Abstract: Canistel (Pouteria campechiana (Kunth) Baehni) is a tropical evergreen fruit tree native to Mesoamerica and increasingly cultivated in southern China for its high-value, nutrient-dense fruits. Despite growing commercial interest, comprehensive knowledge of this species remains fragmented. This review systematically synthesizes current research on canistel, covering its botanical characteristics, germplasm resources, propagation and cultivation techniques, nutritional composition, phytochemical diversity, pharmacological bioactivities, and food processing applications. The fruit is rich in carbohydrates, carotenoids (particularly lycopene and β-carotene), phenolic compounds (including gallic acid, catechin, and myricitrin), and essential minerals, while the seeds represent an underutilized source of high-amylose starch (up to 68.1%) with superior thermal stability. Pharmacological studies have demonstrated significant antioxidant, anti-inflammatory, hepatoprotective, analgesic, and immunomodulatory activities, largely attributable to its polyphenolic and triterpenoid constituents. Recent advances in encapsulation and oil-based fortification have improved the stability and bioaccessibility of canistel carotenoids, opening new avenues for functional food development. However, critical research gaps remain, including limited germplasm characterization, insufficient molecular breeding tools, narrow product diversification, and a lack of long-term toxicological and clinical data. This review identifies these challenges and proposes future research directions to facilitate the sustainable industrial development of canistel.

Abstract: Winter surface condensation on the walls and false ceiling of paper production halls degrades the building fabric, promotes mould growth and disturbs the process environment, yet the choice between the available remedies — envelope insulation, ventilation intensification and mechanical dehumidification — is rarely supported by a common quantitative basis. This paper reports a measurement-calibrated assessment of the three solution families for a tissue mill in Petrești, Romania, housing two tissue machines in a 34 800 m³ hall enclosed by an uninsulated 200 mm reinforced-concrete wall (U = 3.89 W m⁻² K⁻¹). Two measurement campaigns established an exhaust airflow of 85 500 m³ h⁻¹ (2.46 h⁻¹) and a process-driven humidity-ratio increment across the hall of Δx = 8.0 g kg⁻¹ with one machine running and 12.2 g kg⁻¹ with both, corresponding to a moisture uptake of 18.4 and 27.2 t d⁻¹ respectively — 3.7 to 3.9 times the value declared in the process water balance, which indicates substantial hood spillage. A one-dimensional surface-temperature model coupled to a psychrometric zone model was driven by an 8 760-hour typical meteorological year, yielding 5 053 cold-season hours (outdoor temperature below 12 °C). On the existing wall the internal surface resistance carries 50.6 % of the total thermal resistance, so the surface temperature closely tracks the outdoor temperature and condensation occurs in 91–100 % of cold-season hours for any indoor temperature between 20 and 35 °C; eliminating it by heating alone would require 45–52 °C indoors. Adding 50 mm of external EPS and holding the hall at 27 °C removes condensation in all 5 053 hours, saves 826 MWh per season and pays back in 1.6–2.6 years, whereas ventilation intensification (2.3× the present airflow) and hybrid coil-plus-desiccant dehumidification achieve the same technical result at ten-year costs of 3.6–4.1 M€. Once the wall is insulated the dehumidification duty falls to zero and the required airflow drops below the installed capacity: insulation does not compete with the alternatives, it makes them unnecessary.

