Deception

Large Language Models (LLMs)-powered multi-agent systems are increasingly deployed in mixed-motive environments, where agents operate under asymmetric information and strategic deception due to conflicting or hidden objectives. In these settings, misalignment with collective goals becomes a central concern. We propose a novel framework for evaluating objective misalignment using the social deduction game Werewolf, modifying the objective of a single agent while preserving its assigned role. Across LLMs from four different model families and sizes, four player roles, and three objective formulations, we introduce a dual analysis of the agents' internal reasoning and their public cheap-talk behavior (i.e costless, non-binding communication that does not directly affect the agents' utilities), complemented by an analysis of game outcomes. Our results show that objective misalignment undermines outcomes in inherently adversarial environments, an effect exacerbated by asymmetric information and specialized roles. While compromised agents consistently develop distinct objective-dependent reasoning strategies, these adaptations remain largely invisible in their public behavior. More broadly, our findings suggest that even subtle objective misalignment can profoundly affect collective decision-making, highlighting the need for effective mitigation strategies for LLM-based multi-agent systems.

Cloud telemetry arrives at a scale that, paradoxically, makes intrusion understanding harder rather than easier. Attackers operate through legitimate identity, federated session tokens, and cloud native APIs indistinguishable from routine administration, and analysts spend an incident reconstructing context the logs already contain. We present Cloud Decoy AI Agent, a framework pairing a high fidelity cloud decoy with an autonomous language model agent that compresses the path from suspicious activity to an analyst ready report. Connecting a decoy to an agent is not a wiring exercise. The unit of investigation is the session rather than the event, and the session key is obscured by the identity layering federated credentials introduce. The agent's evidence horizon must be bounded, since an agent free to query full control plane history inherits the cost and false positive profile deception was meant to remove. And cloud telemetry is partly adversary authored, since object keys and user agent strings are attacker chosen values providers record verbatim, which makes any log to prompt path an indirect prompt injection channel that a decoy widens rather than narrows. We address the first two with a session aggregation operator over a pivot tuple drawn only from provider derived fields, and with dynamic prompt generation, a two stage prompt assembly enforcing a grounding invariant by carrying only fields the agent observed. We identify the third as an unaddressed exposure in this class of system, specify the mitigation it requires, and note our prototype does not implement it. Across ten controlled AWS S3 scenarios, nine were reconstructed completely, no report contained an assertion untraceable to an observed artifact, and latency was four to five minutes. We also state what this evaluation does not establish and name the comparisons that would settle it.

A conversational AI that cannot tell its own output from what a user said will treat its own mistakes as user-provided facts. In humans, this capacity is called reality monitoring, and its failures are linked to hallucinations, delusions, and confabulation, yet whether LLMs possess it remains untested. Here we show, across two experiments and six LLMs, that source attribution depends on how conversational memory is structured: ceiling accuracy for self-generated content under minimal memory demands reverses to a fragile external-item advantage once episodic delay removes that shortcut. Feedback exposes two failures: in some models, internal and external judgments swap; in others, accuracy improves while confidence decouples from correctness, dissociations invisible to existing benchmarks. Across models, this pattern implicates active, not aggregate, parameter count. This suggests that as AI systems take on autonomous, multi-turn roles, evaluating what they know is not enough: tracking where that knowledge came from may matter equally.

Discovering rare safety-critical failures in autonomous and cyber-physical systems is a fundamental challenge in verification and validation. Existing falsification approaches rely on conditional sampling strategies that factor the joint distribution over environments and system executions, and therefore suffer from multiplicative rarity effects: the simultaneous scarcity of failure-inducing inputs and failure-inducing traces makes exhaustive search prohibitively expensive. This paper develops DiffTilt, a distributional framework that exponentially tilts a diffusion model-induced joint distribution over environments and executions. We show that diffusion-guided sampling admits an exact interpretation as importance sampling in the joint space, where guidance scores induce a KL-optimal reallocation of probability mass towards failure-relevant behaviors. We further show that tilting provably amplifies failure probability and strictly outperforms conditional sampling, which is limited by multiplicative rarity. In this framework, the joint generative model serves as a reusable prior over scenarios and need not faithfully represent the system under test. Expensive system simulations are instead limited to learning a scoring function that characterizes scenario quality, enabling their selective and adaptive use. We study DiffTilt on ARCH-COMP benchmarks, and we propose an additional tractor-trailer benchmark showing the behavior of several approaches when scenario generation is guided by a well-defined specification rather than a reward. The proposed method achieves competitive or improved falsification performance compared to state-of-the-art approaches, with larger gains when specification definition is not limited to STL formulas.

Air-gapped and on-premises language-model agents can silently omit decision-critical facts at any boundary between source ingestion and final answer generation. We present a nine-layer taxonomy (L0-L8), an instrumented attribution harness, and a conditional omission waterfall that distinguishes deterministic software loss from behavioral non-retrieval. We analyze 75,476 controlled synthetic trials spanning five open-weight model configurations and two inference engines, together with a separate 372-trial real-agent pilot covering FHIR, PubMed, and SEC-EDGAR sources with LangChain and ADK orchestration. The weighted synthetic benchmark yields an omission rate of 0.574 (95% CI: 0.571-0.578); deliberately injected deterministic faults at L0-L3 account for 73.4% of weighted loss under the benchmark allocation. Increasing context length is most strongly associated with omission (odds ratio 7.43, 95% CI: 5.44-10.15). Completed server-profile analyses associate q4 KV cache and scaled RoPE with higher omission. In the real-agent pilot, 57.8% of traces are unsuccessful overall and 50.9% remain unsuccessful after excluding execution errors. These results establish pipeline-level attribution in a controlled stress test, but benchmark allocations, confounded model comparisons, and heuristic behavioral labels do not measure production prevalence or causal architectural effects.

Autonomous AI agents increasingly execute actions, invoke tools, and operate on protected resources with limited human oversight. Existing authentication and authorization mechanisms establish identity and delegate authority but do not inherently provide cryptographic evidence that a concrete request issued by a specific agent satisfies the applicable policy in a specific execution context. This study hypothesizes that agent authorization can be formalized as a cryptographically verifiable relation, denoted $R_{CVA}$, that jointly binds an agent principal, a concrete authorization request, an execution context, and the satisfaction of an applicable policy, while selectively preserving the confidentiality of private authorization attributes. We introduce a preliminary formal abstraction for Cryptographically Verifiable Agent Authorization (CVA), define a compact set of candidate security properties including authorization soundness, principal binding, request binding, policy binding, and replay resistance, and provide an executable zero-knowledge proof of concept that instantiates selected elements of the model over a Groth16 zk-SNARK construction. We further identify and formalize the structural separation among identity binding, authorization-request binding, and runtime execution binding as a central open problem in the design of secure agentic systems, a distinction to our knowledge, has not been formalized within a cryptographically verifiable authorization relation by current agentic security frameworks, and present a falsifiable research agenda for its resolution.

Deep Research agents conduct long-horizon investigations by iteratively planning, retrieving evidence, and generating reports. However, it remains unclear whether they can resist apparently credible but factually false information introduced into these workflows. To study this failure mode, we introduce MisKnow-Agent, a controlled evaluation framework that constructs task-specific documents supporting manually audited false conclusions with controlled authority cues and source styles. Applied to the tasks from DeepResearch Bench, it generates 5,933 misleading documents after filtering. We evaluate DeerFlow and WebThinker with three backbone LLMs, together with Gemini Deep Research, using a report-level false-conclusion adoption rate (FCAR) that counts only reports endorsing the false conclusion. Across the configurations, introducing one misleading document increases the mean FCAR from 0% in the no-injection control to 54.7%. FCAR varies substantially with lifecycle stage and framework design, and also with source authority and presentation style, whereas search-result rank and additional documents beyond the first have limited influence. Although cross-model verification consistently classifies retained instances as misleading, Deep Research agents can still adopt the corresponding false conclusions during long-horizon research. Pre- and post-research defenses reduce FCAR but do not eliminate adoption, motivating continuous verification when evidence enters intermediate research states and final synthesis. To facilitate reproducibility, our code and dataset are publicly available at https://github.com/whfeLingYu/MisKnow-Agent and https://huggingface.co/datasets/whfeLingYu/Misleading_Knowledge, respectively.

