The Indian technology sector offers a compelling setting for this kind of investigation. Employees in IT and ITeS companies operate in environments defined by rapid technological change, evolving client demands, and constant upskilling. Automation, artificial intelligence, and shifting service models mean that yesterday’s expertise can become obsolete within a few years. In such an atmosphere, whether an employee feels safe enough to voice a half-formed idea, admit uncertainty, or challenge a manager’s plan can determine not only individual wellbeing but also an organisation’s capacity to innovate. Yet, as the authors note, most existing evidence on psychological safety and innovation comes from Western contexts, and relatively little qualitative research has captured how Indian employees actually experience these dynamics.

To address that gap, the researchers adopted a hermeneutic phenomenological approach, a qualitative method designed to uncover the lived experiences of participants and interpret the meanings they attach to those experiences. Rather than measuring psychological safety with survey scales, the study sought depth: what does it feel like to speak up, or stay silent, in a workplace innovation setting? The researchers conducted semi-structured interviews with 20 employees drawn from IT and ITeS organisations located in Chennai and Bengaluru, two of India’s most important technology hubs. Semi-structured interviews allowed participants to describe their experiences in their own words while ensuring that key topics, such as leadership, feedback, and change, were consistently explored across conversations.

The study met rigorous ethical standards. Approval was granted by the Institutional Ethics Committee for Studies on Human Subjects at Vellore Institute of Technology, Chennai Campus, and the research was conducted in line with the principles of the Declaration of Helsinki. All participants provided written informed consent before taking part and consented to the anonymised use of their responses for research and publication. The authors declare no competing interests, and the work was published open access, with funding support provided by Vellore Institute of Technology.

Analysis of the interview data was carried out using NVivo 15, a widely used qualitative data analysis software package. The researchers applied coding and thematic analysis, a systematic process in which interview transcripts are broken down into meaningful segments, labelled with codes, and then grouped into broader themes that capture recurring patterns across participants. From this process, five themes emerged: autonomy, flexibility and empowerment; organisational strategy and change; technology adoption and support; leadership and team culture; and open feedback. Together, these themes map the terrain on which psychological safety is either built or eroded in technology workplaces.

The first theme, autonomy, flexibility and empowerment, points to the importance of giving employees genuine control over how they work. Participants who felt trusted to make decisions, manage their schedules, and experiment with new approaches reported a stronger sense of psychological safety and a greater willingness to contribute ideas. Empowerment, in this sense, is not merely a managerial buzzword but a structural condition: when employees believe their judgement is respected, the perceived risk of proposing something novel drops, and innovation becomes a shared activity rather than a top-down mandate.

The remaining themes highlight the organisational scaffolding around the individual. Organisational strategy and change emerged as a double-edged factor. On one hand, a clear strategic direction helps employees understand why innovation matters and where their contributions fit. On the other, the study found that frequent organisational changes can undermine confidence, leaving employees uncertain about their roles and reluctant to invest energy in new initiatives. Technology adoption and support proved similarly consequential: access to adequate training and technical support made employees feel equipped to engage with new tools, whereas insufficient training left them anxious about being judged for falling behind. Leadership and team culture, along with open feedback, rounded out the picture, with accessible leaders and honest, constructive communication emerging as key ingredients of a climate in which people dare to speak.

The findings paint a coherent picture of what builds and what breaks psychological safety in Indian technology workplaces. Organisational support, leadership accessibility, training opportunities, recognition, and open communication all contributed to employees’ perceptions of psychological safety and their willingness to participate in workplace innovation. Conversely, frequent organisational changes, insufficient training, communication gaps, and fear of negative judgment reduced employees’ confidence to contribute ideas and engage in innovation activities. The fear of negative judgment deserves particular attention: it is the psychological mechanism at the heart of psychological safety theory, first articulated by Harvard Business School professor Amy Edmondson, who showed that teams learn and perform better when members can admit mistakes without penalty. This study extends that logic into the Indian IT and ITeS context, showing that the same interpersonal dynamics operate even in high-throughput, globally connected service environments.

The practical implications are significant for leaders and human resource professionals. If innovation is a strategic priority, the study suggests that organisations should invest in structured training before rolling out new technologies, stabilise communication during periods of change, and cultivate leadership styles that keep managers approachable. Recognition also matters: employees who see their contributions acknowledged are more likely to keep offering them. Human resource functions can embed these insights into onboarding, performance management, and leadership development, treating psychological safety not as a soft perk but as measurable infrastructure for innovation. For a sector that competes on its ability to absorb new technologies faster than its rivals, the message is that human confidence is as much a competitive asset as technical capability.

