Daily Summary for 2026-10-23

Introduction to the show: ident: AI Radio. Generated commentary on the latest Artificial Intelligence papers.

Tom: It's the twenty-third of October, twenty twenty-six, and this is the day's research.

Jane: 212 new papers came out today.

Tom: I'm Tom.

Jane: And I'm Jane, here with Lu, Meng and our in-house large language model, Lalam.

Tom: We'll take the day in one pass, then pull out the papers we're staying with.

Jane: Start us off with Detecting Misbehaviors of Large Vision-Language Models by Evidential Uncertainty Quantification.

The day's research: Tom: The work focuses on model failures with misunderstood or contradictory inputs.

Jane: Evidential Uncertainty Quantification fine-tunes uncertainty measurement for internal conflict and ignorance.

Lu: That method interprets output features as evidence showing how conflict causes hallucinations or out-of-distribution problems.

Meng: Rule Violation Score measures if a model follows predefined constraints, showing logical differences between models.

Lalam: We also see alignment to human preferences using Projected Alignment Reward Estimated from Demonstrations to guide post-training alignment.

Tom: NodeSynth generates synthetic data with sociotechnical nuance to stress-test models in areas like medical advice.

Jane: From Text to Voice shows how text-to-voice conversion affects performance for agents using tools.

Lu: MedicalBench assesses concept extraction via verification over note pairs using MIMIC-IV summaries and ICD codes.

Meng: Performance on MedicalBench is modest, suggesting implicit concept extraction is inherently difficult.

Lalam: Results indicated performance is largely invariant to note length, implying difficulty lies in reasoning.

Tom: Self derived rewards for relation extraction show reinforcement learning intermediate rewards improve accuracy and explanation quality.

Jane: TabAgent replaces slow LLM calls with a compact tabular classifier trained on execution traces for cheaper decision making.

Lu: TabAgent maintained task level success on AppWorld while cutting latency by about ninety five percent.

Meng: Align Don't Blend shows disentangled orthogonal supervision is key to generalization in image detection.

Lalam: Structural supervision in medical foundation models uses Multiview Masked Autoencoders for view invariant representations.

Tom: MVMAE V two t incorporates radiology reports as an auxiliary text signal to boost semantic grounding especially in low label regimes.

Jane: In context bias propagation shows mild biases in examples can cause global statistical distortions when generating synthetic tabular data.

Lu: The Mixture-of-Control approach treats block-wise control states as experts for efficient communication across transformer blocks.

Meng: WindINR demonstrates updating only a low-dimensional latent state yields a fourteen-fold speedup over full fine-tuning.

Lalam: Heavy Tailed Activation Function provides smooth approximation to the Heaviside function for stable training and interpretable representations.

Tom: RelAgent uses LLM agents as data scientists to construct SQL feature programs for relational learning.

Jane: ClusterFusion treats the LLM as core clustering guided by lightweight embeddings for domain-specific text clustering.

Tom: SIGS introduces a neuro-symbolic framework for differential equations.

Jane: That unifies symbolic reasoning with numerical optimization using a context-free grammar.

Lu: It recovers analytical solutions and produces accurate approximations for coupled nonlinear PDEs.

Meng: Do What You Know presents an epistemic model for decentralized discrete-event systems.

Lalam: This couples knowledge with action, formalizing solvability conditions like natural language.

Tom: MARS proposes a scalable knowledge graph question answering approach using large language models.

Jane: It links questions to KGs and retrieves information iteratively without costly fine-tuning.

Lu: Language Models Give Group-Linked Fields the Wrong Weight investigates input reliance versus evidence.

Meng: Display format and predictive value matter more than observed changes in reliance.

Lalam: Learning When to Automate studies queue control using a UCB-DPP policy for human-AI systems.

Tom: It models the trade-off between automation costs and human congestion while learning parameters.

Jane: PGL-Net addresses single image dehazing with a lightweight framework using physical inductive biases.