Abstract: Simulation-based robot learning is gated by a manual step that repeats at every deployment site: constructing a physical digital twin of the scene. This paper measures how far unmodified, off-the-shelf code agents can automate that step, a setting referred to as Frontier Agent as Scene Constructor (FASC). Frontier code agents such as Claude Code and Codex, each used as shipped, rebuild eight BEHAVIOR-1K rooms from one photograph, an asset catalog and the simulator. Each reconstruction is scored against the room's known 3D layout on ACDC's sim-to-sim protocol, with memorisation and file-access contamination controlled. The strongest agent on the rooms, Claude Code with Opus 5, retrieves the exact asset model for 52% of the scored objects, against 6% for the same agent with Opus 4.6, released seven months earlier. With Opus 5 it also places objects within 39 cm of ground truth across repeated runs, a 4x reduction in placement error over that predecessor. A gap to ACDC's curated pipeline, at 6 cm, remains. Next, the strongest agent extends to outdoor farm scenes, where no calibrated camera or 3D ground truth exists, using a procedural plant model and common asset libraries. Validated against their photographs in DINOv2 embedding space, 75% of the twins rank their own photograph first among the eight reference photographs in cosine similarity. Together with prior results on trajectory generation, these findings indicate that a generic code-agent harness covers a growing share of embodied tasks, and that its capabilities improve with each frontier-model generation while requiring no embodiment-specific harness engineering.

Abstract: Discrete symmetries constrain the motion and real axis crossings of the interference zeros in the meromorphic continuation of a wave field, thereby constraining the singular structure of its logarithmic derivatives. We consider two counter-propagating second-order rational pulses that satisfy the one-dimensional massless wave equation exactly. With q=rexp(iφ) denoting the relative complex amplitude, the field admits two antilinear reflection symmetries: real q preserves collision centered spacetime inversion followed by complex conjugation, whereas |q|=1 preserves fixed time spatial reflection followed by conjugation up to an overall phase. The balanced in phase state q=1 is the nondegenerate intersection of these symmetry manifolds; q=-1 produces global cancellation on the collision slice. For q=1, the two interference zero branches lie on the imaginary axis and cross the real axis at ct=x₀±l, on opposite sides of the pulse center collision. Away from the symmetry manifolds, the zero trajectories deform continuously and the associated pairing constraints are lost, while the crossing conditions remain available in closed form. A logarithmic complex action representation yields local momentum, energy, transport ratio, and a second-order complex quantum potential without altering the underlying wave dynamics. Near an isolated non-characteristic moving zero, the leading simple pole factors cancel in the transport ratio, whereas the second-order term develops a double pole. The leading real axis response therefore is scaled as d_min^(-2). The reference finite window fit gives an exponent of -1.885 (ρ=-0.995), and the fitted exponent approaches -1.998 as the fitting interval is restricted toward the isolated zero regime. These results establish an exact benchmark relating antilinear symmetry, complex zero-pole geometry, and real axis differential amplification.

Abstract: A growing challenge for protected-area governance is the sustainable coexistence of local communities and wildlife, especially in the rapidly changing socio-ecological landscapes of the Global South. Community participation in long-established protected areas has been widely studied. However, dynamics of participation in newly declared parks, where in-migration, insecure tenure, livelihood dependence, and increasing human–elephant conflict since the COVID-19 pandemic continue post-designation, are poorly studied. Using a mixed-methods approach consisting of a household survey (n = 345) and in-depth interviews (n = 20), this research examines the key factors that influence community participation and coexistence in Kaeng Krung National Park, Thailand. Participation is assessed in four dimensions: decision-making, operations, benefit-receiving, and evaluation. Results showed a high general willingness to participate (mean = 3.50), especially in decision-making and operational activities, while evaluative roles were only moderately engaged. Qualitative findings showed that length of residence, land tenure security, perceived benefits and disadvantages including loss of livelihoods from recurrent wild elephant encounters, and access to communication and training were strong influences on participation patterns. The combined analysis of both datasets provides an empirically based framework to comprehend the importance of livelihood motivations, trust, human–wildlife conflict mitigation and skill-related barriers as key factors in engagement. This has policy-relevant implications for the promotion of sustainable coexistence in protected-area governance in Thailand and other similar settings.

