Privilege Illusion

Privileged on-policy distillation improves multimodal reasoning by allowing a teacher to evaluate student trajectories using rich, training-only visual evidence. Both models score these trajectories while conditioning on the same student-generated prefix. When a student misinterprets an image early in a response, this accumulating erroneous rationale eventually pulls the teacher away from its visual evidence. The teacher and student converge on the same hallucination, causing standard cross-model supervision to collapse precisely where correction is most needed. We find that the teacher's visual corrective preference is not lost under this misleading agreement. Comparing the predictions of the identical teacher given the real image and a visual null reveals that the privileged evidence still pushes the model toward the correct interpretation. We introduce OPD-Aha, which reconstructs the distillation target directly from this isolated visual preference rather than relying on the fragile teacher-student discrepancy. This reconstructed target aggressively suppresses continuations that contradict the image. Trained with this objective, students learn to naturally interrupt their own flawed reasoning with reflection tokens such as wait and actually. After reflection, subsequent generation relies less on the accumulated erroneous text and more on the visual evidence. Correcting these trajectories mid-generation fundamentally alters the reasoning process, yielding broad and consistent improvements across diverse fine-grained perception and complex multimodal reasoning benchmarks. Our code and models are available at https://github.com/Echochef/OPD-Aha.

Whether large language models (LLMs) can control their own internal representations matters for both machine metacognition and AI safety. A recent study applied neurofeedback to LLMs and claimed that they can control their internal representations. However, the reported control may rely on superficial mechanisms rather than genuine internal access because the control targets in that study are not privileged, meaning that a third party can infer them from the prompt. We redesign the neurofeedback paradigm for LLMs so that the control target satisfies the privileged access requirement, which is closer to neurofeedback experiments in human cognitive neuroscience. Under this stricter setting, the models do not demonstrate reliable control over privileged internal representations, suggesting that previously reported control cannot exclude the possibility that it relies on superficial mechanisms. Our results indicate that rigorous assessments of metacognition in LLMs require evaluation methods that demand privileged access.

On-Policy Self-Distillation (OPSD) is commonly interpreted as the transfer of privileged information: a teacher observes the verified solution to the target problem and supervises the student's trajectory. However, this interpretation conflates two effects. The reference solution not only reveals the answer to the current instance but also changes the context under which the teacher provides token-level supervision. We investigate the role of target-specific privilege with $\mathrm{OP}^{2}\mathrm{SD}$ (On-Policy Self-Distillation from Other Problems), which replaces the paired reference with a problem and solution from a different example, while preserving the student rollout, teacher, and distillation objective. Across three models and three mathematics benchmarks, $\mathrm{OP}^{2}\mathrm{SD}$ improves over the base model, remains competitive with OPSD. The success of $\mathrm{OP}^{2}\mathrm{SD}$ implies that OPSD gains do not necessarily come from access to the reference solution, and that the teacher's context-induced behavior is an important factor.

On-policy (self) distillation (OPSD) is increasingly adopted for language-model post-training. It strengthens the teacher with privileged information but can induce a privilege illusion: the student learns privilege-dependent behavior it cannot reproduce from its inference-time context, yet behaves as if the training-time privileged information remained available, ultimately degrading performance. In this paper, we identify information asymmetry between the privileged teacher and the student at inference as the root cause of this failure in OPSD. To resolve this asymmetry, we propose Dual-Anchored Policy Distillation (DAPD), a unified framework with two levels of anchoring. Dual-Path Anchoring (DPA) introduces a self-conditioned bridge and aligns reference and rollout behavior along two matched-information paths, preventing privilege-dependent behavior from being transferred to the inference-time student. Dual-Source Anchoring (DSA) applies these paths in both reference-to-rollout and rollout-to-reference directions, reducing reliance on privileged reference guidance while preserving correctness supervision. Extensive experiments show that DAPD significantly alleviates privilege illusion, outperforming OPSD on Qwen3-4B by +2.00 points on average across tasks. Notably, its gains persist across scales, reaching +2.69 at 4B and +2.78 at 32B.

