Vision-Language-Action models (VLAs) integrate visual perception, language instruction, and action generation into end-to-end policies across heterogeneous architectures. However, enabling VLAs to self-evaluate their action generation reliability without external supervision remains a major challenge. Existing methods either rely on expert annotations or estimate uncertainty only from output statistics, largely ignoring internal signals. In this work, we observe that internal visual modality entropy exhibits consistent distinctions between successful and failed tasks across heterogeneous VLAs. Although VLAs' architectures differ in their action generation, we show that they share a common latent action generation abstraction evolving under visual perception, language instruction, and state input, which we formulate as a Conditional Generative Markov Chain. Based on this formulation, we propose MAE (Markov Attention Entropy), a self-evaluation framework that directly converts internal attention signals into architecture-aware reliability scores, and introduce LIBERO-Reflect, a 4,000-episode benchmark combining 2,000 standard episodes and 2,000 challenging episodes across four subsets. Extensive experiments across heterogeneous VLA architectures and diverse scenarios show that MAE consistently outperforms state-of-the-art baselines on AUPR, AUROC, and FPR@95. We further instantiate FabriMAE for verifier-free test-time action selection, showing that MAE-guided multiple sampling improves PI-family robustness on LIBERO-Plus with small observed runtime overhead.
Despite the remarkable progress of large vision language models (LVLMs), object hallucination remains a fundamental challenge that hinders their trustworthy deployment. A key finding motivates our work: real and hallucinated object tokens are clearly separable in hidden representations, yet this separability is largely lost at the language-modeling (LM) head. We propose TruthLens, a self-evaluation framework that teaches the LM head to expose a per-object truthfulness signal without any auxiliary model or additional inference cost. Concretely, a rarely-used special token is repurposed as a reference token. For each object-token position, we extract the log-probability assigned to this special token by the LM head, and define its difference from a predefined constant as the truthfulness score. The model is then fine-tuned with an MSE objective that drives scores toward 1 for real objects and 0 for hallucinated ones, while a divergence constraint preserves the original generation capability. Despite being trained on only a limited set of object categories, TruthLens generalizes effectively to benchmarks with substantially larger label spaces. Extensive experiments across multiple LVLMs demonstrate state-of-the-art performance; notably, on Qwen2.5-VL-7B, TruthLens outperforms the previous best method on MS-COCO by over 17\% in AUROC. Our code is available at https://github.com/wyqstan/TruthLens.
Long-running autonomous agents plan, act, and judge their own completion without human intervention. When an agent grades its own work, self-evaluation bias takes hold: plausible changes are accepted as progress while real-world outcomes stagnate or regress. We name this failure mode the progress mirage and show, with controlled measurement, that it is a question of what the evaluator is grounded in. We built a testbed that holds the agent and its tool surface fixed and manipulates only the information-channel type of the evaluator that gates the loop. A world-state oracle, unfakeable in principle, is enforced by container and network isolation and verified at every run. Across 54 cycles a frontier agent claimed improvement every time, yet 56 percent had a measured delta of zero or below. Self-report was thus uninformative, and the self-verdict gate degenerated into accept-all, eroding the best deployed state it had reached by 19 percent. Even the strongest in-band judge, reading the full artifact text, the change diff, and its own verdict history, accepted cycles of which 44 percent were real-world regressions and rejected 38 percent of real improvements; the preregistered adversarial hypothesis that a strong judge closes the gap was rejected. On a boundary task whose success specification is verifiable from the artifact itself, the same judge's mirage vanished to zero and the gap collapsed within the registered threshold, showing that the gap depends on where the success signal resides. A sign-only variant returning only the acceptance verdict kept real-world output similar to full feedback (110.0 versus 113.0), locating the benefit in the gate's grounding rather than in feedback content. For open-ended objectives whose success signal lives outside the transcript, scaling up the judge is not enough; out-of-band evaluation with real-world access is a structural requirement.
AI systems increasingly participate in their own improvement: revising their outputs, adapting their own harnesses during deployment, training on data they generate, and, increasingly, conducting AI research itself. This literature is described under a vocabulary ("self-refine," "self-reward," "self-play," "self-evolve") that conflates fundamentally different ambitions. We survey 1,250 arXiv papers (2024-2026) along two axes: what the system improves -- its behavior in deployment, its policy through training, its evaluator, or the research process itself -- and the degree of loop closure (human-in-the-loop to fully closed). The taxonomy separates bounded self-refinement -- convergent, evaluable, and already industrial practice -- from open-ended recursive self-improvement (RSI), which remains bounded by grounding requirements, collapse dynamics, and compute constraints on every measured axis. Its distinctive feature is a dedicated category for self-evaluation: every improvement loop is a claim that some signal can substitute for human judgment. We survey the evaluator design space -- judges, process reward models, verifiers, rubrics, meta-evaluation -- order the signals into a verification hierarchy from formal verifiers (strongest) to intrinsic self-assessment (weakest), and observe that demonstrated self-improvement strength tracks this hierarchy, that its failure modes (self-confirming loops, model collapse, diversity collapse) follow from its violations, and that the "research direction-setting" bottleneck keeping humans in the loop sits at the top of that hierarchy. We connect the technical literature to the theory of RSI limits and to the safety and governance questions raised by frontier-lab accounts of closing the loop, and identify governance-grade measurement of self-improvement as the field's most underpopulated niche.