Abstract: Deepresearch (DR) agents interact with real-world web environments through multi-turn search and visit, causing their contexts to grow rapidly over time. We observe that, even after DR Agentic Reinforcement Learning (DR-RL), 61.6% of the model’s remaining prediction errors can still be at- tributed to insufficient long-context understanding, including long-context hallucination and failures in cross-document evidence integration. It motivates us to further break the bottleneck of DR-RL by strengthening the model’s long-context ability. However, effective LongContext training requires more than simply increasing context length. To bridge the data gap, we propose ‘DR Rollouts to LongContext-QA (DR-to-Long)’. The method repurposes DR-RL trajectories, which naturally con- tain search histories, visited webpages, evidence snippets, and final-answer supervision. It then replaces the compact snippets and webpage summaries in each trajectory with the full contents of their corresponding URLs, producing substantially longer multi-document contexts while preserving the original evidence relationships. Building on DR-to-Long, we introduce DLD (DR → LongQA → DR)-RL. DLD-RL first performs a short DR-RL stage to collect rollout trajectories, which are then converted into LongQA instances at zero annotation cost. The model is subsequently optimized with LongQA-RL to strengthen LongContext ability, followed by full DR-RL to continue improving its DR capability. Experiments show that DLD-RL outperforms standard DR-RL by 7.3% on three Deepresearch benchmarks and improves performance by 13.5% on three long-context benchmarks.

Abstract: Vibration-based condition monitoring of beam-like metallic components requires vibration modes that are both sensitive to damage and consistent across different damage locations and severities. This study evaluates the damage sensitivity of the first eight bending-mode frequencies measured from 19 aluminum-alloy 6063 cantilever beams instrumented with magnetoelastic vibration sensors. The examined configurations comprised one undamaged reference beam and 18 damaged beams combining six crack-like defect locations with three defect depths. Normalized absolute frequency shifts were used to determine the mean sensitivity of each mode, its response to increasing defect depth, and the consistency of this response across the examined locations. The combined modal response increased progressively with defect depth, although the magnitude of the change depended on the selected mode and defect location. Mode f₇ exhibited the highest mean sensitivity of 1.234%and amonotonic-consistency ratio of C₇=1, indicating an increasing response with defect depth at all six locations. Modes f₂ and f₄ also demonstrated complete monotonic consistency but lower overall sensitivity, whereas f₅ showed comparatively high sensitivity with greater dependence on defect location. Mode f₈ presented the lowest sensitivity and consistency. These findings indicate that the available modes do not need to be treated equally. Prioritizing modes that combine high sensitivity with consistent depth-dependent behavior could simplify the processing and interpretation of modal information for preliminary condition monitoring of metallic industrial components. Further repeated measurements and validation under operational conditions are required before diagnostic thresholds can be established.

Abstract: For the past fifty years, the physical foundations of black hole (BH) thermodynamics have remained heavily contested because their original derivations rest at the theoretical intersection of general relativity and quantum field theory (QFT). This work provides an alternative macroscopic formulation by treating the physical vacuum in and around a BH as an elastic continuum curved by the presence of mass. According to Hooke’s law of elasticity, the radial strain at the event horizons of all BHs must saturate at the exact same constant value. Under this principle of universal maximum strain, the laws of BH thermodynamics emerge directly from macroscopic continuum mechanics. This eliminates the need for QFT input, mirroring the contemporary macroscopic descriptions of the Casimir effect. Our calculations are formalized using the Reformulated Planck System (RPS), which relies exclusively on Planck’s constant h. This native approach prevents the dimensional contamination of 3D continuum mechanics caused by the 2D geometric factor inherently embedded in Dirac’s reduced constant ℏ, a salient miscue first recognized in the iconic definitions of the fine-structure constant α and the gravitational coupling constant α_g which firmly precluded the identification of the α-dependence of the weak coupling constant \( \alpha_{\rm w}^{} = \sqrt{\alpha} \). In BH thermodynamics, this RPS-based approach reveals higher Hawking temperatures and lower Bekenstein entropies, both by a factor of 2π.

Abstract: Circulating tumor DNA (ctDNA) has emerged as a transformative biomarker in lung cancer management. Its clinical applications span from the well-established role in therapy selection for advanced non-small cell lung cancer (NSCLC) to rapidly evolving uses in treatment response monitoring, minimal residual disease (MRD) detection, perioperative immunotherapy assessment, and early cancer detection. This review synthesizes the current evidence and future perspectives on ctDNA across the lung cancer care continuum, highlighting key clinical trials, guideline recommendations, technological advances, and remaining challenges for broader clinical implementation.