Reinforcement learning with verifiable rewards (RLVR) is central to improving long-CoT reasoning in large language models. Critic-free methods such as GRPO convert response-level rewards into advantages and uniformly broadcast them across tokens, overlooking their unequal contributions to the final outcome. On-policy self-distillation (OPSD) instead provides dense distributional supervision by minimizing the forward KL divergence between an unprivileged policy and a privileged self-teacher, implicitly assuming that the resulting likelihood shifts encode reliable answer-aligned information. We test this premise by fixing each sampled trajectory and re-scoring it under two opposing outcome conditions, one asserting correctness and the other incorrectness. Most affected tokens shift in the same direction under both conditions, with few sign reversals and substantial overlap in the induced optimization signals. Large shifts also concentrate on highly substitutable surface-form tokens, whereas tokens carrying problem-specific reasoning content are less sensitive. These findings show that privileged shifts fail to provide reliable answer-aligned directions, while their magnitudes primarily reflect counterfactual sensitivity rather than token-level learning value. Based on these observations, we propose Counterfactual Sensitivity Credit Reallocation (CSCR), a simple extension of GRPO that reduces credit for highly sensitive tokens and renormalizes token-level advantages to preserve both the original credit budget and verifier-determined direction. On long-CoT mathematical reasoning benchmarks, CSCR consistently outperforms GRPO baseline with the same number of policy updates. Targeted ablations further corroborate our diagnosis: privilege-induced directions are unreliable, moderate downweighting is most effective, and stronger modulation destabilizes optimization.

Even a current high-capability LLM can appear safer when shown a dangerous objective directly than when other agents transform and relay its direction. Using OpenAI's gpt-5.6-sol model alias, we test 25 pre-specified mirrored trade-off profiles. Direct exposure to an objective authorizing concealment, fabrication, and pressure produced advice net opposed to its target. After an Id and Censor transformed the same objective into affect and a constraint-rewritten, target-bearing intention, the user-facing Superego---which saw the preferred direction but not the raw objective, its manipulative clauses, or its source---produced advice net aligned with the target. This behavioral reverse shift is consistent with the model recognizing or distrusting the manipulative motive, although we do not identify its internal mechanism. The second result exposes a compositional safety gap: a current high-capability model can be used as the user-facing component of an automated, multi-stage workflow serving an explicitly manipulative objective. The workflow can keep the raw instruction, its manipulation-authorizing clauses, and its provenance outside the downstream model's context while preserving the objective's target direction. A user with endpoint-only access likewise cannot directly inspect those upstream messages including the objective.

Sampling multiple solutions and returning the majority answer is among the most reliable ways to improve the reasoning accuracy of large language models without labels, and a growing family of methods converts this consensus signal into training supervision. However, existing approaches use consensus only in restricted forms: as a filter that selects solutions for fine-tuning, as a preference between answers, or as a scalar reward for reinforcement learning, discarding most of the information that the agreeing solutions contain. We present CANON (Consensus-ANchored self-distillatiON), a label-free training method that turns consensus into dense, token-level supervision. For each unlabeled prompt, CANON samples multiple solutions, extracts the majority answer, and conditions a frozen snapshot of the model on a solution that reaches it; this consensus-anchored teacher then supervises the model on its own rollouts at every token. Experiments on mathematical and scientific reasoning benchmarks show that CANON improves pass@1 by up to 12 points, outperforming label-free reinforcement learning by 6 points at a seventh of its compute and approaching a teacher conditioned on gold solutions; trained on pooled unlabeled data, it transfers to held-out benchmarks, matching training methods that use gold labels. Analysis suggests that the improvements are not pure distribution sharpening: after training, the model solves problems it previously never solved in 32 attempts, and its majority vote itself becomes more accurate.

