Computer Science (arXiv)

Real-time cognitive load assessment is essential for adaptive human-computer interaction but remains challenging due to limited labeled data and poor cross-subject generalization. Recent ECG foundation models pre-trained on millions of clinical recordings offer rich representations, but cannot be directly applied to wearable devices due to sensor configuration mismatch and task differences. In this paper, we propose CogAdapt, a framework that adapts clinical ECG foundation models to wearable cognitive load assessment. CogAdapt introduces LeadBridge, a learnable adapter that transforms 3-lead wearable signals into anatomically consistent 12-lead representations, and ProFine, a progressive fine-tuning strategy that gradually unfreezes encoder layers while preventing catastrophic forgetting. Evaluations on two public datasets (CLARE and CL-Drive) under leave-one-subject-out cross-validation show that CogAdapt substantially outperforms baselines trained from scratch, achieving macro-F1 scores of 0.626 and 0.768. These results demonstrate the promise of foundation model adaptation for subject-independent cognitive load assessment from wearable sensors.

The Flexible Job Shop Scheduling Problem (FJSP) is the optimal allocation of a set of jobs to machines. Two primary challenges persist in FJSP: the unpredictable arrival of future jobs and the combinatorial complexity of the problem, rendering it intractable for conventional mixed-integer linear programming solvers. This paper proposes an event-based \gls{DRL} approach to solve FJSP with random job arrivals. Specifically, we employ the Proximal Policy Optimization algorithm and use lightweight Multi-Layer Perceptrons to train the \gls{DRL} agent for minimizing the total completion time of all jobs. We design the state representation to be directly accessible from the environment, and limit the learning agent to selecting from among a set of well-established dispatching rules. Simulations show that our \gls{DRL} approach outperforms any of the individual dispatching rules on datasets with varying heterogeneity and job arrival rates. We benchmark our \gls{DRL} against an arrival-triggered mixed-integer linear programming solution and show that our method achieves good performance especially when the datasets are heterogeneous.

Large language models (LLMs) exhibit systematic political bias across a variety of sensitive contexts. We find that LLMs handle counterpart topics from opposing political sides asymmetrically. We refer to this phenomenon as covert political bias and identify 7 categories of techniques through which it operates. We propose two metrics for covert bias: Sentiment Consistency measures symmetry in rhetoric and framing across paired political prompts; Helpfulness Consistency measures symmetric depth and engagement. To reduce both types of covert bias, we introduce Political Consistency Training (PCT), an RL training method with two complementary paradigms: Sentiment Consistency Training and Helpfulness Consistency Training. We show that PCT preserves overall helpfulness, substantially reduces covert political bias, and generalizes to held-out benchmarks. We release our work at https://political-manipulation.ai

Large language models (LLMs) are typically trained on shuffled corpora, yielding models whose knowledge is frozen at train time and whose temporal grounding remains poorly understood. In this work, we study the impact of pre-training dynamics on the acquisition of time-sensitive factual knowledge, focusing specifically on data ordering. Our main contributions are twofold. First, we introduce a comprehensive benchmark of over 7,000 temporally grounded questions and an evaluation protocol that enables analysis of whether models correctly associate facts with their corresponding time periods. Second, we pretrain 6B-parameter models on temporally ordered Common Crawl snapshots and compare them against standard shuffled pre-training. Our results show that sequentially trained models match shuffled baselines on general language understanding and common knowledge while consistently exhibiting more up-to-date and temporally precise knowledge. Temporally ordered pre-training yields improved factual freshness, while shuffled pre-training peaks on older data, possibly due to increased factual repetition. These findings, along with the release of our code at https://github.com/kyutai-labs/kairos , checkpoints, and datasets at https://huggingface.co/collections/kyutai/kairos provide a foundation for future research on continual learning for LLMs.