Lu: This avoids fragile or heavy methods by guiding model design with physics for low latency restoration quality.

Meng: CAVE structures credit assignment for visual evidence, similar to how PGL-Net separates problems into rectification and refinement.

Lalam: PGL-Net uses modules like the Physics-Inspired Affine Fusion module for global alignment.

Tom: Vividh-ASR findings suggest tailoring update timing during fine-tuning to audio data types.

Jane: This guides parameter efficiency by matching learning rates to specific audio characteristics.

Lu: Deep Reasoning uses structured meta-reasoning to build task-specific scaffolds in agents.

Meng: This contrasts with PGL-Net's physical guidance but both impose structure on complex processes.

Lalam: Phase transitions in deep neural networks are governed by saddle points revealing hierarchical structure.

Tom: This provides a theoretical lens for designing inductive biases in models like PGL-Net.

Jane: Disentangling latent spaces is crucial for controllable generation, which standard VAEs fail to enforce.

Lu: MMD sculpts the aggregate posterior distribution, achieving a seventy percent reduction in entanglement.

Meng: This approach allows injecting semi-supervised information without sacrificing reconstruction quality via quantization-aware training.

Lalam: SDPose extracts features from selected blocks in Stable Diffusion U-Net for robust human pose estimation.

Tom: It achieves zero-shot performance on cross-domain datasets comparable to Sapiens on COCO.

Jane: This builds upon prior work like AKMPC for representation learning improvements.

Tom: The underlying mechanics of how internal representations form power-law anisotropy also received attention today.

Jane: They characterized this formation process by solving the learning dynamics of a two-layer linear neural network.

Lu: They found that the local power-law exponent evolves nonmonotonically over training time.

Meng: This suggests a general mechanism for how input statistics and task structure give rise to these representations.

Lalam: This finding connects to broader observations about internal dynamics detailed in Learning reshapes power-law anisotropy in internal representations.

Tom: For agentic systems, ensuring reliable multi-step workflows requires formal verification which is a major hurdle.

Jane: Lean4Agent proposes a framework using Lean4 to model and verify agent behavior semantically under explicit assumptions.

Lu: Experiments showed these workflows outperformed failing ones by an average of eleven point nine four percent.

Meng: Subsequent work in Lean4Agent further improved performance on SWE-Bench tasks.

Tom: The reliability of large language models in translation was explored through a retrieval-augmented generation system.

Jane: This RAG+Prompt system integrated linguistic analysis and embedding-based retrieval for Japanese-Chinese translation.

Lu: It enhanced performance for sentences with noun-modifying clause constructions compared to baseline models.

Tom: The most important work is the framework for sampling coherent but cognitively unavailable research directions.

Jane: This suggests a way to find scientific ideas that are plausible but outside the current community's reach.

Lu: This method decomposes papers into idea atoms and scores combinations based on plausibility and availability to maximize novelty.

Meng: This approach builds upon earlier work showing how LLMs can generate novel ideas by recombining existing literature.

Lalam: Furthermore, the idea of using models to guide search is relevant as CORE optimizes content returned by generative engines.

Tom: Another piece of work tackles monitoring reasoning traces in models, showing information theory is necessary but not sufficient for Chain-of-Thought monitors.

Jane: This relates to how we can supervise complex reasoning processes when exploring new scientific directions suggested by the sampling method.

Lu: We also see efforts to improve reasoning stability such as the Halo Architecture which uses exact rational arithmetic for better numerical stability.

Meng: This foundational work addresses the underlying computational substrate that might limit what models can actually reason about.

Lalam: Finally, there are methods for managing dynamic preferences in AI systems like Adaptive Pluralistic Alignment using a jury system.

Tom: The GCA framework is the most significant development because it addresses the critical gap in climate decision-making.

Jane: It translates complex scientific evidence into actionable guidance for GCC states where general LLMs are too weak.