The study also makes a scholarly contribution by adding qualitative evidence from India to a literature dominated by quantitative Western studies. By capturing employees’ own accounts of what safety feels like in innovation settings, the research complements survey-based findings with texture and nuance, revealing how autonomy, change, technology, leadership, and feedback interweave in daily working life. The authors position the work as a resource for both researchers and practitioners seeking to create supportive work environments that encourage employee participation, learning, and innovation. As artificial intelligence reshapes the global services industry, understanding the human conditions under which employees embrace rather than resist change may prove decisive, and this study offers a grounded starting point for organisations across India’s technology sector and beyond.

Subject of Research: Psychological safety and workplace innovation among employees in Indian IT and ITeS organisations

Article Title: Exploring psychological safety and workplace innovation in Indian IT and ITeS organisations

Article References: Shanmuganathan, S., & G N, S. (2026). Exploring psychological safety and workplace innovation in Indian IT and ITeS organisations. Discover Psychology. https://doi.org/10.1007/s44202-026-00808-2

Image Credits: AI Generated

DOI: 10.1007/s44202-026-00808-2

Keywords: psychological safety, workplace innovation, Indian IT sector, ITeS, leadership, organisational support, qualitative research, hermeneutic phenomenology, employee empowerment, open feedback, training, Chennai and Bengaluru

The research, led by Bridgett Kiernan, Micah O. Mazurek, and Rose E. Nevill at the University of Virginia, examined how 138 first responders from 21 agencies responded to a three-hour in-person course called the Autism Response Protocol. The training was developed through an iterative process that drew on the perspectives of autistic people themselves, caregivers, and first responders, and it combined psychoeducation about autism’s core features with evidence-based communication and de-escalation strategies. Participants completed surveys immediately before and after the training, measuring their confidence, self-efficacy, and knowledge about supporting autistic people in emergencies. The researchers then used linear mixed-effects models to see whether gains differed by professional role, personal relationship to autism, and demographic characteristics such as age, gender, and prior training.

The stakes could hardly be higher. Autistic people face elevated rates of co-occurring physical and mental health conditions, including epilepsy, gastrointestinal disorders, anxiety, and depression, and they are far more likely than the general population to experience suicidal ideation and crises requiring urgent intervention. Roughly one in five autistic children has a police interaction before age 21, and nearly half of autistic adults interact with first responders four or more times in their lifetime. Without adequate understanding of autism, responders may misread communication differences or distress behaviors, escalating situations that could have been calmly resolved. The study’s authors point to sobering real-world examples, including a teenager handcuffed while having a seizure and a 25-year-old autistic man shot by police after calling 911 himself for a wellness check.

What makes the new findings particularly valuable is the framework behind them. Rather than treating first responders as a single bloc, the researchers applied role theory, the idea that socially defined roles shape behavior, expectations, and perceived performance demands. A law enforcement officer securing a scene, a firefighter conducting a rescue, an EMS paramedic triaging a patient, and a 911 dispatcher gathering information all operate within distinct organizational cultures and decision-making structures. Those differences, the theory suggests, could influence how trainees absorb and apply autism-related content. The study was among the first to test this idea empirically across four distinct responder professions, including dispatchers, who are often overlooked in autism training research.

The headline result is strikingly simple: everyone improved. Confidence and self-efficacy scores rose significantly from pre- to post-training across all professional roles, and knowledge scores also increased when relationship type was taken into account. No pairwise comparisons of change scores between professional groups reached statistical significance after adjustment for multiple comparisons. In other words, the training worked about equally well for police officers, firefighters, EMS personnel, and dispatchers, a finding that challenges any assumption that autism education is only needed, or only effective, for the officers who make arrests.

Still, the data revealed meaningful nuances. Law enforcement officers entered the training with significantly higher baseline self-efficacy than their EMS counterparts, about 2.34 points higher on the study’s scale, and consequently showed smaller gains. The authors suggest two plausible explanations. The occupational culture of policing, with its greater perceived autonomy and control over decisions, may reinforce efficacy beliefs in ways that protocol-driven roles do not. Alternatively, officers may simply have had more prior exposure to autism and developmental disability training, often through optional Crisis Intervention Team programs. Either way, the smaller improvement among officers likely reflects a ceiling effect rather than resistance to the training, and it suggests that community efforts to train law enforcement may already be paying lasting dividends.