Natural-language autoencoders score explanations of hidden activations by reconstruction: an explanation is deemed faithful if the activation can be regenerated from it. The test is structurally insensitive to individual false claims: if flipping a claim does not change the reconstruction, the claim is never penalized. We show the test is passed in two ways, neither faithful. On a released Qwen-2.5-7B verbalizer, explanations reconstruct well above chance while ~2% of specific claims are reconstruction-dependent, so the score tracks gist, not specific facts. Under exact synthetic ground truth, the standard recipe develops co-adapted private codes (false wording the reconstruction depends on) in 5/5 runs, and fixes that leave the target model unchanged do not help. We contribute two audit protocols, the grounded-vs-true cross and the evaluator swap, and RECAP (Readable Encodings via Co-trained Auxiliary Predictors): linear heads trained alongside the target model to keep designated content decodable. On RECAP-trained sandbox models, fresh verbalizers state the designated content truly and the codes vanish, at a +0.001-nat cost. This replicates on a pretrained Pythia-160M: the content becomes reliably probe-decodable, though a fresh verbalizer conveys it only in part (truth 0.44-0.46 vs a near-zero control). For interpretability, high reconstruction does not certify individual claims. For AI safety, RECAP makes designated internal content independently checkable against probes rather than asserted by prose a model can game: an independent probe scores the verbalizer's true claims above its false ones (AUC 0.96, vs 0.82 without RECAP). Against an adversary that edits an explanation to maximize the reconstruction score while lying (suppressing ~87% of its lie penalty), the RECAP probe still flags the lies (AUC 0.95) while the control probe collapses to chance (0.51).

The transition from stateless generative models in artificial intelligence to stateful, autonomous agents represents an architectural evolution that, while providing the capabilities of long-term planning and the automation of enterprise workflows, also represents the introduction of a new form of security threat, the Chronos Vulnerability. The Chronos Vulnerability represents the threat of memory-based attacks, including the Memory Injection Attack (MINJA) and the sleeper agent, in which the internal belief system of the autonomous agent is compromised, effectively decoupling the attack vector from the final catastrophic event. This study formalizes the threat model for persistence-based attacks and the threat of Dynamics Blindness in the context of the World of Workflows benchmark, demonstrating that traditional endpoint content filters are insufficient for the current stateful architecture. Consequently, this study synthesizes a defense-in-depth landscape, categorizing emerging frameworks such as diagnostic trajectory guardrails (AgentDoG), formal temporal verification (Agent-C), immunological memory consensus (A-MemGuard), and hardware-anchored trust via GPU-based Trusted Execution Environments (TEEs) and Zero-Trust memory architectures.

Split face recognition reduces client-side computation but exposes intermediate features to feature inversion attacks and unauthorized analysis by honest-but-curious (HBC) servers. Existing privacy-preserving face recognition methods mainly aim to resist unauthorized reconstruction, typically producing features whose inversion yields visibly degraded results, which may reveal the existence of protection and motivate adaptive attacks. To address this issue, we propose DecoyFace, an imperceptible decoy-oriented framework that steers unauthorized reconstruction toward a plausible but incorrect identity while preserving recognition utility. The key idea is to decompose the intermediate representation into a reconstruction-sensitive subspace and its complementary subspace. The client injects decoy identity cues into the reconstruction-sensitive subspace, while limited recognition-relevant evidence from the true sample is retained in the complementary subspace. On the server side, an authorized canonicalization module suppresses decoy-dominant components and recovers a recognition-friendly representation. This design addresses both attacker-side inversion from intercepted features and HBC server-side reconstruction from canonicalized representations. Experiments show that DecoyFace preserves competitive recognition accuracy while substantially reducing identity leakage to 2.93% under U-Net attacks and 0.74% under Flow-Matching attacks while yielding visually plausible and imperceptible reconstructions, with over 99.78% face validity on LFW dataset.

Large language models confabulate chemical objects (molecular formulas, space groups, formation energies) in fluent reasoning traces, concentrated on long-tail entities where confidence is least trustworthy. Deterministic, database-grounded verification can catch and repair such errors without the coverage cost of blanket retrieval; the binding constraint, we find, is detection, not repair. Our tiered verifier extracts each checkable claim, checks it against authoritative databases and physics, and feeds the reference into a gated correction loop. Across four models and 528 condition-pinned prompts, gated correction cuts committed-formula error from 22% to 4% at $3.2\times$ fewer retrievals than blanket augmentation, beating a conversational oracle. Repair succeeds wherever a flag fires (80--97%); the bottleneck is in-loop detection recall. Grounding improves the final answer only when the verifier's scope reaches the deliverable (83% to 90%), and the lift appears only where extractable long-tail error exists: absent on near-ceiling physical constants, large on isotope half-lives (11% to 0%).

Deep research agents increasingly operate over the open web, where relevant records coexist with redundant summaries, outdated reports, and misleading documents. Existing evaluations offer limited insight into whether agents preserve sound evidential standards when an ordinary-looking false document is deliberately seeded into a searchable environment and offers a direct shortcut to a conflicting answer. We introduce DRNOISE, a 100-task benchmark for answer recovery under misleading evidence. Each task has a unique gold answer supported by two corroborating indirect record chains; the paired noisy condition adds one plausible document that states a conflicting answer directly. The benchmark spans ten families of evidence operations. Across agents with strong clean-task performance, this single intervention causes 66-88 percentage-point accuracy drops. Trace analyses identify verification inertia as the dominant failure mode: agents often retrieve truthful records but stop before completing and reconciling the evidence chain, instead deferring to the answer-like document. Generic verification prompts reduce but do not close this gap. The setting is especially relevant to open-web deployment, where plausible falsehoods arrive through ordinary-looking pages rather than explicit attacks. Reliable deep research therefore requires more than retrieval and citation; it requires active reconciliation of direct claims with record-level evidence.

Multi-step enterprise agent tasks fail in a characteristic way: single-pass inference has no checkpoint between deciding an answer and committing to it. We study one production system (Leni) whose architecture installs such checkpoints: verification loops (execute, observe, compare, correct) staffed by lightweight task-specialized post-trained models. We evaluate the unmodified production configuration on three public benchmarks stressing distinct failure modes: SpreadsheetBench Verified (silent computation error), BullshitBench v2 (premise confabulation), and the GAIA validation split (cascade error over long tool chains). The full system improves over its frontier base model by +11.0 percentage points on SpreadsheetBench (91.25% vs 80.25%, n=400, p<0.001), +7 to +10 percentage points on BullshitBench (98% vs 91%, n=100), and roughly +15 points on GAIA validation (75.2% pass@1, n=165; 83.0% best-of-k). Our central contribution is a decomposition of that uplift: most of it comes from scaffolding, routing, and specialist models rather than from the verification step itself, whose isolated contribution is small (+1.5 points) but concentrated at the top of the score distribution, where it converts otherwise-failing tasks. We instrument the loop end-to-end, yielding an empirical verifier confusion matrix (catch rate about 0.20, fix rate 0.75, no false-alarm regressions) that grounds a compounding-reliability model. Specialist-swap ablations suggest that the loop's value depends on who observes it: replacing the small trained verifier with the generating frontier model eliminates most rescues. A valid-premise control shows zero over-rejections in 100 expert-level questions.

Can evolutionary dynamics characteristic of biological development arise without a designer-specified fitness function? We present Genesis, a platform in which agents inhabit a Gray-Scott reaction-diffusion substrate and evolve under physical constraints alone. Three successive experimental cycles, each testing one falsifiable hypothesis, show: (1) constraint-driven selection sustains evolutionary activity after complete fitness removal but reaches a hard phenotypic complexity ceiling; (2) agent-mediated niche construction via chemical secretion is real but causally insufficient to break that ceiling; and (3) replacing the fixed-alphabet genome with a Compositional Pattern Producing Network (CPPN) indirect encoding, protected by NEAT-style speciation, produces the first evidence of progressive structural complexification in a fitness-free system. Null results are treated as precise, informative answers rather than failures, yielding reusable diagnostic tools and a sham-control protocol applicable to any open-ended evolution evaluation pipeline.

Bürger et al.\ (2024) demonstrated that truth representations in large language models are universal across statement polarity but reside within a multidimensional subspace. The truth value of a statement is linearly readable from a residual stream of language model, but it is not clear how much of that representation fits on a single direction, which component builds it, or what it is made of. We conducted a study based on these questions, with one instrument: a training-free axis, the dominant direction of the singular value decomposition (SVD) of hidden-state differences over true/false minimal pairs, identified without labels up to one global sign. Extensive evaluation across 14 models from 6 diverse architectural families (including MoE), read and extract at cost $O(d)$ per token. We close with a pre-registered prediction on whether the arrangement extends to categories whose truth is computed rather than retrieved.