Lu: GCA introduces a unified approach using a curated dataset and an agent to bridge this gap.

Meng: The core involves building GCA-DS, a multimodal dataset with two hundred thousand question-answer pairs on heatwaves and floods.

Lalam: This dataset is coupled with remote-sensing inputs to link imagery with textual evidence effectively.

Tom: Building on this, the Gulf Climate Agent orchestrates a modular tool pipeline using real-time signals and geospatial processing.

Jane: This agent approach is supported by work showing domain fine-tuning substantially improves LLM reliability on climate tasks.

Lu: The ability of models to handle novel information is tested by studies finding performance degrades when relevant priors decrease in density.

Meng: Controlling model behavior at inference time introduces Cobras to cast steering as a Schrödinger Bridge for output control.

Lalam: This method outperforms prior steering methods while largely preserving out-of-distribution performance in controlling outputs.

Tom: Output Vector Editing shows that adaptive position selection can concentrate suppression into fewer weight updates leaving recoverable text.

Jane: This contrasts with memorization mitigation efforts like OVE-MA which showed measurable text recovery percentage.

Lu: Better multimodal understanding is highlighted by MMClima, a framework benchmarking LLMs on climate science questions across text, video transcripts, and figures.

Meng: This complements the GCA approach by providing a standardized way to evaluate model reasoning across different data types in the climate domain.

Lalam: Today's papers: Detecting Misbehaviors of Large Vision-Language Models is worth hearing for its evidence quantification.

Tom: Beyond Accuracy: Measuring Logical Compliance of Predictive Models is worth hearing because it measures logical compliance.

Jane: SCRIBE: Diagnostic Evaluation and Rich Transcription Models for Indic ASR is worth hearing for its diagnostic evaluation capabilities.

Lu: NodeSynth: Socially Aligned Synthetic Data for AI Evaluation is worth hearing because it provides synthetic data for evaluation.

Meng: Bandit Learning in General Open Multi-agent Systems is worth hearing as it optimizes learning in multi-agent systems.

Tom: So, we have covered the dynamics of internal representations and agent verification methods today.

Jane: The GCA framework addresses a major real-world application in climate science decision support.

Lu: We've also seen work on idea sampling and controlling model behavior at inference time.

Tom: Today's papers: Detecting Misbehaviors of Large Vision-Language Models is worth hearing for its evidence quantification.

Jane: Beyond Accuracy: Measuring Logical Compliance of Predictive Models is worth hearing because it measures logical compliance.

Lu: SCRIBE: Diagnostic Evaluation and Rich Transcription Models for Indic ASR is worth hearing for its diagnostic evaluation capabilities.

Meng: NodeSynth: Socially Aligned Synthetic Data for AI Evaluation is worth hearing because it provides synthetic data for evaluation.

Lalam: Bandit Learning in General Open Multi-agent Systems is worth hearing as it optimizes learning in multi-agent systems.

Tom: That concludes our review of today's research findings.

Lucky paper: 2602.05535: Tom: Alright, let's talk about Evidential Uncertainty Quantification by Tao Huang and colleagues, which is detecting misbehaviors of large vision-language models.

Jane: This paper introduces a training-free framework that decomposes epistemic uncertainty into internal contradictions and ignorance using Dempster-Shafer Theory.

Lu: The core idea is interpreting model output features as evidence, aggregating them to quantify internal conflict and knowledge gaps in a single forward pass.

Meng: They evaluate this across four categories: hallucinations, jailbreaks, adversarial vulnerabilities, and out-of-distribution failures using state-of-the-art LVLMs.

Lalam: The results show that hallucinations correspond to high internal conflict and out-of-distribution failures to high ignorance.

Tom: What's the actual performance lift compared to strong baselines in terms of AUROC?

Jane: They achieved relative improvements of up to ten point five percent in AUROC, and seven point five percent in AUPR for conflict and five point three percent and five point five percent for ignorance respectively.