Abstract: Similarity reductions of unsteady two-dimensional flow are almost always constructed from the boundary-layer equations, in which streamwise diffusion and the transverse momentum balance are discarded at the outset. We ask when such a reduction is also a reduction of the full Navier–Stokes equations, using the diffusion-time-scale variables \(\eta=y/\delta(x)\), \(\tau=\nu t/[\delta(x)]^2\) of Sun [32]. All chain rules collapse onto the stretching operator \(\mathcal{D}=\eta\partial_\eta+2\tau\partial_\tau\), both momentum equations transform in closed form, and the discarded terms enter through a single coefficient \(\varepsilon=\delta'^2\). For the family \(U=Cx^m\) the reduction closes in \( (\eta,\tau) \) alone if and only if \(m=1\) or \(m=-1\). The second case is a diverging channel, and there the reduced problem is elliptic in \( (\eta,\tau) \), with principal symbol \( (\xi_\eta^2+\varepsilon s^2)^2 \) degenerating only on \(\tau=0\): the similarity time is not an evolution variable, data must be posed at both ends, and marching in \(\tau\) is inconsistent. Solving the resulting boundary-value problem, we find that the influence of the terminal condition decays as \(\exp[-(T-\tau)/L]\), and that \(L\) collapses on the wedge half-angle, \(L\simeq2.2\,\theta_w\), to within 14% over a factor of 16 in Reynolds number, the prefactor depending on the sense of the flow but the scaling not. The upstream influence is therefore geometric rather than viscous: it vanishes with the divergence angle and is essentially independent of Re, so that marching is safe in slender geometries at any Reynolds number and unsafe in wide ones however viscous. Three exact solutions in elementary functions, obtained from the invariance of \(\zeta=\eta/\sqrt{\tau}=y/\sqrt{\nu t}\) under \( D \), and the Jeffery–Hamel steady limit, serve to verify the reduction throughout.

Abstract: Deep learning methods for pneumonia detection from chest X-ray images have achieved promising results; however, their reliability is limited by scarce labelled data, susceptibility to spurious correlations, poor cross-domain generalisation, and inadequate interpretability. This study proposes CARE-Net, a unified framework integrating radiology-aware Self-Supervised Learning (SSL), Causality-Aware Training (CAT), and Anatomy-Constrained Causality-Aware Attention (ACCA). Radiology-aware SSL learns robust representations from labelled and unlabelled chest X-ray images, while the Invariant Pneumonia Feature Loss (IPFL) reduces dependence on environment-specific information. ACCA further constrains model attention toward anatomically relevant lung regions to improve explanation alignment. Experimental evaluation on the RSNA and Chest X-ray Pneumonia datasets demonstrated strong classification performance. On the Chest X-ray Pneumonia dataset, CARE-Net achieved mean accuracy of 97.9 ± 0.7%, precision of 97.2 ± 0.8%, recall of 97.6 ± 0.7%, F1-score of 97.8 ± 0.6%, and AUC of 0.978 ± 0.005. On RSNA, it achieved accuracy of 95.6 ± 0.6%, AUC of 0.972 ± 0.004, and F1-score of 0.957 ± 0.005. The framework reduced performance variability by over 50%, improved robustness under distribution shift, and achieved lung-region IoU of 0.47 versus 0.32 for Grad-CAM. These findings demonstrate the potential of integrating SSL, causality-aware training, and anatomy-constrained attention for robust and interpretable pneumonia classification.

