cells covering seven model families and three domains ($n{=}30$ paired tasks per cell), relabeling the claim from \role{<thought>} to an external role lifts the explicit-correction rate by 23 to 93 percentage points, with 10 of 13 cells reaching $p{<}0.001$. Further experiments confirm that the effect is asymmetric, mechanistically decomposable, and robust across domains. The failure to self-correct is not a cognitive deficit; it is a chat-template artifact. We exploit this artifact by designing a prompt-structure-only intervention that requires no training and no model modification, with its strongest role label being domain-dependent: \role{<memory>} dominates on math, while a plain \role{user} message dominates on logical deduction.

A sentence like "The authors that no critics recommended have ever received acknowledgment for a best-selling novel" is sometimes rated as acceptable even though, strictly speaking, it is ungrammatical because the negative polarity word "ever" is not licensed where it is. This behavioral effect is sometimes called a "negative polarity illusion". Here we propose that the lossy context surprisal theory of Hahn et al. (2022) -- whereby people have an imperfect encoding of complex sentences -- might explain this effect. We hypothesize that people have poor memory representation of the determiners in the main-clause and embedded-clause subjects and could entertain a determiner exchange that licenses ever. We propose that more similar determiners in those positions would trigger stronger illusion effects. Acceptability judgment tasks with six novel determiner pairs (e.g., "few" and "many", "few" and "most") support our proposal, showing, specifically, that a novel sentence, "Many authors that few critics recommended have ever received acknowledgment for a best-selling novel", triggered a much stronger illusion than the canonical one even without time pressure. These results offer further support for the suggestion that human language processing is imperfect and resource-rational: in face of working memory limitations, humans rationally reconstruct what is most likely from noisy linguistic input to facilitate downstream processing.

Drawing on Ullmann-Margalit's concept of opting (transformative, irrevocable, and shadowed by foreclosed alternatives), we show that current AI systems raise a profound ethical problem that existing AI ethics has not fully captured: the illusion of opting, in which persons and groups encounter the deceptive appearance of meaningful consequential choice while the agency needed to become genuinely capable of choosing is weakened. Against approaches that treat AI primarily as an optimizer of already given ends, we argue that AI systems should be evaluated by whether they protect and cultivate meta-capacity against the illusion of opting: the socially and institutionally scaffolded agentive capacity through which means and ends can be formed, contested, revised, and owned. This reframing is especially urgent for disadvantaged populations, who are least able to absorb the costs of the illusion of opting when AI-mediated pathways misdirect behavior and action. We propose three normative imperatives for AI-mediated consequential decisions: existential honesty, which acknowledges the limits of prediction; ecological rationality, which situates guidance within heterogeneous lived ecologies; and counterfactual reparation, which acknowledges and repairs foreclosed alternatives when AI-mediated decision-making pathways fail.

In modern RAG pipelines, query augmentation methods such as HyDE and query expansion are applied to every query, resulting in substantial LLM inference costs and increased end-to-end latency. The empirical justification for this overhead in real production traffic remains largely unexplored. We present a case study of the Danish National Encyclopedia, evaluating five retrieval workflows over 20,000 query-workflow pairs from production traffic and synthetic conditions. In this system, synthetic queries suggest that LLM augmentation is needed for over 90% of queries to achieve high retrieval coverage. However, under our production deferral policy, only 27.8% of real user queries need LLM augmentation. We call this gap the Coverage Illusion and attribute it to a structural mismatch between synthetic and real query distributions. Pre-retrieval routing cannot resolve this gap, as the need for LLM augmentation is only revealed after searching the index, a result confirmed by our evaluation of four machine learning paradigms. The coverage gap, undetectable from the query alone, motivates a post-retrieval cascade that runs workflows in cheapest-first order and escalates to LLM augmentation only when a step returns no documents. Operating entirely without training overhead or secondary serving infrastructure, the cascade improves quality by +0.140 Composite Overall points over Always-HyDE, reduces latency by 31.8%, and serves 72.2% of real user queries without LLM augmentation.

Large language models (LLMs) show potential as simulators of human behavior, offering a scalable way to study responses to interventions. However, because LLMs are trained largely on observational data, interventions in experiments with LLM-simulated synthetic users can induce unintended shifts in latent user attributes, causing user drift where the implicit simulated population differs across treatment conditions, potentially distorting effect estimates. We formalize the confounding or selection bias that can arise due to user drift and show how intervention-dependent shifts can inflate or attenuate observed differences in user responses under intervention. To diagnose confounding, we propose using negative control outcomes--attributes that should remain invariant under intervention--to identify distribution shifts across intervention conditions, providing evidence of user drift. To mitigate drift, we study adjusting the persona specification by eliciting additional confounders, finding that targeted, setting-relevant confounders can substantially reduce bias across survey-style and multi-turn agent evaluations.