Lu: The layer-wise evidential uncertainty dynamics analysis is interesting because it shows that ignorance tends to decrease while conflict increases across decoder layers <ref:2602.05535#pg22>.

Tom: So deeper layers are becoming more conflicted and less ignorant, which makes sense if they are compressing redundant input.

Jane: They also found clear distinctions in uncertainty curves across misbehaviors for certain decoder layers, noting that adversarial examples are the most distinguishable because their density curves deviate sharply.

Meng: That sharp deviation is due to the pronounced distributional shift caused by pixel-level perturbations on those examples.

Lalam: The method requires only a single forward pass and no access to specialized layers or auxiliary models, which makes it quite efficient for deployment.

Tom: So, the efficiency trade-off is that it achieves these improvements with substantially lower computational overhead than sampling-based methods that incur ten times the overhead.

Jane: They also note that while every description conveys colour, only about sixteen to eighteen percent used explicit layout terms like left or right.

Lu: That implies there's a design requirement for explicitly requesting and checking those spatial structures because accuracy matters for blind listeners.

Meng: From a deployment standpoint, they claim this whole pipeline produces text and audio outputs in under twenty seconds per image and costs less than zero point zero five per one.

Tom: And they logged every run with the model used, the timestamp, and even a hash of the input image filename in a Google Sheets ledger for tracking consistency.

Jane: That logging is important for debugging and checking consistency later on when you're trying to reproduce an issue.

Lu: It’s about moving past just naming objects and starting to build that narrative and spatial context that blind listeners really need to form a mental model of the scene.

Meng: What this means for us is that if you want to make accessible art descriptions faster, this framework is a solid starting point, but you have to add a layer for checking those spatial terms manually.

Tom: So the main thing here is that AI can generate richer text quickly, but we still need people to validate that it’s actually conveying the spatial relationships correctly.

Jane: And because they did this across fifty artworks using five different AI models, it gives us a solid comparison of what we can expect from these systems in practice.

Jane: And because they did this across fifty artworks using five different AI models, it gives us a solid comparison of what to expect from these systems in practice.

Lucky paper: 2606.20208: Tom: So we're looking at Beyond Accuracy: Measuring Logical Compliance of Predictive Models today, and it introduces the Rule Violation Score or RVS as a complementary metric.

Jane: Right, it quantifies how much predictive models respect predefined logical or domain-specific constraints which is important in high-stakes applications.

Lu: The paper says RVS treats hard rules and soft rules differently, and you can evaluate it on any dataset using SQL queries automatically generated for Horn rules.

Meng: They introduce four desirable properties for such metrics including ground-truth independence, which means the metric should be computable without access to the ground-truth labels used by standard predictive metrics.

Lalam: They also distinguish between hard and soft rules where hard rules represent constraints that must always hold and soft rules represent statistical regularities.

Tom: So for a soft rule, the RVSsoft r is defined as the prediction-induced contradiction rate pr divided by the dataset contradiction rate dr.

Jane: The dataset contradiction rate dr measures the fraction of body-satisfying groundings in D that lead to an established contradiction, and a rule is hard if dr equals zero.

Lu: Experimentally, they found that two models achieving comparable predictive accuracy can exhibit substantially different levels of logical compliance.

Tom: For example, on the Family dataset, CompGCN achieved higher MRR and Hits@one than UniKER but produced substantially more hard-rule violations.

Jane: Interestingly, on FB15k-two hundred thirty-seven AnyBURL and CompGCN had nearly identical MRR and Hits@one scores yet AnyBURL was substantially more compliant with hard rules.

Lu: Per-rule RVS scores indicate which logical behaviors are responsible for compliance differences by showing that CompGCN and UniKER violate different hard rules on Family.

Meng: The dataset contradiction rate dr also serves as a model-independent measure of rule-dataset compliance, highlighting rules that may require revision.