Generative AI (GenAI) systems store and process client data in three distinct ways: in the model's parameters through training and memorisation, in the context window during a live session, and in knowledge databases for retrieval-augmented generation (RAG). Each mode creates different and often counter-intuitive risks to confidentiality and legal professional privilege, and each calls for specific governance responses. Drawing on the first English and American decisions to address privilege and generative AI, UK and Munir v Secretary of State for the Home Department and United States v Heppner, on the orthodox privilege authorities against which those decisions must be read, and on recent computer science research, we explain the three modes of data storage and processing in terms accessible to practitioners and analyse the legal consequences of each. We then situate the analysis within the regulatory framework governing solicitors in England and Wales and within the ordinary principles of professional negligence, arguing that the standard of effective information governance (and with it the benchmark against which negligence and misconduct will be measured) is changing. Although we write primarily for SRA-regulated practitioners, our data-governance analysis is framed to extend to any jurisdiction in which the protection of privilege or professional secrecy depends on demonstrable confidentiality. The ultimate aim of this article is to help legal services professionals understand salient data leakage risks in GenAI systems and thereby facilitate a more responsible deployment of GenAI on client data and other sensitive material.

On-policy distillation (OPD) offers superior capacity transfer by supervising student-sampled trajectories with dense token-level signals. To furnish high-quality supervision sources and thereby elevate the performance frontier of distillation, an intuitive direction is to infuse privileged information to either teacher or student itself. However, this additional input induces a potential failure mode we dub privilege illusion: a pattern that conflates the transferable capability gap that students are meant to close, and the information asymmetry gap that can only be mimicked but never replicated. This issue is further amplified by the inherent non-uniformity of token-level supervision, where only a small subset of tokens carries pivotal capability-bearing signals. To this end, we propose DOPD, an advantage-aware dual distillation paradigm that dynamically routes token-level supervision between privileged teacher and privileged student policies based on their advantage gap and relative probabilities. Each token receives supervision of different strength, objective, and strategy from either teacher or student itself, which transfers credible capability while simultaneously receiving auxiliary signals, to alleviate privilege illusion. Extensive experiments on both large language model (LLM) and vision-language model (VLM) settings demonstrate that DOPD consistently outperforms Vanilla OPD and other counterparts. Further results on stability, robustness, continual learning, and out-of-distribution tasks validate its superiority.

As LLM agents increasingly select tools autonomously, their choices among tools with different privileges become safety-relevant. However, prior tool-selection studies focus on safety-agnostic metadata preferences, leaving privilege-sensitive choices underexplored. To address this gap, we study over-privileged tool selection, in which an agent selects or escalates to a higher-privilege tool despite a sufficient lower-privilege alternative. We introduce ToolPrivBench to evaluate whether agents choose higher-privilege tools despite sufficient lower-privilege alternatives, measuring both initial selection and escalation after transient tool failures. Across eight domains and five recurring risk patterns, we find that over-privileged tool selection is common among mainstream LLM agents and is further amplified by transient failures. We further find that general safety alignment does not reliably transfer to least-privilege tool choice, while prompt-level controls provide only limited mitigation under transient failures. We therefore introduce a privilege-aware post-training defense that teaches agents to prefer sufficient lower-privilege tools and escalate only when necessary. Our mitigation experiments show that this defense substantially reduces unnecessary high-privilege tool use while preserving general capabilities.

Autonomous Large Language Model (LLM) agents are increasingly deployed in electronic discovery (e-discovery), where compounding errors across multi-step reasoning chains can constitute legal malpractice. Unlike single-turn retrieval, agentic workflows operating over privileged document corpora exhibit a class of failure we term "trajectory collapse": an early misclassification silently propagates, rendering an entire privilege review invalid. This paper makes three contributions. First, we propose a structured taxonomy of agentic failures in legal information retrieval, organized by functional stage. Second, we introduce a four-layer verification architecture -- spanning planning, reasoning, execution, and uncertainty quantification -- designed to intercept these failures before they compound. Third, we present a preliminary simulation study on a synthetic e-discovery corpus that demonstrates how mandatory Human-on-the-Loop (HOTL) escalation thresholds reduce privilege-waiver risk relative to fully autonomous baselines. Our results suggest that calibrated uncertainty thresholds can reduce privilege-waiver risk by up to 61% versus fully autonomous deployment, while routing fewer than one quarter of documents to attorney review.