While Multi-Modal Large Language Models (MLLMs) demonstrate impressive capabilities in general reasoning, their embodied spatial intelligence remains hampered by a "Cartesian Illusion" - a reliance on text-based probability distributions that lack grounded, 3D topological understanding. This limitation is starkly exposed in multi-agent environments, which demand more than just scene perception; they require second-order Theory of Mind (ToM). Specifically, an Agent A must be able to infer Agent B's belief about the environment, governed strictly by Agent B's physical orientation and sensory limitations. In this paper, we probe the limits of two-stage spatial inference in MLLMs through a novel audio-visual task: requiring Agent A to predict Agent B's estimation of A's relative location. To solve this, we propose an Epistemic Sensory Bottleneck module that abandons rigid, rule-based coordinate transformations. Instead, we introduce an Anchor-Based Embodied Spatial Decomposition Chain-of-Thought (CoT). This guides the MLLM through a "geometric-to-semantic" projection, forcing it to first establish B's local coordinate system and then dynamically weight visual and auditory modalities based on whether A falls within B's visual frustum. Extensive evaluations reveal that while current MLLMs fundamentally struggle with spatial symmetry and out-of-view ambiguities (establishing a rigorous zero-shot baseline of 42% accuracy), our sensory-bounded reasoning chain robustly outperforms pure egocentric and allocentric baselines. By systematically benchmarking these perceptual bottlenecks, our work exposes the current limits of MLLM spatial reasoning and establishes a foundational paradigm for epistemic, modality-aware inference in Embodied AI.

End-to-end LLM trading agents have moved quickly from research curiosity to a small ecosystem of named systems, including FinCon, FinMem, TradingAgents, FinAgent, QuantAgent, and FLAG-Trader. Several of these report headline Sharpe ratios that would be material if read at face value on a deployment desk, and associated benchmarks such as FinBen report trading-task Sharpe statistics in the same range. The gap between architecture research and deployment claim has been crossed too freely on both sides of the academia--industry divide. We take a position on that gap: reported alpha from end-to-end LLM trading agents should not be treated as deployment evidence. Before such returns can support claims of deployable trading capability, they must survive structural validity tests for temporal integrity, real-world frictions, counterfactual robustness, predictive calibration, numerical execution, and multi-agent disaggregation. Current public evidence cannot yet distinguish robust predictive ability from temporal contamination, unmodeled frictions, short-window Sharpe uncertainty, narrative fitting, and parametric priors. The problem is not only evaluative but structural. Language confidence is not tradable probability, narrative reasoning is not numerical execution, and model priors may become undisclosed implicit factor exposures. We contribute a minimum reporting protocol suite, P1--P6, with tiered applicability by claim strength, and a conservative modular alternative that uses LLMs as auditable information interfaces upstream of independent calibration, risk, and execution modules. Code and reproduction harness: \url{https://github.com/hj1650782738/Trading}.

Vision-Language Models (VLMs) often produce self-reflective statements like "let me check the figure again" during reasoning. Do such statements trigger genuine visual re-examination, or are they merely learned textual patterns? We investigate this via VisualSwap, an image-swap probing framework: after a model reasons over an image, we replace it with a visually similar but semantically different one and test whether the model notices. We introduce VS-Bench, 800 image pairs curated from MathVista, MathVerse, MathVision, and MMMU-Pro. Experiments on Qwen3-VL, Kimi-VL, and ERNIE-VL reveal a striking failure: models overwhelmingly miss the swap, with accuracy dropping by up to 60%. Counterintuitively, thinking models are nearly 3x more vulnerable than their instructed counterparts, and scaling offers no mitigation. Multi-turn user instructions restore visual grounding, but self-generated reflective statements during continuous generation do not. Attention analysis explains why: user instructions substantially elevate attention to visual tokens, whereas self-reflection does not. Current VLMs tend to say rather than actually see when claiming to perform visual re-examination. Our code and dataset are available at the project page: https://visualswap.github.io