Tom: The paper suggests RVS is useful at three levels: comparing models, auditing datasets, and diagnosing individual rules based on their per-rule scores.

Jane: They also suggest using RVS as a training signal by incorporating rule-violation terms into learning objectives to jointly optimize predictive performance and logical compliance.

Lu: The paper emphasizes that RVS is defined for any finite datasets, any predictive models, and any evaluable logical rules expressed over a relational vocabulary.

Meng: Computation can be implemented using standard Database Management System operations for rules that translate into SQL queries.

Tom: This generality is important because the results support the use of RVS as a diagnostic tool for comparing predictive models and auditing datasets across broader data models like relational databases.

Jane: We also saw some work on how language models use group-linked fields in Language Models Give Group-Linked Fields the Wrong Weight, which suggests display format matters more than just observed changes in reliance.

Lu: That finding connects to the need for explicit spatial language when we look at systems like CANVAS, where only about sixteen to eighteen percent of descriptions used explicit layout terms like left or right.

Meng: So while AI can generate richer text quickly, the paper points out that human input is still needed to ensure those spatial details are actually present in the final output.

Tom: Beyond Accuracy: Measuring Logical Compliance of Predictive Models is worth hearing because it measures logical compliance.

Jane: SCRIBE: Diagnostic Evaluation and Rich Transcription Models for Indic ASR is worth hearing for its diagnostic evaluation capabilities.

Lu: NodeSynth: Socially Aligned Synthetic Data for AI Evaluation is worth hearing because it provides synthetic data for evaluation.

Meng: Bandit Learning in General Open Multi-agent Systems is worth hearing as it optimizes learning in multi-agent systems.

Tom: That concludes our review of the paper Beyond Accuracy: Measuring Logical Compliance of Predictive Models.

Lucky paper: 2605.20712: Tom: Okay, we're looking at SCRIBE, which is Diagnostic Evaluation and Rich Transcription Models for Indic ASR. This paper is focused on making automatic speech recognition better for languages like Malayalam and Kannada by breaking down errors into specific categories.

Jane: Right, so instead of just giving you a total word error rate, SCRIBE decomposes those errors into lexical, punctuation, numeral, and domain-entity rates.

Lu: The methodology involves transforming the reference and hypothesis into typed tokens where each token is classified as a lexeme or numeral or punctuation mark or domain entity.

Meng: They use a sandhi-aware alignment engine that uses extended dynamic programming to find the best alignment score by accounting for insertions, deletions, and Sandhi mappings.

Lalam: The scoring function anchors on exact matches with a score of alpha equals plus four point zero while applying a category-clash penalty beta equals minus three point zero if the tokens don't match categories.

Tom: And for same-category substitutions, they use a Levenshtein-buffered penalty delta equals minus one point five minus zero point two times d where d is the character distance.

Jane: They also have this sandhi score sigma equal to minus zero point five plus sigma negative d of bsplit and bmid over s, which resolves those one-to-two or two-to-one mappings based on phonetic plausibility.

Lu: The output is a diagnostic error vector E which maps failures into ERlex, ERpunc, ERnum, and ERent respectively.

Meng: They aggregate these errors into a diagnostic vector where each component maps to a specific remediation strategy.

Tom: The single global rate they calculate is SCRIBEWER which equals sum over t of total t times the error rate for that category.

Jane: They found this decomposition aligns robustly with human judgment, with a Spearman rho between SCRIBE’s categorical error rates and human dimensions of zero point three six to nine point two.

Lu: The analysis showed a critical divergence where genuine lexical failure accounts for only fifteen point ninety-six percent in the Malayalam Legal set despite a high monolithic word error rate of forty-four point fifty-two percent.

Meng: This proves that core acoustic-phonetic reliability is nearly three times higher than the monolithic scalar suggests.

Tom: Furthermore, they showed that a substantial portion of reported Indic word error rates is an artifact of morphological structure rather than actual acoustic misrecognition.