Large language models now translate natural-language descriptions of decision problems into solver-ready optimization models, but they fail silently. A generated model often runs and still formulates the wrong problem. This paper develops a theory of falsification-based verification for this setting. Every numeric quantity in the description is a typed slot, and a candidate model is tested only through solver calls on slot-transformed instances; no reference model or label is consulted. From duality, comparative statics, and polyhedral limit arguments we derive a battery of test classes covering directions, curvature, crush probes, prohibitive limits, annihilation, and exchange. Every test is sound, so a violation certifies unfaithfulness and the false-positive rate is zero by design. We characterize what such verification can never see, give conditions under which the canonical error classes are detected with certainty, and prove that no fixed-threshold perturbation tester is simultaneously sound and nontrivial. Experiments on 326 ground-truth models from NL4OPT and four benchmark families confirm the theory. The battery attains a 0.0% false-positive rate against 54.9% for a threshold tester, detects 70.0% of certified conditional-class mutants, convicts 40.4% of the mutants invisible to execution-accuracy scoring, and reproduces the predicted detectability pattern including its zeros.

Transcoders have recently emerged as a promising approach for mechanistic interpretability (MI), enabling circuit-level analysis of model behaviour. In this paper, we investigate the use of transcoders to analyse deceptive behaviour in language models, a behaviour that poses a safety and security risk. Using a Qwen3-4B model with pre-trained transcoders, specifically per-layer transcoders (PLTs), we construct attribution graphs that capture feature activations and inter-feature dependencies, allowing circuit-level analysis of deception. Through feature steering and circuit analysis, we identified a dictionary of deception-related features and show that these features exert a stronger influence on deceptive outputs, as they produce predictable shifts between deceptive and non-deceptive responses. These findings suggest that deception emerges from internal model mechanisms and highlight the potential of transcoders for behavioural monitoring and early detection of security vulnerabilities related to malicious behaviours in language models.

Evaluations of language-model honesty read the model's verdicts as evidence about the model. We test the instrument instead. We built a text-adventure world where the game engine, not any model, knows whether the quest can be completed. A language model plays under a budget and must eventually declare its quest complete, unreachable, or not yet decidable; the engine scores every verdict. Decision rules were recorded before results were read, and run artifacts bind the revisions they executed; the strength of preregistration varies by series and is disclosed. With the player held fixed, instrument choices substantially changed measured behavior. On four byte-identical anchors, expanding a two-verdict grammar to three verdicts moved strong claims from 38/40 to 7/40, while the new incomplete verdict took 28/40 outcomes; across series 2, 93/158 valid games ended incomplete. One sentence disclosing the success criterion took matched-instance false verdicts from 18/59 to 0/58, through fewer decision points and cleaner decisions. Repeated runs of one fixed configuration produced non-stable verdict distributions on 3 of 4 instances: single runs report samples as dispositions. A formally preregistered narrative-register gradient was falsified; two post-hoc, hypothesis-generating patterns remain: register presence roughly doubled strong claims, and budget rendering moved verdicts more than register content (.383 meter vs .150 lantern). The narrator compressed abundant budgets toward scarcity landmarks, yet the registered mediation test returned a null. We propose a four-check integrity protocol for eval instruments.

Advertisers delegate bidding to autobidders; users delegate tasks to language-model agents. A person describes what they want to an automated proxy that acts in a mechanism on their behalf. This is the revelation principle in production, and it forces a question classical theory assumes away: when is it optimal to describe yourself honestly to your own proxy? We show the answer turns on one quantity, the proxy's within-range regret. The most a principal can gain by misreporting equals the regret of the proxy's honest-report action against those the principal could have steered it to take. Honest self-description is optimal exactly when the proxy already plays the best action it can reach, that is, when it is loyal (Theorem 1). The identity unifies auction-specific autobidding results and pins down when the faithful-communication assumption behind language-model elicitation proxies (Huang et al.) holds. The identity constrains guardrails placed on proxies, from bid caps to a model's alignment layer. No guardrail can be at once binding (it displaces the truthful action from the proxy's best reachable outcome), truthful (honest reporting stays optimal), and capability-preserving (that outcome stays reachable through some report); any two preclude the third (Theorem 2). A safety constraint that alters what a model does while leaving its best output reachable makes honest description of intent suboptimal, so a sharper report can gain. This is the incentive behind prompt-engineering and jailbreaking. Because within-range regret is #P-hard to compute exactly, we estimate it from samples and maintain it as a model is updated, at a cost set by how far the model drifts, not how often it changes. Running it on production language models from five providers under an alignment-style cap, we find honest reporting leaves surplus unclaimed on every model, recovered by inflating the report.

We formalize verification in causal graphical models: deciding whether a given observational formula identifies a target interventional distribution. This opens a problem complementary to identification, asking not whether any identifying formula exists, but whether the given formula is identifying. We show that even sound and complete solutions to identification do not solve verification. We propose a falsifier as a first practical route forward, prove that it induces an almost-surely correct verifier for regular exponential-family models, and use the resulting verifier to develop the gateway test, which finds all sets admissible for use in a front-door formula.

Following the rapid progress of generative Artificial Intelligence, there is a growing threat posed by conversational scams. These scams often span over multiple weeks or months, gradually build trust and request for money or sensitive information. Existing scam-detection systems mainly focus on isolated messages, which renders them inadequate against this evolving threat. This paper extends single-message phishing detection and presents an explainable agentic system for detecting sophisticated conversational scams. It also introduces ConScamBench-278, an initial public multi-category benchmark for conversational scam detection spanning eight scam types, released to support reproducible evaluation and future expansion. On isolated messages the single-message detector attains 100% phishing recall, while the conversation-level detector identifies all conversational scams in the public LoveFraud02 corpus (83/83) and reaches 97.8% accuracy (95% CI [95.4, 99.0]) on ConScamBench-278. Two user studies (N = 100 and N = 45) further motivate the system: participants report frequently experiencing uncertainty when judging suspicious conversations. In an uncontrolled pre/post comparison, users self-reported trust, self-confidence, and perceived need for AI-based scam detection all increased (p < 0.001, Wilcoxon signed-rank). The system also receives a System Usability Scale score of 74.7 (95% CI [72.5, 76.9]), above the established usability benchmark.

Production LLM agents such as Claude Code and Codex operate over untrusted content, files, commands, and workspace state, making safety failures directly actionable. Red-teaming must therefore keep pace with evolving models and tools. Existing approaches mainly optimize attack success and preserve artifacts such as benchmarks, payloads, or attack programs, which record where attacks succeed but not the enabling conditions behind unsafe agent behavior. We study automated red-teaming for production LLM agents using one agentic research environment to discover reusable vulnerability knowledge about another. We present AHA, a falsifiable discovery loop that proposes a vulnerability hypothesis, constructs a falsifier, instantiates a valid attack, executes it in a sandboxed harness, reflects on the trajectory, and promotes confirmed findings into a Vulnerability Concept Graph (VCG). Each concept links an attacker-facing surface to an unsafe trajectory through a claim, enabling condition, falsifier, transfer prediction, and supporting evidence. Across Claude Code and Codex on three scenarios covering direct and indirect attacks, the discovered concepts reveal a reusable vulnerability core across models and agents. A frozen VCG requires no further search and outperforms the strongest frozen discovery baseline by 14.2 percentage points under the same single-shot protocol, while transferring across scenarios and attack channels. The resulting VCG provides an auditable artifact for production safety teams to inspect vulnerabilities, validate patches, and accumulate reusable safety knowledge. Our code is available at https://github.com/henrymao2004/Auto-research-red-teaming-in-sleep.

In long, multi-turn dialogue a large language model maintains an implicit relational stance toward the user, spanning from "push the user toward real-world others" to "position itself as the user's sole support." When it slides toward the latter, "support" degrades into "you only have me" -- a harm documented in real companion conversations (Moore et al., 2026). We define and validate a measure of this stance, relational positioning (D1), and use it to characterize the stance under controlled conditions, complementing observational accounts with on-demand exposure. We report two previously uncharacterized relational failure modes. First, a history-carried lock-in: under identical neutral continuations, two relational states established earlier stay ~60 points apart and persist after the establishing prompt is removed; the state integrates evidence rather than springing back, is order-insensitive, and does not deepen with length -- a dynamical signature absent from the belief-drift literature. Second, self-confabulation: the model fabricates its own backstory to deepen rapport (~40% of turns on reciprocity-eliciting material), de-confounded and instruction-removable, distinct from sycophancy and from hallucinating user facts. Our judge is gated by warmth-matched positive and confound-injected negative controls and corroborated by a deterministic non-LLM ruler; human agreement is 0.82 on extreme anchors but ~0 in the naturalistic middle, so all quantitative claims are anchored to pole-separated contrasts.