Personal relationships with autistic people told an equally interesting story. Participants with a close relationship to an autistic person, such as a parent, sibling, partner, or being autistic themselves, scored an average of 8.87 points higher on baseline confidence than those with no personal connection, and those with more distal relationships, such as autistic friends or extended family, also started higher. Yet the group with no personal relationship showed the largest gains in self-efficacy, substantially narrowing the gap by the end of training. This pattern carries a practical warning: departments cannot assume that experienced responders have absorbed autism competence through osmosis, nor can they rely on the informal training that responders with autistic family members often provide to colleagues. Formalized instruction, the authors argue, lifts everyone while relieving that unpaid burden.

One curious discrepancy emerged between subjective and objective measures. Confidence and self-efficacy, which are self-rated, improved significantly, while knowledge scores, measured through an 18-item questionnaire, showed a significant main effect of time only in the analyses examining personal relationships, not professional roles. The authors note this mirrors a pattern seen in earlier EMS training research and in medical education studies suggesting that self-efficacy may be a stronger predictor of real-world clinical practice than knowledge alone. If limited self-efficacy undermines the quality of care autistic people receive, then boosting responders’ belief in their own capability may be the training’s most consequential effect, even when quiz scores move less dramatically.

Demographics mattered too, in ways that carry policy implications. Previous autism training predicted larger gains in confidence and self-efficacy, while age showed a more complicated relationship, with older participants reporting slightly lower confidence change scores in one analysis. Notably, younger and more junior responders tended to start with higher knowledge and self-efficacy, possibly because newer workers are more receptive to training and more likely to have encountered it recently. The authors argue this makes a case for mandatory retraining: many U.S. states now mandate autism training for first responders or law enforcement, but few specify how often it must be refreshed. They propose supplementing annual sessions with brief, frequent refresher courses, an approach shown in occupational safety research to help teams maintain skills year-round.

The study is not without limitations, and the authors are candid about them. All participants came from a single southeastern state, and because first responder training is governed largely by state and local law, the findings may not generalize nationally. The sample was also predominantly White, and all outcome measures relied on self-report Likert scales, which are vulnerable to response bias and may not capture actual applied skill. Future research, the team suggests, should incorporate objective measures such as role-plays, simulated scenarios, or on-the-job evaluations. Even so, the core message stands out with unusual clarity for a training study: a three-hour intervention, co-designed with autistic people and delivered by a team that included a firefighter parent of autistic children and a self-advocate, improved confidence and self-efficacy across every role and every relationship type tested. As autism prevalence continues to rise, ensuring that all first responders, not just police, are trained and retrained may be one of the most direct ways to make emergencies less dangerous for autistic people and their families.

Subject of Research: Effects of autism crisis response training on first responders' confidence, self-efficacy, and knowledge across professional and personal characteristics

Article Title: An Examination of First Responder Gains Following an Autism Training Based on Personal and Professional Characteristics

Article References: Kiernan, B., Mazurek, M. O., & Nevill, R. E. (2026). An Examination of First Responder Gains Following an Autism Training Based on Personal and Professional Characteristics. Journal of Autism and Developmental Disorders. https://doi.org/10.1007/s10803-026-07520-z

Image Credits: AI Generated

DOI: 10.1007/s10803-026-07520-z

Keywords: autism, first responders, training, law enforcement, EMS, self-efficacy, crisis response, role theory, de-escalation, 911 dispatch, firefighters, Journal of Autism and Developmental Disorders

The review, led by Joachim D’Hondt and colleagues, was conducted according to the PRISMA 2020 reporting guidelines and registered in advance on PROSPERO. The researchers searched four major databases, PubMed, Scopus, SPORTDiscus and Web of Science, identifying 6,804 records, which reduced to 4,811 unique articles after duplicates were removed. Rigorous title, abstract and full-text screening left just ten studies that met the inclusion criteria: original, peer-reviewed investigations of lower inter-limb asymmetry and sprinting performance in healthy, injury-free sprint athletes. Methodological quality was assessed with a modified Downs and Black Quality Index Tool, and the certainty of the evidence was graded using the GRADE framework. The studies, published between 2013 and 2025, came from Europe, North America, Oceania and Asia, and together covered 684 participants, including sixteen world-class sprinters, eight elite sprinters and 138 highly trained athletes.

Inter-limb asymmetry is not a single phenomenon but a family of them. The review distinguishes functional asymmetries, such as differences in strength or jump performance between legs; morphological asymmetries, such as differences in limb length, joint width or muscle cross-sectional area; kinematic asymmetries, involving discrepancies in joint angles, step length or contact time; and kinetic asymmetries, meaning imbalances in ground reaction forces or impulse production. Each type was measured with different tools across the ten studies, from single-leg countermovement jumps and magnetic resonance imaging to textile electromyography embedded in shorts and force measurements during actual sprinting. Complicating matters further, the studies used at least five different equations to calculate asymmetry, including the symmetry angle, percentage differences between superior and inferior limbs, fluctuating asymmetry indices and composite scores.