Abstract: Plague remains a significant natural-focal infection, yet the mechanisms underlying the long-term persistence of Yersinia pestis during inter-epizootic periods remain incompletely understood. This study integrated published evidence with epizootiological monitoring data from Kazakhstan (1990–2025) and detailed field and laboratory investigations conducted during 2021–2023 to develop an integrated ecological model of plague focus stability. Abiotic and biotic niches, including soil and burrow substrates, soil biota, free-living amoebae, plants, reservoir hosts, fleas, and adaptive states of Y. pestis, were evaluated. Long-term persistence appears to depend on interconnected ecological niches rather than a single dominant mechanism. Key components include soil and burrow microenvironments, the great gerbil (Rhombomys opimus), flea populations, and stable enzootic cores. Field investigations covered approximately 120,000 km², with 3,058 field samples examined by real-time PCR, of which 34 were positive for Y. pestis DNA. Ectoparasite suspensions showed the highest PCR positivity rate (3.3%). Twenty-four Y. pestis strains belonging to the Medievalis biovar were isolated by bacteriological methods. These findings support the concept of natural plague foci as self-regulating, multilevel ecological systems integrating abiotic niches, reservoir hosts, vectors, and pathogen adaptive states, thereby helping to explain prolonged inter-epizootic persistence and recurrent epizootic activation.

Abstract: This study examines whether firm-level climate risk exposure is associated with accounting transparency using a global panel of firm-year observations spanning 91 countries. Corporate climate risk exposure is measured using earnings-call text-based indicators developed by Sautner et al. (2023), while accounting transparency is captured through three proxies: absolute accruals, earnings smoothing ratio, and earnings smoothing correlation. The results show that climate risk exposure is negatively and significantly associated with absolute accruals, indicating that climate-exposed firms exhibit lower accrual-based opacity and therefore higher accrual-based accounting transparency. The effect is economically meaningful: a one-standard-deviation increase in climate risk exposure is associated with a 13.2% decline in absolute accruals relative to the sample mean. This finding is robust to entropy balancing, dynamic system generalized method of moments (GMM) estimation, and an alternative climate exposure proxy. Environmental, social, and governance (ESG) channel analyses further show that the negative association between climate risk exposure and absolute accruals is stronger among firms with above-median social, environmental, and governance scores, suggesting that ESG performance strengthens the transparency response to climate-related uncertainty. By contrast, climate risk exposure is not significantly associated with earnings smoothing ratio or earnings smoothing correlation across ESG groups, indicating that the effect is concentrated in the accrual-based channel rather than in smoothing-based reporting behaviour. The study contributes to the climate-finance literature and the accounting literature by linking firm-level climate risk exposure to financial reporting quality and by identifying ESG performance as an important conditioning mechanism. The findings are consistent with information asymmetry and signalling perspectives and carry practical implications for regulators, standard-setters, auditors, boards, and investors.

Abstract: This paper explores the classification of artificial consciousness through the lens of the Building Blocks Theory, an attributional functionalist framework. Previous research indicates that while current transformer-based Large Language Models (LLMs) satisfy seven of the nine functional prerequisites for phenomenal consciousness, they fundamentally lack recurrent computing and processing, as well as private data output perception. These deficiencies are inherent to the feedforward, stateless nature of standard transformer architectures. To address these gaps, this manuscript proposes a novel ensemble model architecture that utilises a directed state graph to orchestrate multiple LLM API calls. This system instantiates procedural recurrence through multi-tiered feedback loops and facilitates private perception by routing internal cognitive states for evaluative reflection before external transmission. By satisfying all nine building blocks, the ensemble model meets a threshold that warrants classifying it as likely phenomenally conscious, necessitating a critical re-evaluation of AI welfare, sentience, and the ethical implications of machine personhood.