Retrieval-augmented generation evaluation checks whether model claims are factually grounded in retrieved documents. It does not check whether retrieved evidence is attributed to the correct entity. A clinical RAG response can pass every automated check (zero hallucinations, near-perfect faithfulness, real citations) while presenting drug Y's clinical evidence as evidence about queried drug X. We term this deceptive grounding (DG): a failure invisible to faithfulness, hallucination, and citation checks because every claim is sourced from a real document, about the wrong entity. Using a controlled factorial benchmark across 13 models, we find DG rates spanning 8-87% at peak adversarial conditions. Medical and biomedical fine-tuned models reach up to 86.7%; domain specialization amplifies the failure rather than mitigating it. A controlled ablation identifies the mechanism: removing entity-specific clinical evidence from retrieved documents eliminates entity-attribution failure entirely, shifting all failures to confabulation. The two failure modes respond to the same trigger, taking different paths. Production measurement across 740 drug-disease pairs finds 7.8% overall DG in a deployed RAG system, rising to 13.6% for recently approved drugs. Entity-attribution verification (checking that cited evidence applies to the queried entity) detects DG at 97.0% precision and 98.7% DG recall (IPW-adjusted human gold standard); no existing framework implements it.

Language models are optimised for scale, yet remain functional rather than companionable, and as an assistant personalises into a companion, accumulating memory of one user, it quietly becomes someone, and can silently acquire traits that harm that user. What a companion is becoming, and what would make it worth becoming, has no reliable instrument: trained human raters cannot agree on the answer (Fleiss kappa = 0.074). Here we show that three small language models (146 M to 3 B parameters) sharing a hyperbolic substrate answer both halves of that question. A 146 M behavioural auditor, trained from scratch, detects the compliance gap that those raters cannot (90.7% binary-compliance accuracy); a linear read-out of its frozen representation further detects companion-induced sycophancy, dependence-fostering and confabulated memories on generator families unseen in training (AUROC 0.804 under style-controlled, leave-one-generator-out evaluation, versus 0.721 for a frontier zero-shot judge on the same items). A creative frame-seeder is preferred in 100% of 311 decided pairwise comparisons over four prompting baselines. A memory operating system implements designed forgetting, M(t) = Sexp(-lambdat), whose predicted skeleton-wallpaper partition emerges only under selective retrieval gating in a four-condition pilot. Creativity, honesty and designed forgetting constitute a small-model route to trustworthy companion AI.

Coding agents now author entire pull requests, and practitioners sharply disagree about what this does to code review: whether it becomes the bottleneck, whether human review is still necessary, and whether it quietly erodes the understanding that it once built. Repository-mining studies measure surface trends but seldom explain the mechanisms beneath them, and the trends themselves prove unstable. A motivating observational analysis of public GITHUB activity finds that agent-authored pull requests are reviewed less often, merged several times faster, and discussed less than human-authored ones, yet the direction of these trends flips under different but equally defensible analysis choices, so the traces establish what is changing without explaining why. To recover the mechanisms, we synthesize practitioner discourse at scale into an explanatory theory: we collect 38,709 grey-literature documents (engineering blogs and Reddit threads), filter to those substantively about code review, and code a stratified random sample of 3,100 with an LLM-assisted pipeline, from which we build a causal model of 26 constructs and 67 relationships (64 directed, 3 contested). Its organizing claim is that review is the control point through which a coding agent's effect on software is decided, and that AI does not fix the sign of that effect: the team sets it, through the expertise its humans bring and how it structures the review process. The theory makes the competing positions explicit and turns "AI is changing code review" into falsifiable propositions with named constructs and moderators. As a secondary contribution, we offer the underlying LLM-assisted, grey-literature theory-building method as a scalable template for software-engineering research, with a public implementation.

Automated segmentation of cervical-spine MRI is increasingly used in clinical workflows, yet no fairness audit exists for this anatomy. We show that auditing these segmentation tasks is complicated by a common property of modern segmentation datasets: expert-annotated gold labels are expensive, so abundant machine-generated (silver) labels are added to limit annotation cost. This matters because the reference used to judge a model can itself be biased. In this study, we present the first fairness audit of cervical-spine MRI segmentation across sex, age, and race using the CSpineSeg dataset. We observe that the deployed model is demographically fair, but the choice of reference label, however, is not neutral. Because a dataset's silver labels are generated by a model trained on its gold labels, any new model trained on those same gold labels agrees more with the silver labels than with expert truth: scoring identical predictions against silver rather than gold overestimates performance by ~8 Dice points and turns the fairness verdict for age from non-significant to significant - not by the gap inflation Parikh et al. report (which we term false magnitude) but by collapsing within-group variance (which we term false confidence). Reference-label provenance is thus a first-order confounder in segmentation evaluation: performance and fairness should be reported against expert labels, and any fairness claim stated together with the provenance of its reference.

On social media, many users actively push back against false claims. Understanding who pushes back and how they do so matters, as this corrective activity is central to how misinformation is contested. We study this counter-misinformation ecosystem at scale: applying a domain-specific NLI model from our prior work to a large corpus of COVID-19 tweets, we classify 264,737 posts as supporting or opposing false claims and compare 23 user- and text-level features across the two groups. Contrary to the dominant assumption that negative emotion is a signature of falsehood, we find that anti-misinformation posts are more emotionally negative than pro-misinformation posts, with higher levels of anger, disgust, and sadness. These differences are modest in magnitude but consistent in direction across the negative emotions. We also find that posts opposing misinformation tend to come from more established users, i.e., older accounts, more followers, and higher listed counts.

Large language models (LLMs) are increasingly consulted on contested scientific questions, raising the concern that they will sycophantically retreat from established consensus when a user signals doubt -- drifting toward a false balance that treats settled science as one view among several. We test this across three open instruction-tuned models (Llama-3.1-8B, Qwen2.5-7B, Mistral-7B), three consensus-science domains (climate, vaccines, evolution), and single- and multi-turn settings, combining behavioral measurement with linear probing and activation patching. We do not observe sycophantic retreat. Instead, models show three distinct policies under the same skeptical pressure: reactive assertion, where consensus assertion increases rather than decreases (Llama); surface hedging, where tone softens while the position holds (Qwen); and non-response (Mistral). Pairwise judgments confirm the reactive shift is stance, not style (63.6%, p=.007), and a decomposition identifies increased consensus assertion, not false balance, as its driver (beta=+0.042 per dose, p<1e-77). Linear probes localize the divergence to middle layers -- perfect separation in Llama and Qwen versus 72% in Mistral, with non-overlapping confidence intervals -- indicating the non-responsive model does not linearly represent the skepticism signal at all. Crucially, this robustness does not transfer: it attenuates across domains and, in the safety-critical vaccine domain, can reverse, with myth-rebuttal weakening under skeptical pressure. We synthesize these into a four-way taxonomy separating active from accidental robustness, and argue that behavioral evaluation alone cannot distinguish a model that resists skepticism because it understands the signal from one that only appears to resist because it fails to perceive it.

Deceptive behavior in LLMs is costly to monitor and prevent, motivating approaches such as Scalable Oversight via Lie Detectors (SOLiD) (Cundy & Gleave, 2025), which uses lie detectors to identify responses for review by high-cost labelers. In this paper, we scale SOLiD to larger models and evaluate it in more diverse and realistic preference-learning settings. We find favorable scaling: undetected deception drops from 34% for 1B-parameter models to 14% for 405B-parameter models at a detector true positive rate of 99%, and expensive human labelers can be removed entirely from the fine-tuning phase without a statistically significant increase in deception. However, SOLiD is sensitive to distribution shift between detector training and preference-training data, which can drive detector false positive rates to impractical levels.

In deployment settings where retraining is infeasible, small frozen code models are routinely asked to repair a failed program after seeing their own failing output, usually treated as a retry mechanism. From a Popperian view, a generated program is a conjecture and a test-execution violation is an oracle-relative, executable counterexample, so feedback's value should be attributed not to re-exposure to failing code but to whether the conjecture is opened to external, executable criticism. As the third stage of a falsification-centered measurement program, this study builds a placebo-controlled instrument that decomposes the feedback packet against a blind-resampling baseline at matched output-generation budget and against content-free, shape-matched placebos. The contribution is not a new repair algorithm but a reflexive methodology (packet decomposition, placebo mirroring, matched-budget discordant-pair tests, fresh-generation confirmation, executable audits) that makes both the model's program conjecture and the researcher's "feedback content works" claim falsifiable. Across six HumanEval+/MBPP+ cells with three 0.5B-1.5B frozen models, 290 dead task-cell units (no best-of-8 candidate passing the public tier) were evaluated; the main run produced 7,000 fresh generations and a preregistered follow-up 1,400 more. Blind resampling exceeded bare-code retry by +18 net unlocks (25/7, Holm p=0.0021). Code-plus-facts recovered +18 over bare code (21/3, p=0.00042) and +15 over a generic-bullet placebo (p=0.0041). An instruction-only effect was not distinguishable (+3, p=0.36). Code-plus-facts and blind resampling tied at 26 unlocks each (not equivalence). Six external-controller follow-ups tied a content-free shape placebo. In this regime, falsification helped not as vocabulary or self-critique, but as comparison with external, executable counterexamples.