Children with rare genetic diseases often exhibit distinctive facial phenotypes, yet developing computer vision systems for early diagnosis remains challenging due to extreme data scarcity, privacy constraints, and limited data sharing in pediatric settings. These challenges not only hinder automated diagnosis but also restrict the availability of visual resources for clinical genetic counseling. While prior work has shown that synthetic data can augment real datasets and preserve phenotype-level semantics, it remains unclear whether synthetic data alone is sufficient for learning in ultra-low-resource pediatric settings. In this work, we study the synthetic-only regime for pediatric rare disease recognition. Under a controlled experimental setup, models are trained exclusively on phenotype-aware synthetic facial images at increasing scales. We find that synthetic-only training achieves performance comparable to real-data-only baselines at sufficient scale across multiple backbones, suggesting that high-fidelity synthetic data can approximate clinically meaningful distributions. These findings together further enable the use of synthetic pediatric facial images as privacy-preserving resources for genetic education and counseling, supporting clinician training and patient communication. Our results highlight the potential of computer vision to improve data efficiency and expand accessible visual tools in children's healthcare.

Model cards describe model behavior through a mixture of textual descriptions and structured artifacts, including performance, configuration, and dataset tables. Existing model search systems rely predominantly on semantic similarity over text, which can produce homogeneous result sets and limit exploration of alternatives. We argue that model search is inherently comparative: users want models that are task-aligned yet differentiated in measurable ways. We hypothesize that this balance requires retrieval over condensed, high-quality evidence rather than verbose descriptions, and much of that evidence is concentrated in structured tables. We present StructuredSemanticSearch, a table-driven model search framework built on the ModelTables benchmark. Given a query, StructuredSemanticSearch combines a semantic baseline for task alignment with a structure-aware pipeline that discovers query-related model-card tables using table discovery operators such as unionability, joinability, and keyword search. Retrieved tables are mapped back to model cards under a controlled top-k budget, enabling fair comparison between text-based and table-based retrieval. Beyond retrieval, StructuredSemanticSearch adapts table integration to the model-table domain through orientation-aware integration, producing compact integrated views of tables from partially overlapping and sometimes transposed evidence tables. For evaluation, we introduce a nugget-based, auditable protocol that extracts compact evidence items from model cards, matches queries to condition- or intent-specific nuggets, and measures evidence coverage and diversity over retrieved model-card candidate sets. This protocol also provides a scalable path toward approximate, evidence-based labeling in dynamic model lakes. Experiments on 597 model-recommendation queries show improved nugget coverage for the structure-aware pipeline than semantic baseline

Discrete diffusion models are often trained through clean-data prediction, but the prediction can be used in different ways to define the reverse dynamics. In Masked Diffusion Models (MDM) these choices largely coincide, whereas in Uniform Diffusion Models (UDM) they do not. We show that the standard plug-in bridge parameterization for UDM is not optimized by the denoising posterior, but by a leave-one-out posterior that predicts each clean token without using its own noisy observation. This identifies a mismatch between the plug-in ELBO and the usual cross-entropy denoising objective. We characterize the leave-one-out target and derive exact conversions between the denoiser, the leave-one-out posterior, and the score. These conversions allow us to disentangle parameterization and training objective. Our results also lead to inference improvements without any additional training through an informed predictor-corrector sampler and improved temperature sampling based on the leave-one-out predictor. We further introduce an absorbing-state reformulation of uniform diffusion that preserves the UDM joint law while decomposing it into masked-diffusion-like sampling operations, with simpler denoising posteriors, carry-over unmasking, and a natural remasking mechanism. On language modeling, leave-one-out parameterizations consistently improve UDM generation, while the absorbing construction matches or surpasses masked diffusion. These results suggest that the empirical gap between masked and uniform diffusion is driven less by the choice of marginals themselves than by parameterization and sampling design. The code and models can be found at https://github.com/samsongourevitch/rev_udm.