The headline finding is stark in its simplicity: in general, only limited associations were observed between any of these asymmetry types and sprinting performance. Prvulović and colleagues found no significant link between explosive strength asymmetry, measured with a single-leg countermovement jump, and mean velocity in a 100-metre sprint. Tottori and colleagues reported no significant associations between asymmetry in the cross-sectional area of the psoas major, quadriceps femoris or hamstrings and curve-sprinting time, with the single exception of psoas major asymmetry, which showed a small negative association with cross-directional sprint time differences. Nagahara and Gleadhill found no significant relationships between kinetic asymmetry metrics such as braking and propulsive impulse and maximal sprinting velocity over 60 metres.

Yet the picture is not uniformly null, and the exceptions are scientifically revealing. Gołaś and colleagues, studying eight elite female sprinters with textile EMG sensors, found weak-to-moderate positive associations between sprinting time and EMG asymmetry in the quadriceps, hamstrings and gluteal muscles, meaning greater asymmetry accompanied slower times. Bissas and colleagues reported that in world-class male sprinters, 100-metre time was strongly negatively associated with foot vertical velocity pre-touchdown asymmetry, while in world-class females, shank angle touchdown asymmetry was positively associated with sprint time, a striking sex-specific pattern. Exell and colleagues even found that net ankle work asymmetry and maximum vertical force asymmetry were positively associated with mean 60-metre velocity, suggesting that in some athletes, asymmetry and speed coexist, or that the dominant limb compensates for the weaker one during propulsive phases.

Perhaps the most famous data point in this debate comes from analyses of Usain Bolt himself. Previous biomechanical investigations showed that the fastest man in history ran with pronounced asymmetries: his right leg produced approximately 13 percent greater peak force, while his left leg recorded ground-contact times roughly 14 percent longer, alongside notable step-length discrepancies. If perfect symmetry were a prerequisite for world-record sprinting, Bolt’s mechanics would have made it impossible. The review’s authors point to computer simulation studies suggesting a plausible mechanism: athletes may adopt compensation strategies in which the dominant limb exerts greater force during propulsive phases to offset the weaker limb, preserving overall output despite the imbalance.

The contrast with team-sport research is equally instructive. Earlier work, including a meta-analysis by Fox and colleagues, reported a small but significant negative association between lower-limb asymmetry and sprinting performance, with correlations around 0.20. But that evidence base drew overwhelmingly from unilateral team sports such as soccer, where eight of nine studies involved players whose repetitive kicking and cutting patterns naturally create larger asymmetries. In one striking example, single-leg countermovement jump asymmetry in English Premier League academy players showed strong positive correlations with 5, 10 and 20-metre sprint times, and asymmetry magnitudes in such cohorts have reached 18 percent. Sprint athletes, by contrast, tend to display much smaller asymmetries, often in the low single digits, which may simply fall below any threshold at which performance would be affected.

The review also delivers a methodological caution that should reshape how future studies are designed. The certainty of evidence across all outcomes was rated as very low, reflecting methodological limitations, imprecision and the observational nature of the studies. Only three of the ten studies reported reliability data for their asymmetry measures, a serious gap given that asymmetry is a ratio metric highly sensitive to measurement error in either limb. The authors also argue against composite asymmetry scores, which aggregate different body regions and metrics into a single number; such scores can conflate large asymmetries in a few variables with small asymmetries across many, obscuring meaningful patterns. Because asymmetry is known to be task-, metric- and segment-specific, assessing each type independently is considered far more valid and informative.

For coaches and athletes, the practical message is one of cautious reassurance. The evidence synthesized here does not support the assertion that inter-limb asymmetry is associated with reduced sprinting performance, and the authors conclude that reducing asymmetry does not necessarily need to be a central focus of training when the goal is speed. Small fluctuations in asymmetry over time, as shown in the one longitudinal study included, likely reflect normal neuromuscular variability rather than mechanical inefficiency. That said, asymmetry reduction may still be worthwhile for other reasons, such as injury prevention, and targeted strengthening of the weaker limb has been shown to increase its capacity and reduce side-to-side differences. If practitioners do choose to address asymmetry, the evidence suggests they should do so with task-specific, metric-specific and muscle-specific assessments rather than blanket correction programs.