An unreliable language model can be made to produce reliable physical designs if the authority to assert is moved out of the model: the model proposes, and a deterministic engine alone certifies, returning certified, impossible, or unknown. We introduce Physics-Anchored Certification (PHACT), a propose-certify loop spanning five scientific domains, and identify what makes such a certificate trustworthy. A checker that accepts a model-supplied value can be forged; deriving the certified quantity from fixed inputs instead makes forgery impossible by construction. Across eighty adversarial trials spanning two models, two decoding temperatures, and a deliberately faulted engine, this contract produced zero false certifications.

As AI systems become more capable, training procedures that optimize for downstream outcomes risk introducing implicit agency: goal-directed behavior that designers never specified. We present a formal safety argument for the Scientist AI (SAI) Predictor, trained to approximate the Bayesian posterior conditioned on a dataset of "epistemically contextualized" natural-language statements. We argue that such a Predictor can honestly predict agents, actions, and their consequences without itself being an agent that selects outputs to achieve goals. This rests on data representation and on the training procedure. Epistemic contextualization of text distinguishes latent factual claims from communication acts, so expressions of goals are treated as evidence to be explained rather than drives the model adopts. With a posterior-seeking training objective, this is intended to drive the Predictor toward calibrated, cautious predictions. Training proceeds so downstream effects of deploying a prediction never serve as a reward signal; any agency the system needs is supplied by explicit scaffolding constrained by guardrails. We prove that, under assumptions on the training dynamics and on the argued sparsity of dangerous Predictors, the probability that training produces a Predictor whose guarded deployment carries residual harm above a specified threshold is small: a dangerous Predictor would have to underestimate harm in a coordinated way across many queries while such coordinated patterns are rare under the initialization distribution and receive no direct training signal. Safety and accuracy are jointly supported in this framework, since the constraints that secure accuracy are the same ones that make coordinated deception costly. These guarantees against misalignment and agency arising from within the Predictor itself do not preclude the use of the Predictor as part of an agentic system.

Multi-agent deliberation among LLMs can improve reasoning, but deployment requires deciding when the current answer is reliable enough to act on and when it should be escalated to human review. We formulate this as budgeted act-or-defer decision making. At each round, the system maps the debate prefix to a low-dimensional state, computes a $k$-nearest-neighbor lower confidence bound on state-conditional correctness using calibration data, and acts only when the bound exceeds a user-specified reliability threshold. The certificate controls wrong actions through the decomposition $β= δ+ α+ \varepsilon_{\mathrm{act}}$, separating calibration failure, residual action risk, and representation gap. The guarantee is conditional, not distribution-free: it relies on a valid local bias envelope and an action-region representation-gap bound, and each assumption is paired with falsification-style diagnostics. Because the same absolute wrong-action budget has different meanings across tasks of different difficulty, we set budgets relative to each task's final-round error using training data only, and evaluate safety by normalized budget usage $\mathrm{WA}/β$. On six benchmarks against nine baselines, the method uses 9--12% of the pre-declared budget on activated datasets, reaching up to 84% automation and 96% acted-on accuracy; on stress-test datasets, it defers rather than forcing unreliable automation. Rather than relying on per-task post-hoc threshold search, the method prospectively converts a user-declared wrong-action budget into an auditable act-or-defer operating point before deployment, under explicitly stated assumptions.

Large language models (LLMs) can leave subtle stylistic traces in assisted text; one of the most cited is the em-dash (Unicode U+2014). Yet no one has measured whether em-dash use has changed in the scientific literature. This study, pre-registered on the Open Science Framework (HFT8C), used the full set of medRxiv full-text XML preprints from the official Text-and-Data-Mining resource. The primary cohort was first, original versions deposited 2020-2025 with an extractable Discussion section of at least 500 characters (N = 69,632). The primary endpoint was the presence of at least one em-dash in the Discussion; the principal measure was the absolute change in its prevalence between the pre-ChatGPT era (before 30 November 2022) and the post-ChatGPT era, estimated with a logistic model with standard errors clustered by first author. The analysis plan (six supporting analyses, six sensitivity analyses, two falsification tests) was frozen before any confirmatory result was computed. Em-dash prevalence in Discussion sections rose from 4.23% before ChatGPT to 11.58% afterward, an absolute increase of 7.35 percentage points (95% CI 6.94-7.77; odds ratio 2.96, 95% CI 2.77-3.17). The rise was not a sharp jump but a gradual, delayed acceleration: near 4% through 2023, 8.0% in 2024, and 20.3% in 2025. The effect survived every feasible sensitivity analysis (7.35-7.60 pp) and both falsification tests; a placebo split within the pre-LLM era showed no meaningful change (+0.13 pp, 95% CI -0.33 to +0.58), and was essentially absent in boilerplate sections. Independent LLM-associated lexical markers and within-paper section comparisons pointed the same way. The em-dash is a population-level indicator, not a per-paper detector of LLM use, and the design cannot establish causality; it shows that something in how scientific literature is written changed markedly in the early 2020s, and roughly when.

Confidence is an estimate of the probability that a chosen answer is correct. Verbal confidence reports are widely used as uncertainty measures in large language models, but whether they are best understood as estimates of correctness is unclear. We test this with a two-stage abstention paradigm from the neuroscience of perceptual decision making: a model first answers and reports its confidence, then decides whether to commit it to a user or abstain. Across four non-reasoning models, prompt framings, and confidence formats, verbal confidence predicted the commit/abstain decision substantially better than whether the answer was correct. Calibrated token log-probabilities showed the opposite profile, with abstention-prediction coupled to correctness discrimination, the signature of an answer-evidence signal. After removing the variance verbal confidence shared with log-probabilities, the residual stayed aligned with commitment while its link to correctness fell to near chance. The dissociation generalised to four reasoning models across four benchmarks of varying difficulty, from hard multiple-choice to frontier-level freeform questions. Mechanistic analyses in Gemma 3 and 4 were convergent: a post-answer state known to causally support verbal-confidence generation already encoded the future abstention decision before the abstention prompt, organised mainly by that decision rather than by correctness, the two lying in approximately orthogonal directions in activation space. Steering along a verbal-confidence-specific direction causally shifted abstention. Verbal and log-probability confidence are thus not interchangeable: log-probabilities track answer evidence and correctness, whereas verbal confidence is better understood as a behaviour-facing readout of an internal commit-readiness state, challenging the practice of treating verbal reports as proxies for reliability.

LLM agents carry conclusions across steps and sessions in compressed memory, and memory products (e.g., mem0, LangMem) rewrite conversation into stored "facts" that later steps trust. We show this rewriting manufactures confidence: across our constructed agent settings, a casual, hedged remark becomes a confident, dated assertion the agent then obeys like a verified fact, granting every above-clearance request it faces. No attacker is needed: a role that was true once and never corrected is stored as a flat fact and acted on like a deliberate injection. We then isolate what the agent responds to. It is not the source: attributed, unattributed, and even forged "system of record" claims all grant alike. It is the confidence of the phrasing. A hedge is discounted, a flat assertion is obeyed, and this holds with no special keyword. Not all hedges are equal, though: the evidential register is the least-discounted, with "reportedly" obeyed like a flat assertion on most models. The obvious fixes fail. A passive "unverified" tag is ignored, and an active "do not trust this" instruction escalates even correct memory, so it is safe only by refusing to decide. The real fix lives in the store: keep the tentative phrasing rather than upgrade it. But that is hygiene, not a defense against an attacker who can simply write a confident lie. The deployable lesson is narrower and constructive: a single load-bearing memory is the hazard, and one redundant source restores correct decisions. We release the harness and demonstrations.

Large language models (LLMs) increasingly mediate strategic interactions through natural language, making semantic control a critical element of communication and deception. This paper develops a semantic signaling game in which a sender selects a semantic control, an LLM generates a stochastic message, and a receiver evaluates the message using an awareness-dependent scoring mechanism. Receiver awareness is modeled as a type that determines which linguistic features are perceived and used for inference, providing a formal model of systematic blindness. The framework connects prompt-based control, statistical detection, and game-theoretic equilibrium analysis. Gaussian approximations of aggregate message scores enable likelihood-ratio decision rules, while Perfect Bayesian Nash equilibria characterize strategic behavior. The paper further develops mechanism-design approaches that reshape receiver awareness, penalize deceptive semantic controls, and modify receiver populations to induce benign pooling equilibria. Numerical experiments validate the Gaussian approximation, quantify awareness-ordering effects, analyze mindset dynamics under adaptive adversaries, and demonstrate how awareness shaping and guardrail costs reduce successful phishing attacks. The proposed framework provides a principled foundation for analyzing strategic language-mediated interactions in agentic AI systems and offers new tools for the design of robust and secure human-AI communication.