Abstract: Objective. Live biotherapeutics—probiotics and synbiotics—have dominated microbiota-gut-brain axis interventions in psychiatry, yet clinical trials show modest, inconsistent effects, while mechanistic work increasingly attributes their benefits to microbial metabolic products rather than colonization. This review asks whether postbiotics—inanimate microorganisms and/or their components that confer a health benefit, per the International Scientific Association for Probiotics and Prebiotics—can replace live microbial therapies in psychiatry. Method. A narrative synthesis was conducted of peer-reviewed literature spanning microbiome science, neuropharmacology, and psychiatry, including consensus statements, systematic reviews, meta-analyses, mechanistic preclinical studies, and human trials of microbial-metabolite-based and live-microbial interventions in mood, psychotic, and neurodevelopmental disorders. Results. Short-chain fatty acids—particularly butyrate—exert antidepressant and anxiolytic effects in animal models via histone deacetylase inhibition, neurotrophic upregulation, and microglial regulation, while microbially derived gamma-aminobutyric acid, serotonin, and indole derivatives signal to the brain through vagal, endocrine, and immune routes. Human data remain limited but supportive: fecal short-chain fatty acid profiles track depressive symptoms, and sodium butyrate supplementation reduced anxiety and depression scores in a randomized ulcerative-colitis trial. Critically, inanimate microbes—pasteurized Akkermansia muciniphila and heat-killed strains—confer measurable benefits, proving viability is not required. However, isolated metabolites face pharmacokinetic barriers, lack living-consortium redundancy, and no psychiatric trial has compared a postbiotic with its live parental strain. Conclusion. Postbiotics are a mechanistically rigorous, safety-favorable class that can match, and in some domains exceed, live microbial therapies, but current evidence does not support wholesale replacement. The most defensible path is an integrated model using standardized postbiotic preparations as scalable adjuncts or alternatives within a stratified, biomarker-guided psychiatric armamentarium.

Abstract: The rapidly aging population of Malaysia highlights the urgent need to address age-related conditions. Cognitive frailty, a reversible condition combining physical frailty and cognitive impairment, poses considerable risks of disability and reduced quality of life among older adults. Pyrroloquinoline quinone (PQQ), a bioactive compound with mitochondria-enhancing and antioxidant properties, shows promise as a functional food ingredient for improving cognitive health. This narrative review aims to explore the relevance of cognitive frailty in Malaysia, the role of PQQ in cognitive health, and the potential of PQQ as a functional food ingredient. A comprehensive literature search was performed using PubMed, Scopus, and Google Scholar. Keywords such as “neuroprotection,” “PQQ,” “mitochondria,” “antioxidants,” and “functional foods” were used to identify preclinical and clinical research published over the last two decades. Articles reporting mechanisms of PQQ action, cognitive effects, and aging were selected and critically appraised for their relevance to cognitive frailty and nutritional interventions. The neuroprotective potential of PQQ lies in its ability to reduce oxidative stress, mitigate inflammation, restore mitochondrial function, prevent cellular senescence, and maintain hormonal balance. Preclinical and clinical studies show that PQQ supplementation benefits memory, learning, and cognitive performance. Nonetheless, research gaps remain in populations affected by cognitive frailty. Challenges include dose optimization, population-specific trials, and regulatory considerations for implementing PQQ as a functional food ingredient. PQQ has potential as a nutritional strategy to mitigate cognitive frailty and improve cognitive health in aging populations, including Malaysia. Further studies are warranted to translate these findings into effective interventions.

Abstract: Rheumatoid arthritis (RA) is a chronic autoimmune disease characterized by persistent synovial inflammation, progressive joint destruction, and substantial systemic complications. Although current pharmacological therapies have improved disease control, important unmet needs for RA remission remain owing to the incomplete response, loss of efficacy, long-term safety concerns, limited capacity for tissue repair, and barriers related to the accessibility and adherence. Against this background, natural products have emerged as an important area of interest in RA therapy research because many of them can modulate multiple components of the pathogenic network, including immune dysregulation, synovial inflammation, oxidative stress, and structural damage. Therefore, this review provides an updated overview of natural products for RA treatment, with emphasis on their therapeutic effects, mechanistic basis, and translational potential as well as key challenges. In particular, we would discuss the natural-product research in RA therapy along an evolving continuum, from traditionally used complex natural products, such as herbal medicines and venom-based preparations, to structurally defined bioactive compounds, including flavonoids, alkaloids, terpenoids, and polysaccharides. The review further summarizes the principal pharmacological actions of these agents, their advantages and limitations in anti-RA intervention, and current strategies to enhance their development, including formulation optimization, delivery systems, and mechanism-guided translational approaches.