Vision-Language Models (VLMs) increasingly operate on ultra-high-resolution (UHR) Earth observation imagery, yet they remain vulnerable to a severe scale mismatch between large-scale scene context and micro-scale targets. We refer to this empirical gap as a "resolution illusion": higher input resolution provides the appearance of richer visual detail, but does not necessarily yield reliable perception of spatially small, task-relevant evidence. To benchmark this challenge, we introduce UHR-Micro, a benchmark comprising 11,253 instructions grounded in 1,212 UHR images, designed to evaluate VLMs at the spatial limits of native Earth observation imagery. UHR-Micro spans diverse micro-target scales, context requirements, task families, and visual conditions, and provides diagnostic annotations that support controlled evaluation and fine-grained error attribution. Experiments with representative high-resolution VLMs show substantial failures in spatial grounding and evidence parsing, despite access to high-resolution inputs. Further analysis suggests that these failures are not fully resolved by increasing model capacity, but are closely tied to insufficient guidance in locating and using task-relevant micro-evidence. Motivated by this finding, we propose Micro-evidence Active Perception (MAP), a reference agent that decomposes queries into evidence-seeking steps, actively inspects candidate regions, and grounds its answers in localized observations. MAP-Agent improves micro-level perception by making high-resolution reasoning evidence-centered rather than image-centered. Together, UHR-Micro and MAP-Agent provide a diagnostic platform for evaluating, understanding, and advancing high-resolution reasoning in Earth observation VLMs. Datasets and source code were released at https://github.com/MiliLab/UHR-Micro.

Recently, post-training methods based on reinforcement learning, with a particular focus on Group Relative Policy Optimization (GRPO), have emerged as the robust paradigm for further advancement of text-to-image (T2I) models. However, these methods are often prone to reward hacking, wherein models exploit biases in imperfect reward functions rather than yielding genuine performance gains. In this work, we identify that normalization could lead to miscalibration and directly removing the prompt-level standard deviation term yields an optimal policy ascent direction that is linear in the advantage but still limits the separation of genuine signals from noise. To mitigate the above issues, we propose Super-Linear Advantage Shaping (SLAS) by revisiting the functional update from an information geometry perspective. By extending the Fisher-Rao information metric with advantage-dependent weighting, SLAS introduces a non-linear geometric structure that reshapes the local policy space. This design relaxes constraints along high-advantage directions to amplify informative updates, while tightening those in low-advantage regions to suppress illusory gradients. In addition, batch-level normalization is applied to stabilize training under varying reward scales. Extensive evaluations demonstrate that SLAS consistently surpasses the DanceGRPO baseline across multiple backbones and benchmarks. In particular, it yields faster training dynamics, improved out-of-domain performance on GenEval and UniGenBench++, and enhanced robustness to model scaling, while mitigating reward hacking and preserving semantic and compositional fidelity in generations.

Vision-Language Models (VLMs) exhibit systematic bias toward visual illusions, recalling memorized facts rather than perceiving actual visual differences. This paper presents a training-free framework for the 5th DataCV Challenge Task 1 at CVPR 2026, addressing this perception-versus-memory conflict through three complementary strategies:(1) illusion-aware image preprocessing that weakens illusion-inducing context via type-specific transformations (edge extraction, color isolation, morphological processing, and reference-line overlay), (2) anti-illusion prompt engineering guiding VLMs toward qualitative visual comparison, and (3) multi-vote ensemble that further improves robustness. Our method achieves 90.48% accuracy on the official 630-image test set using Claude (claude-opus-4-6) with 5-vote majority ensemble, and 98.41% on a human-verified subset. The approach requires no finetuning, relying solely on visual manipulation and prompt design. Our solution secured 2nd place in the challenge, only 0.47% behind the 1st-place solution. Code is available at https://github.com/jasminezz/sf-illusion-aware-vlm.git.

Success-driven social learning, in which individuals preferentially adopt the ideas and methods that appear most successful, is a foundational principle of collective behavior across systems ranging from ant colonies to scientific communities. But science is a particular kind of collective search -- one in which the quality of an explanation is itself difficult to assess. Is success bias adaptive in this setting? In agent-based simulations of collective theory building, we find that it is not. Scientists in our model systematically overestimate the quality of their own theories, creating an illusion of understanding: a persistent gap between perceived and actual performance. Success bias amplifies this illusion; communities that favor apparently successful theories explore a narrower range of possibilities, efficiently filtering out poor explanations but failing to discover better ones. This effect intensifies with problem complexity, as scientists in more complex environments become increasingly unable to assess how well their theories actually perform. Most strikingly, when agents optimize their social behavior to maximize the perceived success of their theories, they paradoxically undermine their actual performance, and produce levels of inequality that mirror those found in real scientific communities.