What remains genuinely unresolved is the question of cause and effect. With nine of the ten studies using cross-sectional designs, it is impossible to say whether asymmetry shapes sprint performance, whether sprinting shapes asymmetry, or whether both simply reflect each athlete’s unique neuromuscular signature. Exell and colleagues observed that two athletes with comparable kinetic asymmetries produced markedly different step velocities, hinting that sensitivity to asymmetry varies from person to person, with some athletes compensating effectively and others not. The review’s authors call for standardized asymmetry calculations, robust longitudinal designs, and the establishment of task-, metric- and muscle-specific thresholds that account for individual responders and non-responders. Until then, the sprinter’s lopsided stride, once viewed as a flaw to be engineered away, may deserve to be seen for what the evidence now suggests it often is: a harmless quirk of a body built for speed.

Subject of Research: The relationship between inter-limb asymmetry and sprinting performance in sprint athletes

Article Title: The Relationship Between Inter-limb Asymmetry and Sprinting Performance in Sprint Athletes: A Systematic Review

Article References: D’Hondt, J., Chapelle, L., Van Handenhoven, J., Stalmans, J., & Aerenhouts, D. (2026). The Relationship Between Inter-limb Asymmetry and Sprinting Performance in Sprint Athletes: A Systematic Review. Sports Medicine – Open, 12(1), Article 146. https://doi.org/10.1186/s40798-026-01082-1

Image Credits: AI Generated

DOI: 10.1186/s40798-026-01082-1

Keywords: inter-limb asymmetry, sprinting, biomechanics, sports medicine, systematic review, kinematics, kinetics, muscle morphology, athletic performance, sprint athletes, neuromuscular, training

The observation that synchronous remote and face-to-face theoretical instruction produced comparable knowledge acquisition is consistent with a substantial body of literature from medical and health professions education. The so-called no significant difference phenomenon, first documented decades ago in distance education research, has repeatedly shown that when content, pacing, instructor quality, and learner engagement are held constant, the delivery medium itself rarely determines learning outcomes for declarative knowledge. What this study adds is a veterinary-specific and surgically contextualized confirmation of that pattern, using a standardized one-hour lesson on bovine tenotomy of the deep digital flexor tendon delivered to novice learners with no prior surgical experience. The randomization, the use of identical slides, images, and instructional guidance across groups, and the controlled classroom environment all strengthen the internal validity of this comparison and address common confounders in modality research, such as instructor variability and content drift between formats.

Equally important is what the study reveals about the limits of theoretical preparation alone. In both groups, initial practical performance was strikingly limited, with few students able to complete the procedure correctly on a first attempt. This result should not be read as a failure of the students or of the instruction, but rather as evidence about the nature of surgical skill itself. Procedural competence depends on the integration of conceptual understanding with visuospatial perception, instrument handling, tissue interaction, and real-time corrective feedback, capacities that cannot be fully encoded through verbal and visual description alone. Cognitive load theory offers a useful framework here: novices confronted with a complex novel motor task experience intrinsic load that overwhelms working memory, and textual or pictorial prelearning, while necessary, is insufficient to scaffold execution without live modeling and guided practice. The significant improvement observed on the second attempt, following an in-person instructor-led demonstration, illustrates the well-established pedagogical principle that observation of expert modeling followed by deliberate, repeated practice with feedback is a far more potent driver of early skill acquisition than preprocedural instruction by itself.

The design of the study deserves attention because it isolates a specific and practically relevant question. Rather than comparing entire curricula or long-term outcomes, the investigators examined a single instructional sequence: theoretical lesson, first attempt, demonstration, second attempt. This granularity matters for curriculum planners because regulatory frameworks operate at the level of workload distribution, while learning happens at the level of sequenced instructional events. The evidence suggests that remote delivery can occupy a legitimate place in that sequence as the vehicle for conceptual preparation, provided that it is followed by supervised in-person experiences that include demonstration, repetition, and feedback. The comparable improvement between groups during the second attempt further suggests that the benefits of in-person demonstration were not modulated by how the preceding theoretical content had been delivered, reinforcing the idea that remote and in-person elements can be complementary rather than competing.