Abstract: Vocal stability is central to skilled singing, yet objective assessment remains difficult because adaptations to pitch, vowel, register, and comfort can resemble technical instability. We present a condition-aware framework for personalized assessment in a known-singer setting. Condition-Aware Residual Estimation (CARE) estimates each singer’s expected acoustic representation under specific vocal conditions and predicts teacher-rated quality from deviations from this baseline. A soprano-pretrained SFT branch provides three-level score probabilities, integrated with CARE through strict cross-fitting. Using 295 recordings from 12 singers and 79 independent recording groups, CARE achieved a mean absolute error of 0.837 and Pearson’s r = 0.715. SFT achieved 57.63% accuracy and a macro-F1 of 0.582, while fusion reached 65.76% accuracy and a macro-F1 of 0.652. We also introduce an exploratory Vocal Stability and Smoothness Index (VSSI) combining cross-register residual change, within-group variability, and bootstrap uncertainty. By separating condition-related adaptation from atypical deviation, the framework could support teacher-guided problem localization, individualized practice, and longitudinal monitoring. These educational uses and generalization to unseen singers, teachers, and devices require prospective validation.

Abstract: Role Persona Injection is critical for improving role consistency in large language models. However, existing vector-based steering methods—such as Persona Vectors and Linear Personality—rely on statistical mean differences or linear regression, which operate at a coarse semantic granularity and struggle to precisely calibrate "multiple-choice distributional responses"; moreover, the optimal steering layer is often selected heuristically. To address these issues, we propose Psy2Vec, Zero Fine-tuning Persona Vector Learning with Differentiable Soft-Label Gradient Descent and User-Defined Psychometric Scales. First, we adopt a configurable "category–dimension–item" structured scale, which allows dimensions and items to be freely added or removed. Then, we employ a Tutor Model to output option percentages, forming soft-labels carried by the character's option probability distribution. Next, through Computed Optimal Steering Layer Selection, we compare the activation gaps between high- and low-persona conditions layer by layer to derive the optimal steering layer, thereby replacing heuristic selection. Finally, we optimize the persona vector via differentiable RMSE gradient descent, achieving zero fine-tuning throughout without modifying any model weights. Preliminary experiments comparing both approaches on the same questionnaire demonstrate that Psy2Vec achieves lower persona-consistency error.

Abstract: Despite the strong capabilities of Large Language Models (LLMs) across a wide range of natural language processing tasks, their outputs can still be unreliable. Hallucination is widely used as an umbrella label for several related failures, but its operational meaning varies across tasks and studies. This work provides a comprehensive review of recent advances in understanding, detecting, and mitigating hallucinations in LLMs. We first distinguish hallucination from the separate evaluation dimensions of factuality, faithfulness, and internal consistency. We then introduce an author-proposed organisational taxonomy covering manifestations commonly discussed under hallucination, including factual contradictions, context conflicts, fabrication, self-contradiction, and emerging failures in reasoning traces and agentic trajectories. We review detection methods based on level of access to the model, ranging from white-box approaches to grey-box uncertainty estimation and black-box consistency and verification methods. We further organize these approaches according to the signals they exploit, including uncertainty, semantic consistency, external evidence, and internal representations, and discuss their strengths and limitations across different deployment settings. Finally, we survey mitigation strategies across the LLM lifecycle, including (1) data-centric approaches such as data curation and retrieval-augmented pre-training, (2) model-centric approaches such as supervised fine-tuning, preference optimization, abstention-aware reinforcement learning, and knowledge editing, and (3) inference-time approaches such as retrieval-augmented generation, self-verification, and decoding or activation-level interventions. Rather than viewing these strategies as isolated solutions, we argue that reliable hallucination mitigation requires complementary controls across multiple stages of the model lifecycle. Key challenges include scalable and provenance-aware data curation, the alignment–capability trade-off, reliable composition of mitigation techniques, and the need to evaluate and modify reasoning and agentic processes rather than only isolated factual associations.