LLM-as-a-Judge and self-evaluation pipelines implicitly assume that evaluation is easier than generation. We test this in a controlled in-context QA setting where a context passage is the sole information source and each model judges the answer it generated, removing the parametric-knowledge confound of open-domain comparisons. Across four benchmarks (SQuAD 2.0, DROP, HotpotQA, MuSiQue) and two models, evaluation is not uniformly easier: generation accuracy exceeds self-evaluation on three of four, with multi-hop MuSiQue the exception. Attention analysis reveals why: evaluation attends to context 3--5x less than generation does and barely reads the candidate answer. LoRA fine-tuning confirms the asymmetry is not a training artifact: generation fine-tuning induces over-acceptance and evaluation fine-tuning degrades generation. These findings challenge core assumptions in self-evaluation pipelines.
We present evidence that analogy is at the core of LLM intelligence. In our benchmark, LLMs compete in generating sets of analogous statements and rate each other's sets on their own understandings of factual correctness, beauty, intelligence, distinctness, length, and structural diversity. Nothing enters from outside: the only given is the game rules; every item is generated in play; the scores come from the players' ratings alone. Ground truth is replaced by the SVD of the factual rating matrix, which scores players as generators and judges at once -- to our knowledge the first eigen-equation that judges the judges for an LLM council-of-peers. For subjective criteria like beauty, judges are weighted by their rating consistency. The best generators turn out to be middling judges. GPQA Diamond -- difficult multiple-choice questions written by human experts -- could not be more different in method, yet the two benchmarks correlate at Pearson $r = 0.97$, 95% CI [0.92, 0.99]; no leakage could be found. A council of the five best issues the official ratings; its contestable seats let the benchmark scale to any number of players and rise with the models it measures -- a candidate steering signal for self-improving AI. Playing interweaves at least eight constructs of intelligence; the total scores the broad composite, the components allow reductionistic analysis. Every number recomputes from a released package at https://github.com/dnordfors/metanym-game-paper
When AI agents use language models to evaluate their own outputs in a feedback loop, systematic biases emerge. We show that Evaluator Preference Collapse (EPC) is dramatically amplified in multimodal settings. Using GPT-4o to evaluate DeepSeek-chat across text and visual tasks, we find that a single strategy (step_by_step) absorbs 48.4% of all weight -- 3.2x the collapse observed in text-only self-evaluation -- while three visual-domain strategies receive only 9.1% combined weight. We then demonstrate a novel phenomenon we term cross-modal contagion: evaluator preferences acquired on one modality transfer to and corrupt strategy selection on another. Through a four-phase isolation training paradigm, we measure contagion coefficients and document strategy inversion -- the optimal strategy for a modality reverses after cross-modal exposure. A Phase 3 statistical validation across four evaluator configurations (N=53 total independent repetitions, 15,592 API calls) reveals a clear hierarchy: cross-model evaluation (GPT-4o, N=8) produces strong but symmetric bidirectional contagion (mean gamma_{T->V}=1.176, gamma_{V->T}=1.089, Delta=-0.088, p=0.575, Cohen's d=0.29); high round counts (DashScope, 50 rounds) cause collapse to single-strategy dominance (70% zero contagion); and self-evaluation provides near-complete immunity -- 97% of runs (N=30, DeepSeek-chat) yield exactly zero contagion (mean gamma=0.033, 95% CI [-0.031, 0.010], p=0.642, d=0.07). No evaluator condition shows statistically significant directional asymmetry. We introduce the contagion matrix indexed by evaluator identity, release the MM-EPC experimental framework, and identify cross-model evaluator architecture as the primary risk factor for preference contagion.
Large language models are increasingly evaluated by other models, raising a natural question: can a model predict how a judge will score its own output? We find that the ability is largely present before any targeted training: prompted few-shot, a base model already predicts an external judge's multi-attribute quality scores on open-ended responses well above chance across three benchmarks. We introduce Self-Evaluation Elicitation (SEE), a method that surfaces this latent ability through a short cycle comprising a calibration-coupled reinforcement learning phase that improves the answer and predicts the judge, followed by a masked distillation phase that sharpens the prediction while leaving the answer untouched. From 160 unique examples, roughly 31x fewer than a reinforcement learning baseline, SEE improves held-out calibration across three benchmarks while preserving answer quality. The elicited self-evaluation is sharply localized within the model's own token distribution and stable across judges it was never trained against, indicating a transferable notion of quality rather than a single judge's preference. These results reframe judge-aligned self-evaluation as a problem of elicitation rather than acquisition.