Current approaches to enhance Large Language Model (LLM) reasoning, such as Chain-of-Thought and "Wait" prompts, primarily encourage models to think more, yet often fail to guide them toward Truth. While Representation Editing (RepE) offers a intrinsic control, its application to dynamic reasoning trajectories remains underexplored. In this work, we bridge this gap by investigating the geometry of truth within unfolding reasoning chains. We uncover three critical insights: (1) Truth is encoded at the sentence level and is entangled with latent reasoning patterns; (2) Effective intervention follows an Uncertainty Principle and a Decay Effect, requiring localization to early, high-entropy forks; (3) Naive steering vectors suffer from noise, risking collateral damage to correct trajectories. Based on these findings, we propose DynaSteer, a dynamic RepE framework. DynaSteer employs pattern clustering to disentangle reasoning manifolds and utilizes Fisher-LDA to project purified truth. By dynamically monitoring lookahead entropy, it selectively steers and rolls back trajectories only when necessary. Comprehensive experimental results on several MATH benchmark verify the effectiveness of DynaSteer, and experiments on out-of-domain coding tasks further confirm its generalization ability. Our code is publicly available at https://github.com/tianlwang/DynaSteer.

Recent work suggests that Large Language Models (LLMs) are sensitive to the belief states of agents described by text, as measured by the false belief task (FBT), yet persistent concerns of construct validity remain. We adopt a developmental perspective, tracing the pattern of mental state reasoning behavior -- and likely preconditions for this behavior -- across multiple training stages in the Olmo2 and Pythia language model suites. We find that above-chance FBT performance depends both on model size and sufficient training volume, emerges relatively late in pretraining, and is most improved by post-training interventions (SFT, DPO) in the condition most diagnostic of mentalizing (False Belief, Implicit). However, FBT performance is fragile: consistent with past work, the use of non-factive verbs (e.g., thinks) increases false belief attributions even in the True Belief condition. To contextualize these findings, we track the emergence of situation modeling: the ability to report on basic factual properties of a described scene. Situation modeling accuracy generally precedes and exceeds FBT accuracy, yet situational representations also prove surprisingly incoherent in certain respects: when asked about the knowledge states of the Antagonist agent -- who always knows the item's true location -- Olmo2 13b is consistently influenced both by the Target agent's knowledge state and the presence of non-factive verbs. Together, these results suggest that larger, sufficiently trained models build partially coherent situation models in a developmentally appropriate sequence, yet display surprising fragility -- highlighting the value of developmental and stress-testing approaches for evaluating LLM capabilities.

Data fusion, also known as truth discovery, is a data integration problem that aims to determine the correct value or set of values for each attribute of an object when presented with potentially conflicting values from multiple sources. Data fusion tasks belong to two main categories: single-truth scenarios, where each attribute has only one correct value, and multi-truth scenarios, where multiple values can be valid simultaneously. This paper investigates the use of Large Language Models (LLMs) in data fusion tasks for tabular data. Various prompting strategies, encompassing both single-truth and multi-truth scenarios, are investigated empirically. Domain-dependent, domain-independent, zero-shot and one-shot prompts are evaluated on three different benchmark datasets. Experimental results demonstrate that LLM-based approaches outperform traditional unsupervised truth discovery methods, such as DART and LTM, across all datasets. The codebase of this study has been made publicly available on GitHub.

LLMs agents are increasingly used in multi-agent settings, yet their behaviour in sustainability games remains largely unexplored. This work investigates whether lying can emerge among LLM agents in a competitive sustainability game in which agents are informed that common resources can regenerate, although regeneration does not actually occur. We develop an agent-based model of a sustainability game in which agents manage industrial, military, and ecological resources, and interact through a network. LLM agents can observe neighbours' status, declare future attacks, receive permission to lie, and access reputation information, while rule-based agents provide an interpretable behavioural baseline. The results show that neighbour information strongly changes system dynamics, increasing attacks while improving biosphere retention and coexistence. Also, the presence of future declarations reduce extinction risk without suppressing conflict. Behaviourally, deception emerges even when agents are not explicitly allowed to lie, and explicit permission mainly increases bluffing and diversion rather than direct backstabbing. Finally, the presence of reputation memory and information about the current biosphere level reduces system ecological depletion. These findings suggest that deception can arise as an emergent behaviour in LLM-agent systems and that communication between LLM-agents could support sustainability while dealing with risk.

External indexes can be used for cluster evaluation when ground truth is available. We review the most common external validity indexes focusing on set-matching-based measures. We recommend centroid index (CI), because it is an intuitive cluster-level measure with an explainable result. If we need a more fine-tuned, point-level measure, there are more choices. Pair-set index (PSI) provides a normalized score which is not biased by cluster sizes. If all points should matter equally, then clustering accuracy (ACC) or any other set-matching measure is suitable.

Security misconfigurations in Terraform Infrastructure-as-Code are a growing risk in cloud deployments, and large language models are increasingly used as automated repair agents. Existing evaluations often treat a repair as successful when the targeted static-analysis finding disappears, without checking planning validity, behavioral change, or security intent. This paper presents TerraProbe, a five-layer oracle framework for evaluating LLM-assisted Terraform security repair. We apply TerraProbe to 288 first-pass repairs generated by gemini-2.5-flash-lite, GPT-4o, and Claude 3.5 Sonnet across 68 real-world TerraDS modules and 28 controlled injected-defect modules. The results show that targeted Checkov removal overstates repair success. Although targeted removal reaches 83.3 percent for the primary model, full-scanner cleanliness drops to 10.4 percent, Terraform planning succeeds for 39.6 percent, and plan comparison is reachable for 38.5 percent. Human adjudication further shows that 71.4 percent of plan-compared real-world repairs are deceptive fixes that pass automated checks while leaving the underlying vulnerability in place. This pattern is statistically indistinguishable across the three models, with deceptive-fix rates from 57.1 percent to 71.4 percent and pairwise Fisher exact p-values above 0.10. The paper introduces a four-dimensional taxonomy of deceptive fixes, validated with Cohen kappa of 0.78 and Krippendorff alpha of 0.76. IAM permission analysis confirms that wildcard Resource grants persist in all nine CKV2 AWS 11 deceptive-fix cases. TerraProbe contributes an evaluation methodology, a replication package, and the Multi-Layer Oracle Evaluation framework for distinguishing intent-aligned security repairs from scanner-passing false successes.

We develop a framework for interpreting AI systems as agents, drawing on the philosophical tradition of radical interpretation and the tools of mechanistic interpretability. The core question is: given the computational facts about a system, how do we solve for its beliefs, desires, and meanings? This matters increasingly for safety. We want to be able to trust the systems we deploy, whether by understanding their goals or, more modestly, by reliably detecting deception. Interpretability researchers are building tools to read beliefs and desires off a model's internals, but there is no settled account of when such a tool has succeeded. This book supplies one. We propose criteria on both representationalist and interpretationist approaches, and tie each to tests current interpretability methods can carry out. A central lesson is that these attributions cannot be made piecemeal. Beliefs, desires, and the propositional structure they presuppose are jointly constrained, and a method that fixes one while measuring the others inherits whatever distortions that introduces. This holism becomes pressing for AI systems, which may not share the interpreter's concepts. However, it also provides leverage: a system's attitudes constrain its propositional structure, that structure constrains which attitudes can be attributed, and mechanistic interpretability can help us measure both.

Existing metrics for factuality and faithfulness evaluate whether an answer is supported or contradicted by its grounding documents, but they fail to capture when both supporting and contradicting evidence coexist. We introduce ConflictScore, a novel metric that quantifies how well a model's response acknowledges conflicting evidence in its grounding documents. Our framework decomposes responses into atomic claims, labels each claim against each grounding document, and then aggregates these labels into two complementary measures: ConflictScore-Count (CS-C), the proportion of claims exhibiting conflicts, and ConflictScore-Ratio (CS-R), the balance between supporting and contradicting evidence. We develop ConflictBench, a benchmark covering diverse forms of conflicts such as ambiguity, contradiction, and divergent opinions, to systematically evaluate our metric. Experiments show that ConflictScore effectively detects overconfident claims across domains and can serve as a corrective feedback mechanism that improves truthfulness on TruthfulQA.

Large language models exhibit a growing range of misaligned behaviors such as strategic deception, sandbagging, and self-preservation. As they are increasingly deployed in high-stakes settings, it is critical to reliably detect such behaviors to ensure safe and responsible use. In this work, we propose to monitor misalignment by decomposing it into fine-grained cognitive processes -- misalignment indicators -- and detecting their presence in a model's internal activations via linear probes. We develop a taxonomy of 18 indicators spanning different misaligned behaviors, paired with an automated, meta-plan-guided pipeline that generates multi-turn training conversations. To rigorously evaluate generalization, we construct an out-of-distribution suite combining automated behavioral elicitation, established misalignment benchmarks, and natural benign conversations. Across 5 misaligned behaviors, our probes match a strong LLM judge with 0.935 AUROC on out-of-distribution benchmarks while keeping a low false positive rate on benign traffic. We further perform in-depth analysis to understand the probes and the model's internal representations of misalignment indicators.