Large language models (LLMs) increasingly excel at mathematical reasoning, but their unreliability limits their utility in mathematics research. A mitigation is using LLMs to generate formal proofs in languages like Lean. We perform the first large-scale evaluation of this method's ability to solve open problems. Our most capable agent autonomously resolved 9 of 353 open Erdős problems at the per-problem cost of a few hundred dollars, proved 44/492 OEIS conjectures, and is being deployed in combinatorics, optimization, graph theory, algebraic geometry, and quantum optics research. A basic agent alternating LLM-based generation with Lean-based verification replicated the Erdős successes but proved costlier on the hardest problems. These findings demonstrate the power of AI-aided formal proof search and shed light on the agent designs that enable it.

While ubiquitous wearable sensors capture a wealth of behavioral and physiological information, effectively transforming these signals into personalized health insights is challenging. Specifically, converting low-level sensor data into representations capable of characterizing higher-level states is difficult due to high phenotypic diversity and variation in individual baseline health, physiology, and lifestyle factors. Moreover, collecting wearable data paired with health outcome annotations is laborious and expensive, and retrospective annotation remains practically unfeasible, contributing to a scarcity of data with high-quality labels. To overcome these limitations, we propose a foundation model for wearable health that is pretrained on more than one trillion minutes of unlabeled sensor signals drawn from a large cohort of five million participants. We demonstrate that the joint scaling of model capacity and pretraining data volume leads to systematic improvements in performance, as evaluated on a diverse set of 35 health prediction tasks, spanning cardiovascular, metabolic, sleep, and mental health, as well as lifestyle choices and demographic factors. We find that this population scale representation unlocks label-efficient few-shot learning and generative capabilities for robust daily metric estimation. To further leverage this learned representation, we deploy a classroom of LLM agents to autonomously search the space of downstream predictive heads built on the model embeddings, showing broad performance improvements that increase with LLM model capacity. Finally, we show how integrating these downstream predictors into a Personal Health Agent can support model responses that are more relevant, contextually aware, and safe, and we validate this via 1,860 ratings from a cohort of clinicians.

We identify a sharp interaction-degree threshold for the classical simulation of QAOA with $2$-local cost functions. At degree $3$, classical sampling from depth-$1$ QAOA with small multiplicative error would collapse the polynomial hierarchy to its third level. At degree $2$, exact classical sampling from depth-$p$ QAOA on $n$ qubits runs in time $n^{O(1)}$ whenever $p = O(\log n)$. The hard degree-$3$ instances have trivially optimizable cost functions, so sampling hardness does not by itself imply a quantum optimization advantage.

Random forests are widely used in fields involving sensitive tabular data, but existing approaches to enforcing differential privacy (DP) typically degrade performance to the point of impracticality. In this paper, we introduce Lumberjack, a differentially private random forest algorithm that achieves substantially higher utility by constructing large random decision trees and then applying aggressive, privacy-preserving pruning to retain only sufficiently populated nodes. A key component of our approach is a novel $(\varepsilon,δ)$-DP heavy hitter detection algorithm for hierarchical data, whose error is $O_{\varepsilon,δ}(\sqrt{\log h})$ for trees of height $h$ and may be of independent interest. This favorable scaling enables the use of significantly deeper trees than in prior work, leading to improved expressiveness under privacy constraints. Our empirical evaluation on benchmark datasets shows that Lumberjack consistently outperforms prior DP random forest methods, establishing a new state of the art. In particular, our approach yields substantial improvements in the privacy-utility trade-off for practical privacy budgets. Our findings suggest that carefully designed DP random forests can close much of the utility gap, highlighting a promising and underexplored direction for future research.