Vision-Language Models (VLMs) are frequently undermined by object hallucination, generating content that contradicts visual reality, due to an over-reliance on linguistic priors. We introduce Positive-and-Negative Decoding (PND), a training-free inference framework that intervenes directly in the decoding process to enforce visual fidelity. PND is motivated by our finding of an attention imbalance in VLMs, where visual features are under-weighted. Our framework introduces a dual-path contrast: a positive path that amplifies visual evidence and a negative path that constructs counterfactuals to penalize prior-dominant generation. By contrasting outputs from both paths during decoding, PND steers generation toward visually grounded results. Experiments on POPE, MME, and CHAIR demonstrate state-of-the-art performance without retraining.

Adaptive multi-agent systems (MAS) are increasingly adopted to tackle complex problems. However, the narrow task coverage of their optimization raises the question of whether they can function as general-purpose systems. To address this gap, we conduct an extensive empirical study of adaptive MAS, revealing two key findings: (1) topological overfitting -- they fail to generalize across different domains; and (2) illusory coordination -- they achieve reasonable surface-level accuracy while the underlying agent interactions diverge from ideal MAS behavior, raising concerns about their practical utility. These findings highlight the pressing need to prioritize generalization in MAS development and motivate evaluation protocols that extend beyond simple final-answer correctness.

Recent Vision-Language-Action (VLA) models report impressive success rates on standard robotic benchmarks, fueling optimism about general-purpose physical intelligence. However, recent evidence suggests a systematic misalignment between standard benchmark success and true embodied reasoning, raising the question of whether these high scores reflect genuine cognitive capability. To address this gap, we introduce BeTTER, a diagnostic Benchmark for Testing True Embodied Reasoning in robotic policies. BeTTER applies targeted causal interventions (e.g., spatial layout shifts, temporal extrapolation) while enforcing kinematic isolation to explicitly decouple high-level reasoning failures from low-level execution limits. Through systematic evaluation, we reveal that state-of-the-art VLAs catastrophically fail in dynamic scenarios, exhibiting severe lexical-kinematic shortcuts, behavioral inertia, and semantic feature collapse. Crucially, our mechanistic analysis traces these symptoms to fundamental architectural bottlenecks - such as capacity compression and myopic downsampling - which systematically degrade the model's foundational semantic representation. We demonstrate that highly static evaluation protocols effectively mask this degradation by allowing optimization to overfit to sensorimotor priors. Supported by real-world robotic validation, our findings confirm that this representational breakdown is not a simulation artifact, highlighting the critical need for future VLA paradigms to resolve the structural tension between high-frequency control and high-level reasoning.

On-policy distillation (OPD) is an increasingly important paradigm for post-training language models. However, we identify a pervasive Scaling Law of Miscalibration: while OPD effectively improves task accuracy, it systematically traps models in severe overconfidence. We trace this failure to an information mismatch: teacher supervision is formed under privileged context available during training, whereas the deployed model must report confidence using only deployment-time information. We formalize this perspective theoretically, showing that teacher-conditioned success is generally not a valid target for deployment-time confidence and that helpful privileged context induces entropy collapse and a systematic optimism bias. To address this, we propose a calibration-aware OPD framework, CaOPD, that estimates empirical confidence from model rollouts, replaces self-reported confidence with this student-grounded target, and distills the revised response through the same self-distillation pipeline. Experiments across various models and domains show that CaOPD achieves Pareto-optimal calibration while maintaining competitive capability, generalizing robustly under out-of-distribution and continual learning. Our findings highlight that capability distillation does not imply calibrated confidence, and that confidence should be treated as an essential objective in post-training. Code: https://github.com/SalesforceAIResearch/CaOPD

Like visual processing, language processing is susceptible to illusions in which people systematically misperceive stimuli. In one such case--the comparative illusion (CI), e.g., More students have been to Russia than I have--comprehenders tend to judge the sentence as acceptable despite its underlying nonsensical comparison. Prior research has argued that this phenomenon can be explained as Bayesian inference over a noisy channel: the posterior probability of an interpretation of a sentence is proportional to both the prior probability of that interpretation and the likelihood of corruption into the observed (CI) sentence. Initial behavioral work has supported this claim by evaluating a narrow set of alternative interpretations of CI sentences and showing that comprehenders favor interpretations that are more likely to have been corrupted into the illusory sentence. In this study, we replicate and go substantially beyond this earlier work by directly predicting the strength of illusion with a quantitative model of the posterior probability of plausible interpretations, which we derive through a novel synthesis of statistical language models with human behavioral data. Our model explains not only the fine gradations in the strength of CI effects, but also a previously unexplained effect caused by pronominal vs. full noun phrase than-clause subjects. These findings support a noisy-channel theory of sentence comprehension by demonstrating that the theory makes novel predictions about the comparative illusion that bear out empirically. This outcome joins related evidence of noisy channel processing in both illusory and non-illusory contexts to support noisy channel inference as a unified computational-level theory of diverse language processing phenomena.