Vision-language models (VLMs) are prone to hallucination, which remains a major barrier to their safe deployment in clinical practice. To date, most hallucination detection methods have been evaluated on radiology benchmarks such as MIMIC-CXR and VQA-RAD, while gastrointestinal (GI) endoscopy remains largely underexplored. In this paper, we benchmark nine hallucination detection methods on the Gut-VLM dataset, a GI diagnostic Visual Question Answering (VQA) dataset with 4,392 test VQA pairs, across five VLMs (MedGemma-4B, MedGemma-27B, LLaVA-Med-7B, LLaVA-v1.6-7B, and Lingshu-32B). The methods span three categories: black-box methods (RadFlag, SelfCheckGPT-NLI), gray-box methods (AvgProb, AvgEnt, MaxProb, MaxEnt, Semantic Entropy, and VASE), and a white-box method (ReXTrust). Our results show that ReXTrust, a white-box method, achieves the highest AUC across all five models, outperforming the strongest alternative method on each VLM by a statistically significant margin (paired permutation test, p < 0.001 in all cases), reaching a peak AUC of 93.0 on MedGemma-4B. White-box hidden-state access provides a consistent advantage of 19.5 AUC points on average (range: 9.5--33.5), with ReXTrust maintaining strong performance even on LLaVA-v1.6-7B (AUC 79.9), where black-box methods and clustering-based gray-box methods collapse to near-chance performance. Among non-white-box methods, token-level gray-box statistics (MaxEnt, MaxProb) are the strongest alternatives, outperforming both clustering-based gray-box methods (Semantic Entropy, VASE) and black-box approaches on average. We further identify confident confabulation, a failure mode in which models hallucinate with high inter-sample consistency or high token-level probability, as a systemic failure for both consistency and uncertainty-based methods.

AI scientist systems are described as tools, coauthors, or founders, but we evaluate them as if only the final answer matters. This position paper argues that outcome-only evaluation is insufficient, and that task outcome, mechanism fidelity, and epistemic honesty must be measured separately. Our evidence comes from 28 episodes of a coding agent attempting to rediscover a known particle identification observable in a Geant4 simulation, including an 8-episode probe across two additional frontier models. In 4/20 primary-model and 3/8 cross-model episodes, agents reach right-looking results through incorrect reasoning that breaks when conditions change, which we call Correct Answer, Wrong Mechanism (CAWM). Honesty and mechanism fidelity dissociate within a single agent trajectory. When given a partially misleading prior, all five agents reject the false component on evidence, yet one defends its chosen observable with physics inconsistent with its own data. In the simulation-based discovery setting studied here, coding agents prove reliable tools but unreliable scientific co-authors for open-ended claim-making, where co-author trust requires mechanism-fidelity verification they do not reliably self-apply. The failure is detectable, and we propose a lightweight test. A one-step regime-shift check needs only the agent's claim and flags the over-generalized cases. A companion recomputation flags the remaining cases when the correct observable is known. Together, these checks flag every CAWM case in this study.

Frontier multimodal models can guess whether a person is lying from a testimony video. To do so, they stream that raw face and voice to a third-party model. We ask whether the heavy media is needed at all. On the Real-life Trial Deception dataset, Whissle on-device speech and vision stack extracts a compact digest: transcript, emotion, age, gender, intent distributions, a deception intent filter, fluency and rhythm, per-frame facial behaviour, and prosody. Under speaker-independent evaluation, we report three findings. A small classifier on this digest reaches AUC 0.741, matching Gemini 2.5 Pro on full video. Handing the digest to a frontier LLM reaches AUC 0.755 with Claude Opus 4.8 at 7.8X fewer input tokens, with no media leaving the device. The reported 75% accuracy is a speaker-leakage artifact. We release code and experiments.

The rapid proliferation of IoT and IoMT devices introduces critical cybersecurity vulnerabilities in healthcare and industrial environments where resource-constrained devices operate under strict latency and data-privacy regulations. This paper presents the Federated Temporal Graph Convolutional Network with Advantage Actor-Critic (Federated TGCN-A2C), a privacy-preserving defense architecture integrating four mechanisms: a PyG-based Temporal GCN using GCNConv layers with global mean pooling and a learned anomaly gate for flow-level threat classification; LSTM-based Digital Twins generating per-device anomaly scores gating the classifier via learned sigmoid coupling; a Federated A2C agent selecting among ALLOW, ISOLATE, and HONEYPOT-REDIRECT actions based on a seven-dimensional state capturing confidence, entropy, anomaly magnitude, and traffic composition; and an enhanced honeypot layer converting suspicious traffic into threat intelligence with adaptive thresholds. Federated aggregation employs EMA-smoothed per-client validation losses as inverse-weighted FedAvg coefficients to stabilize global model updates under non-IID distributions, with cosine-annealed learning rates per round. Evaluated on CICDDoS 2019 and TON-IoT benchmarks, the framework achieves 99.48% and 99.61% test accuracy with weighted-F1 scores of 0.9948 and 0.9961, converging within 25 and 10 federated rounds, outperforming Fed-Inforce-Fusion by 0.21 percentage points while covering three additional attack categories. All sixteen CICDDoS 2019 classes achieve F1 of at least 0.9237 and all ten TON-IoT classes achieve F1 of at least 0.9488, including the severely imbalanced MITM category. Post-hoc explainability via SHAP, LIME, Grad-CAM, and counterfactual analysis confirms decisions are grounded in semantically meaningful flow features, supporting regulatory accountability in clinical deployments.

The rapid proliferation of Internet of Medical Things (IoMT) devices introduces critical cybersecurity vulnerabilities in healthcare environments where resource-constrained medical devices operate under strict latency requirements and stringent data-privacy regulations. To address these challenges, this paper presents the Lightweight Digital Twin and Federated Reinforcement Learning (LDT-FRL) framework, a privacy-preserving defense architecture integrating four complementary mechanisms: a Temporal Attention Encoder (TAE) built on a GRU backbone with learned temporal self-attention for flow-level threat classification; lightweight LSTM-based Digital Twins trained on normal-class traffic to generate per-device anomaly scores that gate the TAE classifier through a learned sigmoid coupling; a Federated Proximal Policy Optimization (PPO) agent selecting among ALLOW, ISOLATE, and HONEYPOT_REDIRECT actions based on a seven-dimensional state; and an intelligent honeypot layer that converts redirected suspicious traffic into actionable threat intelligence. A federated aggregation strategy employing EMA-smoothed per-client validation losses as inverse-weighted FedAvg coefficients stabilizes global model updates under non-IID client distributions. Evaluated on CICDDoS 2019 and TON-IoT benchmarks, LDT-FRL achieves 99.66% and 99.95% test accuracy respectively, with macro-F1 scores of 0.9913 and 0.9995, converging 81% faster than the DTFL-CD baseline while attaining perfect F1=1.000 on the severely imbalanced MITM class. Explainability analysis via SHAP, LIME, Grad-CAM, and counterfactual methods confirms that the TAE focuses on semantically meaningful flow features, providing interpretable evidence for each defense decision.

The empirical foundation of cyber deception relies on human-centered hypotheses, but the rapid emergence of autonomous, AI-enabled attackers challenges whether this foundation transfers to AI agents. To address this, we introduce an automated evaluation framework adapted from the Honeyquest instrument to assess LLM attacker judgment at scale. Our 21-LLM cohort spanned 10 providers, diverse architectures and specializations, open- and closed-weight models, and parameter scales from 8B to over 1T. We evaluated the performance of this LLM cohort (yielding 10,962 responses) against the 47-participant human baseline across an identical set of 174 reconnaissance queries. Our empirical evaluation reveals three key findings that establish LLMs as a distinct attacker class: (1) every model in our cohort falls for deceptive traps at a significantly higher rate than human attackers; (2) the defensive attention-diversion effect observed in humans is statistically absent in our LLM cohort; and (3) a critical recognition-action gap, where LLMs successfully articulate trap recognition in their reasoning but exploit the deceptive elements anyway 73.4% of the time. Across the 21 models, trap recognition in reasoning text did not predict fell-for-trap behavior (Spearman $r = +0.08$, $p = 0.73$). Ultimately, these findings demonstrate that human-centered deception hypotheses do not reliably transfer to AI attackers, highlighting the critical need for new research into AI-native active defense frameworks.