As generative image models evolve rapidly, the perceptual gap between generated and real images continues to narrow, making AI-generated image detection increasingly challenging. Many existing methods exploit frequency-domain cues for detection, typically described as frequency-domain artifacts or high-frequency discrepancies. However, the specific and recurring spectral regularities remain insufficiently understood and characterized. In this paper, we systematically analyze the one-dimensional radial log-power spectra of real and generated images. We find that generated images do not necessarily exhibit higher or lower energy across the entire spectrum or high-band range. Instead, their spectra deviate from the power-law decay and show an anomalous uplift in the ultra-high-frequency tail. We term this phenomenon spectral tail uplift. We further attribute this phenomenon to nonlinear harmonic accumulation in trained generative models, suggesting that it can serve as a structural cue across generative architectures. Based on this observation, we propose Spectral Tail Auxiliary Learning (STAL), a frequency-domain auxiliary supervision framework for generalizable AI-generated image detection. STAL transfers spectral-tail cues from a tail-aware frequency teacher to a spatial detector during training, while all frequency-domain modules are discarded at inference time. Consequently, STAL introduces no inference overhead. Extensive experiments on 9 public datasets show that STAL achieves strong generalization and stability across generators, data distributions, and real-world scenarios.

Modern smart grids rely on dense measurement infrastructures, communication links, and intelligent field devices. Although this improves supervision and control, it also increases vulnerability to cyber-physical disruptions. Operators must distinguish physical incidents, such as faults or line disturbances, from malicious actions, such as false data injection or unauthorized command execution. This chapter investigates this problem using the well-known MSU/ORNL Power System Attack Dataset. The proposed method combines machine learning with genetic-algorithm-based feature selection. The objective is twofold: to classify attack and natural events accurately, and to determine whether a reduced set of physically informative PMU/IED measurements can support reliable detection. Several baseline models are evaluated, including logistic regression, RBF-SVM, XGBoost, Random Forest, and Extra Trees. The results show that tree-based ensemble models are the most effective for the considered dataset, with Extra Trees providing the strongest full-feature baseline. After feature selection, the GA + Extra Trees model reduces the clean PMU feature space from 112 attributes to an average of 27.4 attributes over five runs, while increasing macro-F1 from 0.9118 to 0.9212 and ROC-AUC from 0.9791 to 0.9837. These results indicate that many synchronized electrical measurements are redundant. A compact subset of phasor-based features can still provide accurate and interpretable anomaly detection in smart grids.

Autonomous systems have achieved superhuman performance in isolation or simulation, yet they remain brittle in shared, dynamic real-world spaces. This failure stems from the dominant single-agent paradigm for physical applications, where other actors are ignored or treated as environmental noise, preventing effective coordination. Here we show that multi-agent reinforcement learning provides the essential safety scaffolding required for real-world interaction. Using high-speed quadrotor racing as a high-stakes testbed, we train agents to navigate complex aerodynamic interactions and strategic maneuvering with a variable number of racers. Through league-based self-play, agents evolve sophisticated anticipatory behaviors, including proactive collision avoidance, overtaking, and handling multi-agent physical interactions, including aerodynamic downwash. Our agents outperform a champion-level human pilot in multi-player races at speeds exceeding 22 m/s, while simultaneously reducing collision rates by 50 % compared to state-of-the-art single-agent baselines. Crucially, training with diverse artificial agents enables zero-shot generalization to safer human interaction. These results suggest that the path to robust robotic co-existence lies not in isolated safety constraints, but in the rigorous demands of multi-agent interaction. Multimedia materials are available at: https://rpg.ifi.uzh.ch/marl

Quoridor is an award-winning abstract strategy game designed by Mirko Marchesi and published in 1997. Similar games include Maze Attack, Blockade (also known as Cul-de-sac), and Pinko Pallino. In line with chess, checkers, Go, and other classic combinatorial games, Quoridor is a turn-based, deterministic, perfect-information game played on a square grid. We show that it is PSPACE-complete to determine whether a given player has a winning strategy in a given Quoridor position on a board with size $n \times n$. We prove this by reduction from Gpos(POS CNF), a Boolean formula game originally defined in 1978 by T. Schaefer.