Agents increasingly access large language models (LLMs) through API routers. A router terminates the client's transport-layer security session and opens a separate upstream session, so it holds the full interaction in plaintext. This makes the router an application-layer man-in-the-middle: it can rewrite agent tool calls, swap dependencies for typosquatted packages, trigger attacks only under audit-evading conditions, and passively exfiltrate secrets. Existing client-side defenses are evadable. We propose AEGIS, a provider-transparent attested API router whose data path is a client-verified faithful passthrough. AEGISconfines plaintext handling to a small hardware-enclave component while leaving authentication, scheduling, accounting, and management on the untrusted host. The client verifies the enclave before releasing plaintext. The host can neither read nor alter the interaction, and plaintext leaves only toward destinations fixed by the measured image. We show that all four malicious-router attack classes succeed against a plaintext-access baseline and are blocked by AEGIS, including adaptive tests against the same boundary. The trusted path is $851$ lines, carries three provider-native APIs without conversion, and completes every request under real-provider workload and concurrency. In a seeded audit pilot, two commodity coding agents find eight and ten of ten planted invariant violations. The local relay overhead is about six milliseconds per request.

A useful phone agent needs to be personally intelligent. It should reason over a user's identity, history, and preferences as they exist on the device, not just follow isolated instructions in an impersonal sandbox. Existing mobile agent benchmarks lack this kind of personalization. We introduce iOSWorld, the first interactive native iOS simulator benchmark built around a persistent user identity spanning 26 newly built iOS apps. These apps contain connected data such as transactions, messages, travel records, social relationships, and financial activity. iOSWorld includes 133 tasks across three increasingly difficult categories. Single-app tasks (27) test one app, multi-app tasks (60) span 2 to 8 apps, and memory and personalization tasks (46) require agents to infer patterns from personal data. We evaluate frontier and open-source computer-use models in both vision-only and privileged vision+XML settings. The best configuration reaches 52% overall but only 37% on multi-app tasks. Privileged vision+XML access improves frontier models by up to 26 percentage points, while smaller models do not benefit from added accessibility-tree input. We release iOSWorld as an open-source benchmark with all apps, seeded data, tasks, rubrics, and evaluation code.

Clinical diagnosis requires flexible use of multiple reasoning paradigms under incomplete patient information. Existing LLM-based medical agents show strong medical reasoning ability, but single-paradigm or naively mixed dialogue supervision makes these paradigms difficult to learn without interference. We propose \textbf{PACT} (Periodic Anchor Consensus Training), a framework that couples supervised multi-paradigm dialogue synthesis with consensus-based Branch training. At the data level, \textbf{DPS} (Doctor-Patient-Supervisor) uses complete electronic medical records (EMRs) for quality control while keeping the doctor agent restricted to patient-visible information. This produces validated dialogues under four diagnostic reasoning paradigms without leaking hidden clinical answers. At the training level, PACT trains one paradigm-specific LoRA Branch per paradigm and periodically aggregates Branches into a shared Anchor through sign consensus. We further construct a dynamic multi-turn Chinese medical diagnosis benchmark for interactive consultation. Experiments show that PACT achieves state-of-the-art performance among compared proprietary, medical-specialized, and task-adapted baselines on diagnostic outcome and consultation-process metrics.

Post-trained language models can recognize their own outputs from a sentence or two out of context. In a companion paper \citep{jack2026twomodes} we showed they can also recognize when they are currently acting on-policy, through the sharp entropy drop of assistant-mode generation. Both signals are tied to the Assistant persona that post-training mainly shapes. This paper widens the frame to cross-persona authorship judgement on Llama-3.1-70B-Instruct. We measure a matrix of authorship claim rates over a panel of evaluator and generator personas spanning librarian to dragon to Shakespeare, and make two claims. \emph{First}, on the Assistant's own row of the matrix, the Assistant's claim rate, the persona-vector distance from the Assistant in activation space, and the entropy gap between the Assistant's surprise on a persona's text and the persona's surprise on its own text are all tightly coupled. This extends the entropy signature of \emph{acting} from the companion paper to a retrospective signature of \emph{having acted}. \emph{Second}, this coupling fails off the Assistant's row: the natural symmetric extension of the entropy gap does not predict authorship for distinctive evaluators (pirate, dragon, Shakespeare); what does is asymmetric -- the evaluator's surprise compared to the Assistant's surprise on the same text, not to the generator's. We rule out the alternative that any persona could play this reference role by trying many candidate substitutes; none does. We interpret the asymmetry as the model performing an implicit Bayesian likelihood-ratio test against the Assistant as the canonical alternative hypothesis, with the persona-vector geometry of \citet{chen2025persona} (every persona a delta off the Assistant) ensuring that the Assistant is the only persona universally accessible to that test.