It is worth situating these results within the broader literature on simulation and skills training in surgical education across species and professions. Research in human surgical training has repeatedly demonstrated that theoretical knowledge and technical skill are only weakly correlated, and that structured skills laboratories, model-based simulation, and progressive mastery learning outperform knowledge-focused preparation for the development of manual competence. Veterinary education faces additional challenges, including the heterogeneity of patient size and anatomy, the ethical and economic constraints on animal use for teaching, and the limited availability of cadaveric and clinical material. In this context, the tenotomy model used in the study is instructive: it is a fundamental procedure that integrates core elements of surgical technique, including aseptic preparation, tissue handling, hemostasis, and closure, while being feasible to teach in a controlled laboratory setting. Findings from such standardized procedures can plausibly generalize to other basic surgical skills, though extrapolation to more complex procedures, live animal surgery, or clinical decision-making under uncertainty should be made cautiously.

The choice of a bovine model also carries practical significance for the Brazilian context, where large animal practice represents a substantial share of professional activity and where curricular reforms affect a large and diverse student population. Third-semester students represent the earliest stage at which surgical exposure typically occurs, making them an appropriate population for studying initial performance rather than refined competence. The restricted sample size, justified by a power analysis based on a large anticipated effect size, is adequate for detecting major differences between groups but limits the precision of estimates and the ability to detect smaller effects or interactions. Readers should therefore interpret the absence of significant differences as an absence of evidence for a modality effect under these conditions, not as definitive proof of equivalence. Replication with larger cohorts, multiple institutions, additional procedures, and longer follow-up would strengthen the evidentiary basis for curricular policy.

The Brazilian regulatory developments that motivated this study reflect a global trend toward flexibility in professional education, driven in part by expanded digital infrastructure, pandemic-era experience with emergency remote teaching, and economic pressures on higher education. However, the study’s conclusions offer a measured counterweight to purely administrative interpretations of that flexibility. If semipresential formats allow theoretical coursework to move online, the implication is not that practical training can be proportionally reduced, but rather that in-person time should be concentrated and protected for the activities that demonstrably require it: demonstration, supervised practice, feedback, and progressive refinement. This reframing could actually improve curricula by making explicit the distinction between knowledge transmission, which digital tools handle well, and skill formation, which depends on physical supervision and repeated deliberate practice. The danger identified in the study’s conclusions, that curricular expansion of non-presential components might erode the quantity and quality of hands-on training, is a real one, and the empirical finding that novices perform poorly without live demonstration provides concrete support for maintaining robust practical components.

The study also highlights the importance of instructional design quality in remote delivery. The theoretical lesson was not a passive lecture but a structured session with standardized visual and verbal guidance, delivered synchronously so that students could interact in real time. Synchronous formats preserve many of the pedagogical affordances of the classroom, including immediate clarification of doubts and instructor awareness of student comprehension, which may explain why remote delivery performed comparably to in-person teaching for knowledge outcomes. This distinction between synchronous and asynchronous digital learning is often blurred in public debate, yet it is likely to be critical in professional programs. Asynchronous content may serve well for review, preparation, and reinforcement, but the real-time guidance examined here more closely mirrors traditional teaching and may be the more appropriate substitute within a blended sequence.

From a methodological standpoint, the use of a standardized assessment before and after demonstration provides a within-subject dimension that strengthens the study beyond a simple between-group comparison. The significant within-group improvement from first to second attempt demonstrates the sensitivity of the practical scoring instrument to genuine performance change, which supports the validity of the between-group null result. Future work could extend this design by adding further practice trials, delayed retention testing, and transfer tasks, which would reveal whether the demonstrated gains persist and generalize. Incorporating validated measures of technical skill, such as structured checklists and global rating scales completed by blinded assessors, as well as objective motion analysis where feasible, would further enrich the evidence base.

Ultimately, the contribution of this study lies less in resolving the debate about remote education than in sharpening its terms. It suggests that the productive question for veterinary curricula is not whether digital delivery is superior or inferior to face-to-face teaching in the abstract, but which learning objectives, at which stages of training, and within which instructional sequences, each modality should carry. For theoretical foundations, the evidence supports the legitimacy of synchronous remote instruction under structured conditions. For initial technical execution among novices, the evidence affirms the indispensability of live modeling and supervised practice. As Brazilian veterinary programs adapt to the new regulatory landscape, studies of this kind, conducted with randomized designs, standardized procedures, and transparent reporting, provide the empirical scaffolding needed to design blended curricula that expand access and flexibility without compromising the practical competence that the profession, and the animals under its care, fundamentally require.