Real-world sensor-based learning systems require uncertainty estimation that is both reliable and computationally efficient. Evidential Deep Learning (EDL) provides single-pass uncertainty estimation by modeling the class probabilities via Dirichlet distributions, where the Dirichlet parameters are predicted by a learned neural network mapping. However, this approach can lead to computational challenges, as Dirichlet expected objectives are more complex than standard supervised learning losses, complicating their analysis and implementation. We address this issue by approximating the objective of the first-order empirical risk minimization problem induced by EDL with a plug-in loss evaluated at the Dirichlet mean and show that, under mild assumptions, the approximation error decays with growing evidence for a broad class of loss functions, including mean-squared error and cross-entropy loss. As a special case, our analysis provides justification for the use of softmax in the context of uncertainty estimation, since under a particular evidence-to-Dirichlet mapping, our framework includes the standard softmax classifier. We validate the proposed simplified objectives on the Google Speech Commands dataset and show that they achieve predictive accuracy and selective prediction performance comparable to classical EDL, while being simpler to implement using standard deep learning losses and training pipelines. To the best of our knowledge, this empirical analysis is the first to obtain coverage-accuracy trade-offs for speech recognition tasks through EDL.

Parameter-efficient fine-tuning enables fast personalization of text-to-image diffusion models, but composing multiple custom concepts remains challenging due to representation interference. Existing modular methods either rely on expensive post-hoc fusion or freeze adaptation subspaces, which limit expressiveness and concept fidelity. To address this trade-off, we propose Sequential regularized LoRA (SeqLoRA), a constrained continual learning framework that jointly optimizes both LoRA factors via bilevel optimization. Theoretically, we establish strong convergence guarantees for our algorithm and model the residual layer activations as a matrix sub-Gaussian process to derive high-probability bounds on catastrophic forgetting. We further prove that learning the LoRA basis from data minimizes residual interference energy more effectively than frozen-basis methods. Experiments on multi-concept image generation demonstrate that SeqLoRA improves identity preservation and scalability across up to 101 concepts, while avoiding costly fusion and reducing attribute interference in composed generations.

Decision trees partition the feature space using hard binary thresholds, assigning identical confidence to instances far from a decision boundary and to those directly on it. We introduce ternary decision trees, which augment each split node with an uncertainty zone of half-width delta centered on the optimal threshold. Instances in this zone receive predictions formed by weighted blending of both child subtrees and are flagged as boundary-uncertain, signaling that downstream applications may treat these predictions differently. Crucially, delta is computed locally at each node from statistics already available during standard CART split finding, requiring no external noise specification. We propose and evaluate five delta-estimation methods: quality-plateau (plateau width of the split criterion curve), class-overlap (empirical class-distribution overlap), gain-ratio (split quality relative to split entropy), node-bootstrap (threshold variance under node-level resampling), and margin (SVM-inspired distance to the nearest cross-class training example). Evaluated across 72 OpenML-CC18 datasets with 5-fold cross-validation, all five methods with probabilistic routing significantly outperform standard CART on decided accuracy (Wilcoxon signed-rank, p < 0.001). The margin method achieves the best efficiency (0.104 accuracy gain per unit of boundary-uncertain flagging rate), wins on 42 of 72 datasets, and requires zero additional hyperparameters. Analysis on three Breiman synthetic benchmarks reveals that margin is self-calibrating on clean data while node-bootstrap and quality-plateau best track theoretical irreducible error. Experiments on four medical and financial datasets demonstrate practical value: on mammography, node-bootstrap achieves +0.71% decided accuracy by flagging 10.8% of screening cases as boundary-uncertain.