Mental imagery vividness is a stable individual trait, yet whether imagined scenarios share relational structure across human and synthetic large language model (LLM) populations remains unknown. We applied psychological network analysis to vividness ratings from two validated questionnaires: the Vividness of Visual Imagery Questionnaire (VVIQ-2) and the Plymouth Sensory Imagery Questionnaire (PSIQ), across geographically and linguistically distinct human samples (Florida, Poland, and London; total N = 2,743) and six large language models (LLMs; Gemma3-12B/27B, their quantization-aware counterparts, Llama3.3-70B, and Llama4-16x17B). Imagination networks were constructed as regularized partial correlation graphs, with node centrality and community structure compared across populations using Pearson correlations and the Adjusted Rand Index (ARI). Human networks showed robust cross-population centrality correlations for expected influence, strength, and closeness (r = 0.31-0.93), and community detection recovered clusters aligned with VVIQ-2 scene contexts (ARI = 0.27-0.40) and PSIQ sensory modalities (ARI = 0.87-1.0). Betweenness centrality was unstable across all populations, consistent with its sensitivity to individual experiential history. LLMs failed to replicate human network structure: LLM-human centrality correlations were weak and largely non-significant after correction, and most LLM configurations produced degenerate single-cluster topologies (median ARI = 0). This failure was consistent across model architectures, parameter scales (12B-272B), and conversational conditions. We posit that these findings may be driven by human imagination networks reflecting memory organization accumulated through embodied experience, a representational structure that linguistic training alone does not reproduce regardless of model scale and conversational memory.

Recent work has argued normatively, on synthetic data, that evaluating survival models by discrimination alone (concordance index) yields systematically misleading model comparisons, because the metric ignores calibration and time-dependent accuracy. Whether this matters for real, published, non-clinical models has not been tested. We reproduce three published survival-ML models across three structurally distinct domains -- hard-drive failure prediction, peer-to-peer credit default, and user disengagement on digital platforms -- validate our instrument against the anchor paper's own synthetic experiment, and test five pre-registered hypotheses under a Holm-corrected family-wise error rate. Three of five reject (though one pre-registered threshold clears by a narrow margin). A model reproducing the published literature's discrimination almost exactly (C = 0.9595 vs. 0.958 reported) fails a formal calibration test at p < 0.001; a broad feature-ablation search finds no single attribute responsible for its discrimination, so the calibration failure is not a trivial shortcut artifact. A lender's estimated default risk is biased upward by roughly two percentage points, growing to nearly four in the riskiest segment, when loan prepayment is treated as non-informative censoring rather than a competing risk. A platform's churn model shows probability estimates that degrade with the horizon even as global discrimination stays within the pre-registered C-index band. A direct test of whether metric choice inverts model preference does not reject, though with limited power given two to three models per domain; the failure mode we document is better characterized as misplaced confidence in a chosen model than as choosing the wrong one. We release a pre-registered evaluation harness with full code and an annotated notebook, so these results can be verified independently and the audit extended.

Psycholinguistics studies show that human readers fall for coherence illusions: an incoherent discourse can seem coherent simply because a distractor matches what comes next. We investigate whether Dutch language models (6 monolingual and 4 multilingual) show the same behavior on texts that link back to earlier context with words such as 'again' and 'too'. First, we find that surprisal at the critical word tracks human acceptability judgments and eye-tracking data. Models are more surprised by incoherent continuations, but a matching distractor in the prior context reduces this surprisal. Second, attention entropy identifies heads that behave differently under coherence vs. incoherence. We find that ablating these heads shows transfer effects across experiments, suggesting a shared mechanism. Third, we introduce energy from the associative-memory literature as a metric to quantify discourse coherence. Taken together, our results show that coherence illusions arise in Dutch LLMs, with entropy and energy exposing mechanisms that operate across settings.