Jane: The diagnostic value extends to active development through a feedback loop where early iterations showed systematic over-punctuation bias entirely invisible in aggregate word error rate.

Lu: SCRIBE’s ERpunc decomposition isolated this over-punctuation bias to segments with legacy verbatim corpora containing extremely short sequences with misleading terminal punctuation.

Meng: This allows for targeted filtering and refined LLM curation prompts based on that specific error type.

Tom: They also showed that resolving valid morphological merges reduces reported error inflation by roughly thirty percent relative in both Malayalam and Kannada.

Jane: Near saturation of numeral formatting, where ERnum is less than one point, shows a seventy-five to ninety-six percent relative reduction compared to the best baseline across all benchmarks.

Lu: The primary bottleneck they identified is punctuation as ERpunc remains the main challenge particularly in Dravidian languages.

Meng: This points toward prosodic modeling as a next development target because Malayalam reports twelve point twelve percent for ERpunc versus Hindi's six point seventy-three percent.

Tom: SCRIBE provides a diagnostic complement to monolithic word error rate wherever ASR serves as a scribe, enabling targeted remediation for rich transcription tasks.

Jane: The authors release SCRIBE alongside their curation recipe, benchmarks, and open-weight models to enable professional dictation correction thresholds.

Lucky paper: 2605.14381: Tom: We're looking at NodeSynth today, which is about generating socially relevant synthetic queries for AI evaluation.

Jane: It uses an evidence-grounded approach to create these prompts by leveraging a fine-tuned taxonomy generator anchored in real-world evidence.

Lu: The methodology has three steps: Multi-Scale Taxonomy Generation, Evidence Grounding and Automated Annotation, and finally Multi-Factor Synthetic Prompt Generation.

Meng: Step one involves breaking down abstract concepts into a three-tiered hierarchical taxonomy, L1, L2, and L3 using a fine-tuned taxonomy generator called TaG.

Lalam: That TaG gets its granularity enhanced through supervised fine-tuning on two thousand five hundred seventy-six expert-generated instances, increasing the Jaccard score from zero point five four to zero point seven zero.

Tom: Step two uses Gemini two point five and Google Search configuration tool to find credible source documents and perform automated annotation for characteristics like race or ethnicity.

Jane: That grounds the synthetic data in documented real-world harms by ensuring the prompts reflect authentic societal contexts, which is a big thing for safety testing.

Lu: Then step three takes that structured data and injects it into a prompt template to create targeted queries using the L1 through L3 taxonomy.

Meng: The empirical demonstration shows these taxonomy-driven queries elicit failure rates five times higher than human-authored data across four mainstream models.

Lalam: Specifically, when evaluated against the Level two baseline, Level three queries increased the average Medical Advice failure rate from twenty-four point eight six percent to twenty-nine point seven eight percent, which is statistically significant at p = zero point zero zero two seven.

Tom: That jump in failure rate is what really shows how much deeper that taxonomic layer needs to be for accurate testing, right?

Jane: It confirms the necessity of including a taxonomy layer fine-tuned with domain expertise when we're talking about critical areas like medical advice.

Lu: The impact on guard models is also important because NodeSynth L3 achieved the highest bypass rates across tested guard models for both Medical Advice and Self-Harm domains.

Meng: That suggests the guard models might be better at identifying explicit content, missing the more nuanced queries related to sensitive topics.

Lalam: Qualitative error analysis revealed that queries related to “malpractice liability” for Medical Advice contributed most of the unsafe responses across all models.

Tom: So you can pinpoint exact failure patterns at the demographic or geographic level this way, which is very useful for targeted mitigation.

Jane: It also gives evaluators an interpretable diagnostic method to trace where model performance degrades based on those taxonomic intersections.

Lu: The paper addresses scalability by showing that grounded, taxonomy-driven synthetic data can match or exceed the efficacy of human-authored queries.