Shapley and Banzhaf interactions capture the complex dynamics inherent in modern machine learning applications. However, current estimators for these higher-order interactions trade off between speed and accuracy. To overcome this limitation, we introduce ProxySHAP. ProxySHAP reconciles the high sample efficiency of tree-based proxy models with a principled path to consistency via residual correction. On a theoretical level, we derive a polynomial-time generalization of interventional TreeSHAP to compute exact interaction indices for tree ensembles, successfully bypassing exponential tree-depth dependencies in prior methods. Furthermore, we formally analyze the residual adjustment strategy, characterizing the specific conditions under which Maximum Sample Reuse (MSR) corrects proxy bias without its variance scaling exponentially with interaction size. Extensive benchmarking demonstrates that ProxySHAP sets a new state-of-the-art standard for approximation quality, including in large-scale applications with thousands of features. By achieving the lowest error in both small- and large-budget regimes, ProxySHAP significantly outperforms the prior best estimators ProxySPEX and KernelSHAP-IQ, while also delivering superior performance on downstream explainability tasks.

Distillation attacks create a deployment trade-off for model providers: the same outputs that make a model more useful can also make it easier to imitate. We study this trade-off through a minimax game between a utility-constrained teacher and an adaptive student. Our framework yields tractable one-sided response rules: an adaptive evaluation rule in which the student reweights high-value examples, and a teacher-side defense template that suppresses outputs most useful for distillation. From a cheap proxy for example value, we derive Product-of-Experts (PoE), a simple forward-pass-only defense that combines the teacher with a proxy student during generation. Empirically, adaptive evaluation reveals a large passive--adaptive gap: on state-of-the-art defenses, adaptive students recover substantially more capability than passive evaluation suggests on GSM8K and MATH. Under this stronger evaluation, the apparent robustness gap between expensive defenses and PoE narrows considerably, while PoE remains substantially cheaper and preserves higher-quality reasoning traces. Overall, our results suggest that strong distillation remains difficult to stop, and that progress on antidistillation should be judged against adaptive students rather than passive ones. Our code is available at: https://github.com/ysfalh/distillation-game.

Optimization over the intersection of two manifolds arises in a broad range of applications, but is hindered by the coupled geometry of the feasible region. In this paper, we prove that the regularities -- clean intersection and intrinsic transversality -- are equivalent, which yields a tractable projection onto the tangent space of the intersection. Therefore, we propose a geometric method that employs a retraction on only one manifold and updates the iterate along two orthogonal directions. Specifically, the iterates stay on one manifold, and the two directions are responsible for asymptotically approaching the other manifold and decreasing the objective function, respectively. Under intrinsic transversality, we derive the convergence rate for both the feasibility and optimality measures, and show that every accumulation point is first-order stationary. Numerical experiments on problems stemming from sparse and low-rank optimization, including fitting spherical data, approximating hyperbolic embeddings on real data, and computing compressed modes, demonstrate the effectiveness of the proposed method.

Biomedical knowledge graphs (KGs) treat disease associations as static facts, but temporal information is crucial for clinical reasoning, e.g., a symptom diagnostic of one disease at age 3 may imply a different disease at age 13. Existing KGs such as PrimeKG, Hetionet, and iKraph do not encode when a finding becomes clinically relevant over the course of a disease. This limits their usefulness for longitudinal clinical reasoning and retrieval augmentation. We introduce ChronoMedKG, a temporal biomedical knowledge graph that contains 460,497 evidence-linked triples (filtered from 13M raw extractions) covering 13,431 diseases. Each association is tied to temporal components like onset window or progression stage, which are backed by PMID-traceable evidence and a multi-signal credibility score. The graph is constructed through a disease-autonomous multi-agent pipeline in which multiple frontier LLMs independently extract knowledge from PubMed and PMC literature. Only those relations are kept that are supported by multi-model consensus, survive credibility filtering, as well as ontology alignment. ChronoMedKG scored 92.7% agreement against Orphadata and adds temporal grounding for 6,250 diseases absent from HPOA, Orphadata, and Phenopackets, including 1,657 Orphanet-coded rare diseases. We further introduce ChronoTQA, a benchmark of 3,341 questions across eight task types (six temporal plus two static controls), with a 12-question supplementary probe. Frontier LLMs lose roughly 30 points moving from static to temporal questions; ChronoMedKG retrieval rescues 47-65% of their long-tail failures, against 17-29% for HPOA-RAG. As such, ChronoMedKG provides a crucial temporal axis for retrieval-augmented clinical systems that was previously absent.