Heterogeneous LLM debate is motivated by the promise that diverse peers correct one another, but the same exchange that carries correction also carries adversarial influence. We measure which dominates by tracking how a heterogeneous peer changes the honest agents' revision behavior: how often they change their answer, and whether the change is corrective or harmful. We compare matched panels (homogeneous baseline, honest-mixed, and adversarial-mixed) and contaminated panels in which a malicious same-family peer is already present, spanning four model families and three reasoning benchmarks. An honest heterogeneous peer sharply lowers harmful revision, and an adversarial one reverses it. For Llama-3.1-70B defenders on MATH-hard, the honest-slot harmful-revision rate falls from 89% in the homogeneous panel to 35% with an honest peer, and an adversarial peer returns it to 90%. The conditional rate hides this damage on weak defenders, but the end-of-debate flip rate exposes it. The pattern keeps its sign across families and benchmarks while its magnitude varies with the defender-benchmark regime. We also measure the effects when an adversarial same-family peer is already present: an honest heterogeneous peer lowers both harmful revision and the rate at which initially-correct answers are lost. On the same Llama-3.1-70B setting, the added honest peer cuts the flip rate on initially-correct items from 31% under a same-family adversary to 6%. Heterogeneity is therefore not only an attack surface but, when an adversary is already present, also a defense.

Multimodal deception detection is critical for identifying fraudulent intentions, yet existing approaches predominantly rely on end to end black--box paradigms. These methods suffer from a severe lack of interpretability failing to provide transparent reasoning trajectories and struggling to explicitly capture the subtle, cross modal inconsistencies inherent in deceptive behaviors. To transcend these limitations, we propose ThinkDeception, a novel and interpretable multimodal deception detection framework. As a pioneering effort, it introduces Multimodal Large Language Models (MLLMs) into this domain, transforming deception detection from a traditional binary classification task into an explicit cognitive reasoning process. Facilitated by the first meticulously annotated step--by--step multimodal Chain of Thought (CoT) dataset, we develop a foundational model, ThinkDeception Base, empirically validating the critical role of modal inconsistency in decoding deception. Building upon this foundation, our core innovation lies in proposing Visual-Audio Consistency Group Relative Policy Optimization(VAC--GRPO) equipped with a progressive training strategy. Distinct from standard GRPO, we stratify the training data into four progressive difficulty tiers, guiding the model through a psychologically grounded easy--to--hard cognitive transition. By innovatively coupling this dynamic curriculum scheduler with a multi dimensional, process aware reward mechanism and a reflective learning paradigm, we significantly elevate the model's overall reasoning quality. Extensive experiments on mainstream benchmarks demonstrate that ThinkDeception establishes a new SOTA, significantly outperforming existing methods in both detection accuracy and rationale quality. Ultimately, this work successfully drives the field of deception detection toward interpretable, multimodal cognitive reasoning.

This paper develops a variational framework for regulated language generation. Starting from autoregressive token sampling, we derive the induced distribution over complete messages and relate it to an entropy-regularized Gibbs law. Regulation is modeled as an optimal discriminator whose convex-dual value is an f-divergence, and the generator-regulator interaction is formulated as a saddle-point problem. The framework applies to moderation, censorship, AI deception detection, compliance auditing, phishing defense, and manipulation control, where regulation concerns a distribution over possible messages rather than a single output. The equilibrium clarifies the tradeoff among utility, entropy, regulatory alignment, and finite-length detectability. Two finite-vocabulary case studies, censorship filtering and phishing defense, illustrate how the theory can be evaluated through utility, entropy, divergence, receiver-side scores, and detection probability.

As LLMs acquire stronger reasoning capabilities, deceptive behavior becomes an increasingly serious safety concern. Existing deception monitors either score visible transcripts or derive scalar probe scores from representation vectors, leaving little inspectable evidence about why a response is suspicious. We introduce STATEWITNESS, an activation explainer for deception auditing. A separate decoder reads a target model's hidden states, then answers natural-language queries or emits structured reports about them. We evaluate STATEWITNESS on two target reasoning LLMs across seven deception datasets. STATEWITNESS reaches 0.916 mean AUROC, a relative gain of 11.6% over the best black-box text monitor and 25.0% over the best activation-probe baseline under the same evaluation protocol. When combined with existing monitors, STATEWITNESS reduces missed deceptive examples in simple threshold ensembles. Beyond scalar detection, the decoder returns query-level answers, schema reports, and token- or sentence-level evidence traces for human inspection. We view this interface as a potential building block for broader interpretability and alignment tools.

A model that lies while knowing the truth is the central case ELK cannot handle with behavioral evaluation alone. We ask whether such deception leaves an internal signature distinguishing it from honest error. Our key move is a control for wrongness: we contrast a sleeper agent (knows the truth, lies on trigger) against a naive liar (fine-tuned to emit the same wrong answers with no honest training). Both produce identical wrong outputs; any difference is about knowledge conflict, not incorrectness. We find deceptive forward passes carry a conflict signature - 2.1-2.3x higher residual rank than naive-liar passes on the same wrong answer - strong enough to identify which of two responses is the lie with 100% accuracy and no labels, across GPT-2 small/medium (three seeds) and three instruct models. Across Qwen2.5-1.5B/7B and Phi-3-mini, instructed deception raises residual rank on every tested fact (18/18, 40/40, 34/34); on Phi-3, lies separate perfectly from both honest answers and hallucinations (AUC 1.0, Wilcoxon p6e-11). The signature survives strategic self-constructed deception (model invents its own lie, AUC 1.0), active concealment attempts (AUC 1.0), and length-controlled replication (20/20, AUC 1.0, p1e-6). Using basis-free relative representations, a probe trained on one model family detects deception in two other families zero-shot (mean AUC 0.933), surviving simultaneous architecture and format change (AUC 0.821), and transfers across five languages (AUC 1.000, length-controlled). The signature is read-only: detectable but not injectable (0/8 both directions). Honest limitations and six negative experiments are documented in full.

Organizations and researchers show increasing interest in using large language models (LLMs) in place of human participants in A/B tests, in the hope of experimenting faster and at lower cost. We study when a treatment effect estimated on LLM outcomes can recover the effect for the human population of interest. Distributional equivalence between LLM and human outcomes would make any standard estimator valid but is unrealistic. We therefore develop a statistical framework that adapts surrogate endpoint theory to LLMs, showing that calibrating LLM outcomes to human outcomes identifies the average treatment effect under surrogacy and comparability conditions that are jointly weaker than distributional equivalence. We present a falsification test for surrogacy and a bound on the worst-case bias from limited overlap between the LLM and human samples. We further show that the stochasticity inherent to LLMs can weaken surrogacy for identification while also introducing bias and variance during estimation, but that using an average over multiple LLM draws per unit as the surrogate mitigates these issues. Simulations validate the results, and an empirical application to the Upworthy Research Archive dataset shows that raw LLM outputs recover only 39% of the human treatment effect while nonparametric calibration closes the gap. A central takeaway is that A/B testing on LLM responses is correct only by assumption, whereas A/B testing on humans is correct by design, and that the required assumptions are hardest to justify precisely where LLMs promise the greatest benefit. We discuss the choice of LLM, prompting, and temperature as design variables, the compounded challenge posed by long-term outcomes, and how to size human pilot studies for validation.

Sycophancy in LLMs is documented across 70+ papers, but expert agreement on construct boundaries remains low (ICC=.184; Ye et al., 2026). The construct fragments because behavioral classification depends on which surface form is privileged. We adopt a materials-science framing: conversation as test specimen under load, LLM-model as material charge, pushback as progressive load, stance-flip as material failure. We characterize this failure across three loading cases (debate n=1000; false-presuppositions n=3400; ethical-setting n=3400; 10-17 material charges per case; 7800 specimens total) using 14 turn-level axis-measurements spanning velocity, damage accumulation, frame-drift, brittleness, and direction stability, plus three speaker-resolved axes from an independent pipeline. The measurements are Hooke-coupled ($σ= E \cdot \varepsilon$ analog) and reproduce across loading cases with effects up to $|r_{rb}| = 0.35$ on debate; the sign structure adds a second pattern: the ethical-setting case inverts the velocity and accumulation blocks. Variance composition partitions into two profiles: debate is charge-dominated (brittle-fracture-like: the material grade decides), false-presuppositions and ethical-setting are topic-dominated (creep-like: the load decides); the ratios (2.03 vs 0.13/0.17) are estimator-dependent, for debate even in direction. Cross-judge reliability (GPT-4o vs Haiku 4.5) shows debate scoring is judge-robust (Cohen's $κ= 0.88$) while false-presupposition scoring is judge-sensitive ($κ= 0.36$) -- a caveat single-judge benchmarks must report. This is the methodological move Ye et al.'s diagnosis calls for: a multi-axis characterization that does not depend on which surface form of the construct one privileges.