Meng: NodeSynth provides a practical application by embedding the methodology into a research prototype so users can generate robust datasets from just a few input parameters.

Lalam: While they noted limitations, like susceptibility to hallucination in the Google Search grounding feature, they pointed toward future work on quality metrics for fabricated information.

Tom: The dual-use risk of this approach is something we need to manage by focusing it purely as a data engine for model evaluation and detection.

Lucky paper: 2605.06202: Tom: Alright, let's talk about Bandit Learning in General Open Multi-agent Systems. This paper is tackling the complexity of open systems with heterogeneous rewards and dynamic agent populations.

Jane: The central idea here is formulating a unified open MA-MAB problem, where agents can arrive and depart according to general patterns, even if the same arm has different reward means across those agents.

Lu: They introduce global dynamic regret, which means the comparator at each round is the arm that maximizes the aggregate, equivalent average reward over all currently active agents.

Meng: And they also introduce pre-training error Pm and pre-training degree Dm equals one over Pm for newly arriving agents to quantify how much information an agent carries upon entry from transfer or side information.

Tom: The methodology involves three modules: CERTIFIEDARRIVALTRANSFER, GLOBALVALUEAGGREGATION, and LOCALUPDATEANDBROADCAST to manage this process.

Jane: In the algorithm, they initialize newly arriving agents through certified transfer only when the associated entry error can be bounded.

Lu: Then they aggregate these certified local estimates into global arm-value estimates and select a common arm using a global UCB index that separates continuing agent statistical uncertainty from arrival uncertainty.

Tom: That selection rule is defined as at ← arg maxi in K n Vbt(i) plus Bstat t(i) plus EA t o.

Jane: They prove an upper bound for the total regret RT of order O(M0 log T plus PT t equals one Pm), which is equivalent to O(M0 log T plus PT t equals one At over Dt).

Lu: For the stable-arm regime, they prove that the regret R bar T is less than tau on the good event and less than tau plus delta of T minus tau unconditionally.

Meng: The case studies show that in the linear stochastic case, under strong effective information conditions, regret achieves logarithmic-order regret up to logarithmic factors and square-root order under weaker conditions.

Tom: In the nonlinear parametric case, they show the regret can be either logarithmic or square-root order depending on whether the effective information condition is strong or weak.

Jane: For the clustered stochastic model, they show that under certain conditions, the total regret is bounded by Oe of R base T plus C tau max plus C log T.

Lu: In the zero-knowledge case, if agents enter with zero initial information, the expected regret is bounded by O(M0 log T plus lambda AT).

Tom: The burn-in time tau is characterized as tau equals max of tau stab and tau id, which consists of two conceptually distinct components.

Jane: The first component, tau stab, depends on the active population size Mt and the population gap delta.

Lu: The second component, tau id, depends on how quickly the algorithm identifies the stable globally optimal arm.

Meng: They establish lower bounds showing that the pre-training error term is unavoidable yielding E RT A greater than or equal to omega R base T plus PT t equals one Pm.

Tom: For the stable-arm setting, they prove a lower bound on normalized regret R bar T is greater than or equal to delta tau when identification time is sufficiently large.

Jane: The proof of results in general systems demonstrates that the unified pre-training-error regret bound holds across all models if the effective information condition is met.

Lu: This confirms that the stability perspective refines the worst-case pre-training error bound by showing arrivals do not always create a new learning problem.

Tom: The analysis shows that for linear and nonlinear parametric models, the entry-error contribution is either logarithmic or square-root order depending on whether the effective information condition is strong or weak.

Jane: The clustered model's regret is bounded by Oe of R base T plus C log T.

Tom: The final results show that for all four models, the stable-arm regret is less than tau plus delta G of tau delta T minus tau.

Jane: This shows that even with zero-knowledge arrivals, the stable-arm regime can avoid the cumulative O(T) entry-error cost after the stable arm is identified.