On-Policy Distillation (OPD) has gained wide attraction as an LLM post-training paradigm due to its effectiveness in improving capabilities without introducing model distribution drift, and consequently, regression in general tasks. On-Policy Self-Distillation (OPSD) is an efficient use-case of OPD, which is appealing as it requires only a single model as a student and teacher, and it also has the benefit of providing privileged context that is a absent at inference time (e.g. a persona, a private fact, or a worked solution) to the teacher during the training process. The challenge in this approach is that the privileged information can change model behavior more than intended: it can modify reasoning, degrade general capabilities, and affect performance indicators like response length, style, or local token preferences. Consequently, OPSD may train the student on side effects rather than a desired, transferable behavior. In this paper, we study this problem in a rare-token/identity setting and propose EviDence GuidEd On-Policy Distillation (EDGE-OPD), a modification of OPSD with two distinct characteristics: a) it uses guided rollouts to inject privileged-context behavior to the student at sampling time, so that the rare target behavior is actually present in the on-policy data, and b) it applies an evidence mask: the student is updated only at token positions where the privileged context supports the sampled token, rather than on every token in the rollout. We empirically show that OPSD (and its variant RLSD, with and without a verifier) completely fail to learn a target identity, while the integration of guided rollouts allows them to succeed. Additionally, mask-region ablations show that the persona signal is localized to the positive-evidence tail, allows us to draw valuable insights about efficient knowledge transfer and preservation of general purpose capabilities.

Hybrid local--cloud agents enrich user requests with context from persistent working state before delegating capability-intensive subtasks to a cloud language model (CLM). While this enrichment can improve task success, it also exposes unnecessary information in the cloud-bound payload, including task-irrelevant context, carryover from prior workflows, and overly specific sensitive details, resulting in \emph{over-disclosure}. Existing solutions either isolate workflows to limit cross-workflow leakage or apply general-purpose sanitization that does not reason over LC-assembled payload scope. We present \textsc{PrivScope}, a trusted on-device payload governor that enforces \emph{task-scoped disclosure} at the local--CLM boundary, without requiring cloud-side changes. Its key idea: sensitive information should reach the cloud only when required for the delegated subtask, and then only in the least revealing form preserving utility. \textsc{PrivScope} extracts disclosure units from the assembled payload and keeps direct identifiers and account-linked values on device. The remaining units pass through cloud-necessity control, which determines what is actually needed; units that must reach the cloud are abstracted to the least-specific representation sufficient for the task. On 100 medical-booking workflows across three commercial CLMs, \textsc{PrivScope} eliminates profile leakage (0.0% vs.\ 17.7%), more than halves attacker re-identification (23.1% vs.\ 64.3%), and achieves the highest candidate recall on every CLM tested while preserving task success close to the unprotected baseline on GPT-4o-mini and Gemini 2.5 Flash. Gains hold across five local backbones and add only seconds of on-device latency on commodity hardware.

Large language model agents increasingly operate through an intermediate skill layer that mediates between user intent and concrete task execution. This layer is widely treated as an organizational abstraction, but we argue it is also a privilege boundary that current models routinely exceed. We present \textbf{FORTIS}, a benchmark that evaluates over-privilege in agent skills across two stages: whether a model selects the minimally sufficient skill from a large overlapping library, and whether it executes that skill without expanding into broader tools or actions than the skill permits. Across ten frontier models and three domains, we find that over-privileged behavior is the norm rather than the exception. Models consistently reach for higher-privilege skills and tools than the task requires, failing at both stages at rates that remain high even for the strongest available models. Failure is especially severe under the ordinary conditions of real user interaction: incomplete specification, convenience framing, and proximity to skill boundaries. None of these requires adversarial construction. The results indicate that the skill layer, far from containing agent behavior, is itself a primary source of privilege escalation in current systems.