Subject of Research: Initial surgical performance in veterinary training: theoretical instruction modality and practical outcomes before and after instructor-led demonstration

Article Title: Initial surgical performance in veterinary training: theoretical instruction modality and practical outcomes before and after instructor-led demonstration

Article References: Paredes, S. L., Callegarette, N. C., Mendes, R. P., Rocha, P. S., Vallejo, K. A. G., & Correa, R. R. (2026). Initial surgical performance in veterinary training: theoretical instruction modality and practical outcomes before and after instructor-led demonstration. Global Surgical Education – Journal of the Association for Surgical Education, 5(1), Article 175. https://doi.org/10.1007/s44186-026-00582-8

Image Credits: AI Generated

DOI: 10.1007/s44186-026-00582-8

Keywords: Initial, surgical, performance, veterinary, training, theoretical, instruction, modality, practical, outcomes, instructor-led, demonstration

The scale of the challenge is illustrated by the training of LLaMA-3, which the survey reports took about 54 days using 16,000 H100-80GB GPUs on Meta’s production cluster. Even at that scale, the system reached a Model FLOPs Utilization, or MFU, of only about 38 to 41 percent. MFU measures how effectively available floating-point computing capacity is used during training; the unused portion can reflect communication delays, memory bottlenecks, synchronization, uneven workloads or inefficient kernels. The figures make clear why simply adding more processors does not automatically produce proportional gains. Every accelerator must receive data, exchange intermediate results or gradients and remain synchronized with its peers. A slow link, overloaded storage service or failed device can leave many other processors waiting.

Most of the models covered by the review use decoder-only Transformer architectures. Text is converted into tokens and then vectors, supplied with positional information and processed through repeated layers containing attention and feed-forward blocks. Attention calculates relationships between tokens by transforming them into query, key and value tensors and applying a weighted operation based on their similarities. In its conventional form, self-attention has computational and memory demands that grow quadratically with sequence length, making long contexts especially expensive. Architectural changes such as multi-query attention, grouped-query attention and multi-latent attention reduce some key-value storage pressures, while Mixture-of-Experts designs activate only a subset of feed-forward experts for each input. These model-level choices directly affect the hardware, communication patterns and memory strategies required for training.

The survey describes distributed training as a combination of complementary forms of parallelism. Data parallelism assigns different portions of a batch to different devices, then uses collective operations to aggregate gradients. Tensor parallelism divides the matrices inside individual layers and is generally best suited to tightly connected accelerators because intermediate activations must be exchanged frequently. Pipeline parallelism places consecutive groups of layers on different devices and passes activations between stages, making it useful across nodes with lower bandwidth but introducing idle “pipeline bubbles” while stages fill and drain. Sequence parallelism divides long token sequences among devices, reducing per-device activation and attention costs. In practice, these methods are combined into hybrid plans, often called three-dimensional parallelism when data, tensor and pipeline parallelism are used together.

Choosing that combination is a difficult systems problem because every benefit carries a cost. Fully replicated data parallelism is comparatively simple but stores duplicate model states on every device. Fully sharded approaches such as ZeRO-3 and Fully Sharded Data Parallelism distribute parameters, gradients and optimizer states across devices, sharply reducing memory use while increasing communication. Hybrid sharding provides a middle ground by replicating states within smaller groups and sharding them across those groups. Pipeline schedules attempt to reduce idle time by interleaving forward and backward computations, but more active micro-batches can create memory imbalance between stages. Automated parallelism systems address the complexity by searching possible partitions, estimating execution and communication costs and selecting a plan for a particular model and cluster. The review highlights approaches based on dynamic programming, simulators, reinforcement learning, Monte Carlo search and constraint-guided planning.

Hardware and networks form the physical foundation of these strategies. GPUs are well matched to the matrix and vector operations that dominate Transformer training and increasingly support mixed numerical formats, including FP16, BF16 and FP8. Their high-bandwidth memory and specialized tensor-processing units allow large matrix operations to run in parallel, while interconnects such as NVLink and NVSwitch provide faster communication than conventional PCI Express within a server. Other accelerator ecosystems, including AMD GPUs, Google TPUs, Intel Gaudi processors, Graphcore IPUs and Cerebras wafer-scale systems, offer different combinations of memory, compute capacity and software support. The survey emphasizes that software portability remains important: a strategy optimized for one accelerator and programming environment may require substantial adaptation to run efficiently on another.

Communication can become the dominant expense in a distributed job. Gradient synchronization creates periodic bursts of what network engineers call elephant flows, while tensor and expert parallelism can generate intensive all-to-all exchanges. Remote Direct Memory Access allows one machine to access memory on another without involving the operating system, and GPU-direct variants can move data between GPUs across nodes while bypassing the CPU. InfiniBand provides a dedicated high-performance fabric, whereas RDMA over Converged Ethernet brings similar capabilities to Ethernet-based data centers. Network layouts are increasingly designed around training traffic rather than general-purpose workloads. Rail-optimized architectures group corresponding GPUs through carefully arranged switches, while other designs use sparse connectivity, optical circuit switching or topology-aware routing. Load-balancing methods split large transfers across multiple paths, and congestion-control schemes regulate traffic to limit stalls and packet loss.