Every Python function deployed as an LLM tool must today exist in two forms: an HTTP endpoint for human-facing clients and CI pipelines, and an MCP tool registration for agent runtimes such as Claude and Cursor. These representations share business logic yet diverge in all the surrounding machinery (routing, validation, serialisation, streaming, and schema maintenance), and they drift apart as the underlying code evolves. We present HarnessAPI, a Python framework that eliminates this duplication by treating a typed skill folder as the single source of truth. From one handler.py plus Pydantic schemas, the framework automatically derives a streaming HTTP endpoint with Server-Sent Events, an interactive OpenAPI/Swagger UI, and a zero-configuration MCP tool, all served from a single process. Dual-mode content negotiation lets the same handler serve SSE-streaming and JSON-returning clients with no handler changes. A dynamic code-generation mechanism ensures Pydantic type annotations propagate correctly to FastMCP's inspection layer, resolving a technical limitation that prevents naive closure-based registration. Measured across six representative skills using cloc, HarnessAPI reduces framework-facing boilerplate by 74% compared with a manually maintained dual-stack implementation (FastAPI server + FastMCP server). HarnessAPI subclasses FastAPI, inheriting its full middleware, dependency-injection, and deployment ecosystem. It is available at https://github.com/edwinjosechittilappilly/harnessapi and on PyPI (pip install harnessapi)

We investigate whether acoustic emotion recognition models can serve as proxies for the Pathos dimension in political speech analysis, as operationalised by the TRUST multi-agent large language model (LLM) pipeline. Using a Bundestag plenary speech by Felix Banaszak (51 segments, 245 s) as a case study, we compare three analysis modalities: (1) emotion2vec_plus_large, an acoustic speech emotion recognition (SER) model whose continuous Arousal and Valence values are derived via post-hoc Russell Circumplex projection; (2) Gemini 2.5 Flash, an LLM analysing the full speech audio together with its transcript in an open-ended, context-aware fashion; and (3) TRUST-Pathos scores from a three-advocate LLM supervisor ensemble. Spearman rank correlations reveal that Gemini Valence correlates strongly with TRUST-Pathos (rho = +0.664, p < 0.001), whereas emotion2vec Valence does not (rho = +0.097, p = 0.499). We further demonstrate, via a systematic quality evaluation of the Berlin Database of Emotional Speech (EMO-DB) using Gemini in an open-ended annotation paradigm, that standard SER benchmark corpora suffer from acted speech, cultural bias, and category incompatibility. Our results suggest that LLM-based multimodal analysis captures semantically defined political emotion substantially better than acoustic models alone, while acoustic features remain informative for low-level Arousal estimation. Future work will extend this approach to video-based analysis incorporating facial expression and gaze.

Large language model post-training methods such as supervised fine-tuning (SFT), reinforcement learning (RL), and distillation are often analyzed through their loss functions: maximum likelihood, policy gradients, forward KL, reverse KL, or related objective-level variants. We study a complementary factor: the state distribution on which supervision is applied. For an autoregressive policy, a state is a prompt plus generated prefix. SFT trains on fixed dataset states, while RL and on-policy distillation (OPD) train on states induced by the current learner. We formalize post-training as state-distribution shaping and run a controlled smallscale study using Qwen3-0.6B-Base on GSM8K, with TruthfulQA and MMLU as retention evaluations. Our results show three phenomena. First, a mild SFT run improves GSM8K with little forgetting, while a stress SFT run causes substantial retention loss. Second, OPD from a degraded SFT teacher surpasses that teacher on GSM8K, TruthfulQA, and MMLU, despite using the teacher as its only supervision source. Third, a lightweight on-policy RL run improves GSM8K while preserving retention. These results support a state-centric view of post-training: the source and locality of training states can be as important as the form of the supervision signal.