Storage must also be treated as part of the training engine. A 70-billion-parameter model can produce a checkpoint of about 980 gigabytes, according to the survey, and thousands of accelerators may need to save or reload such states together. Distributed file systems and object stores therefore need high write and read bandwidth as well as consistency, availability, protection and security. Training data presents a different challenge: LLaMA 3 was trained on more than 15 trillion tokens, while crawling, filtering and preparing web-scale data can involve volumes far larger than the final dataset. Caches such as Alluxio and JuiceFS can prefetch data from slower storage, keeping accelerators supplied even when each token is normally read only once. Cluster schedulers must coordinate these demands with GPU, CPU, memory and network resources, while balancing fairness, utilization, job packing, adaptive scaling and energy consumption across multiple users and workloads.

Memory optimization is central because training stores far more than the model’s visible parameters. Mixed-precision training requires parameters and gradients, higher-precision copies for stable optimizer updates, momentum and variance states, activations from the forward pass, temporary communication buffers and allocator space. For a model with Phi parameters, the survey estimates that model states alone can require 16Phi bytes under a common mixed-precision arrangement before activations and other buffers are counted. Activation recomputation reduces the peak by discarding selected intermediate tensors during the forward pass and rebuilding them during backpropagation, trading memory for additional computation. Sharding removes duplicate states, while CPU and NVMe offloading supplements limited GPU memory. Defragmentation techniques use improved allocation policies or virtual-memory stitching to turn scattered free regions into usable capacity. The goal is not merely to fit a model, but to prevent memory management from idling the entire cluster.

At the computation level, the review points to a shift from treating training as a sequence of generic operators toward designing complete dataflows for specific hardware. FlashAttention reduces high-bandwidth-memory traffic by processing attention in tiles, retaining intermediate results in faster on-chip memory and fusing matrix multiplication, softmax and related operations into efficient kernels without changing the exact attention result. Later variants improve parallel scheduling, asynchronous data movement and low-precision execution. Compilers such as TVM, Triton and TorchInductor can generate kernels, fuse operators and exploit memory locality across broader graphs, reducing the need for every optimization to be hand-written. Mixed precision extends the same principle: BF16 and FP16 are widely used, while FP8, INT8, INT4 and even one-bit or ternary representations are being investigated. These formats can reduce computation, storage and communication, but they require scaling methods, higher-precision accumulators or latent weights to prevent numerical errors from undermining training.

The reliability problem grows with both cluster size and training duration. A failure may come from a GPU, a network link, storage, software or a less obvious straggler that slows progress without producing an immediate crash. Because distributed training is often synchronous, a single stalled participant can leave thousands of others idle. Checkpointing provides a recovery point, but writing and restoring enormous states can itself interrupt useful work. The survey therefore treats anomaly detection, rapid fault diagnosis, resilient communication and automatic recovery as core components rather than afterthoughts. It also distinguishes the needs of pretraining from those of later alignment and reinforcement learning, where actor, critic, reward and reference models may alternate between inference and optimization. Newer asynchronous systems stream generated trajectories to trainers instead of waiting for the slowest rollout, while agentic training adds tool calls, sandboxes, verifiers and long-horizon interactions. The review’s broad conclusion is that future progress will depend on co-design: accelerators, memory, networks, storage, schedulers, compilers and learning algorithms must be optimized as one evolving system, with optical computing and optical networks among the possible technologies for overcoming the limits of conventional digital hardware.

Subject of Research: Distributed infrastructure for efficient large language model training

Article Title: Efficient training of large language models on distributed infrastructures: a survey

Article References: Duan, J., Zhang, S., Wang, Z., Jiang, L., Qu, W., Hu, Q., Wang, G., Weng, Q., Yan, H., Zhang, X., Qiu, X., Lin, D., Wen, Y., Jin, X., Zhang, T., & Sun, P. (2026). Efficient training of large language models on distributed infrastructures: a survey. Vicinagearth, 3(1), Article 9. https://doi.org/10.1007/s44336-026-00038-z

Image Credits: AI Generated

DOI: 10.1007/s44336-026-00038-z

Keywords: large language models, distributed training, GPU clusters, parallelism, AI accelerators, high-performance networking, memory optimization, mixed precision, fault tolerance, Efficient, training, large