Over-Concentration in the AI Conference Ecosystem: What Have We Lost?

A Question from the Research Community

Recently, a researcher posted a topic on the Reddit machine learning community that resonated widely: "Does anyone else miss the old conference ecosystem?"

The poster recalled how conferences like BMVC, ACCV, FG, ICIP, and ICASSP once had vibrant, thriving communities. FG was the central hub for face analysis, ICASSP was the stronghold for signal processing, and BMVC and ACCV regularly produced high-quality papers. But today, all of this is quietly changing — attention and top-tier submissions are concentrating heavily around a handful of flagship conferences.

The question may sound nostalgic, but it touches on a structural shift in today's AI academic ecosystem — one that deserves serious examination. Is this mere nostalgia, or has the research ecosystem genuinely become too concentrated?

From "A Hundred Schools of Thought" to "Winner Takes All"

The Old Era of Vertical Communities

Before the deep learning wave swept everything aside, fields like computer vision, signal processing, and pattern recognition each developed independent methodological traditions. This fragmentation naturally gave rise to a vertical conference ecosystem. Academic conferences exhibited a clear "vertical division of labor" — different conferences served different subfields, forming specialized communities with distinct characteristics:

FG (Face and Gesture): A dedicated stage for face and gesture analysis. Notably, FG attracted not only computer vision researchers but also those from psychology and human-computer interaction, creating a unique interdisciplinary community.
ICASSP: Founded in 1976 as the flagship conference of the IEEE Signal Processing Society, it long focused on highly specialized topics such as Fourier transforms, filter design, and speech coding — a definitive authority in acoustics, speech, and signal processing.
BMVC / ACCV: BMVC has been hosted by the British Machine Vision Association since 1990, known for its rigorous review standards and long regarded as a benchmark for academic quality among regional conferences. ACCV served as an important academic window for computer vision research in the Asia-Pacific region.
ICIP: A major platform in image processing, long serving as the core community for low-level image algorithms and applied research.

These conferences were mid-sized, with community members who knew each other well, and reviewers who were genuine domain experts. This "vertical division" was, at its core, a product of an era with clear disciplinary boundaries — each field had its own benchmark datasets, evaluation metrics, and methodological traditions. Researchers could find highly focused peers here and engage in deep technical exchange. A paper that didn't make it into a top-tier venue could still receive the attention and citations it deserved at an appropriate vertical conference.

Today's Flagship Conference Monopoly

The monopoly of flagship conferences didn't emerge overnight — it was the result of multiple converging forces. The breakthrough performance of AlexNet at the 2012 ImageNet competition marked deep learning's emergence as a universal cross-domain tool. When the same neural network architecture could handle image classification, speech recognition, and signal analysis simultaneously, research communities that had previously been clearly separated began converging onto the same methodological platform.

At the same time, academic incentive structures reinforced this trend: faculty hiring, research grant applications, and student job searches all became increasingly dependent on publication records at flagship venues like CVPR, NeurIPS, and ICML. This "conference brand premium" created a classic Matthew effect — top conferences attracted the best submissions, the best submissions further enhanced the reputation of top conferences, creating a self-reinforcing positive feedback loop. Statistically, CVPR's annual submission count climbed from under 2,000 around 2012 to over 9,000 in 2023, while acceptance rates dropped from roughly 30% to around 25%, with absolute competition intensity rising sharply.

Today, academic resources and attention are rapidly consolidating around a small number of flagship conferences: NeurIPS, ICML, ICLR, CVPR, ICCV, and a few others. This concentration has generated a series of obvious problems:

Explosive submission growth: Top conferences routinely receive tens of thousands of submissions
Severely limited acceptance capacity: Physical and review resources cannot expand at the same pace
Uneven review quality: Reviewers are assigned large volumes of papers, making deep, careful reviewing nearly impossible

This "winner-takes-all" dynamic is fundamentally changing how research findings are disseminated.

Are Quality Papers "Disappearing"?

The original poster raised a particularly alarming observation: in this kind of ecosystem, how many good papers end up as non-archival submissions, posted only to arXiv, or never publicly shared at all?

arXiv (pronounced "archive") was founded in 1991, originally serving the physics community before gradually expanding into computer science and other fields. Its cs.CV (computer vision) and cs.LG (machine learning) sections began explosive growth in the mid-2010s, and today hundreds of new papers are added daily. arXiv broke the constraints that traditional conference cycles imposed on knowledge dissemination — researchers could immediately make their work public while submitting elsewhere, staking an early claim to academic priority. However, this model of "bypassing peer review" also raises concerns: without a unified quality filter, readers must assess reliability on their own; and as arXiv becomes the de facto primary dissemination channel, the core value of traditional conference peer review is partially undermined, while alternative quality assessment mechanisms have yet to mature.

This concern is far from unfounded. When everyone rushes toward a handful of conferences, three crises emerge:

First, the randomness of peer review increases significantly. The academic peer review system faces unprecedented pressure. At NeurIPS 2023, for instance, the conference received over 12,000 submissions, requiring organizers to mobilize thousands of reviewers, each handling an average of more than six papers. Multiple empirical studies — including the well-known NeurIPS 2014 "consistency experiment" — have shown significant randomness in top-conference acceptance decisions. The same paper, when reviewed by different panels, can yield completely different outcomes, with estimates suggesting that roughly 50% of accepted papers might be rejected upon re-review. Massive submission volumes combined with reviewer overload make review outcomes highly variable, leading more and more researchers to simply post their work on arXiv rather than go through formal peer review.

Second, fragmentation in the dissemination of results worsens. After the decline of vertical conferences, researchers in specific subfields lost a centralized forum for focused exchange. Work that would once have found a precise audience at FG or ICASSP is now scattered across a vast sea of preprints, easily overlooked.

Third, the credibility of peer review is diluted. When peer review starts to feel like a "lottery," its authority as a quality-gating mechanism diminishes — which in turn drives more researchers away from formal publication channels, creating a vicious cycle.

The Disappearance of Focused Communities: A Deeper Cost

Beyond the issue of paper dissemination, the original poster also specifically mentioned missing "focused communities" — a point that touches on deeper values within the academic ecosystem.

From a sociology of science perspective, tight-knit professional communities are not merely venues for publishing papers; they are also carriers of tacit knowledge. In mid-sized vertical conferences, informal exchanges between senior researchers and newcomers — corridor conversations, poster-session discussions, dinner debates — serve a knowledge-transmission function that textbooks and papers cannot fulfill: common experimental pitfalls, established evaluation conventions, and research intuitions not yet explicitly articulated all circulate and transfer through this kind of high-density professional interaction.

The value of vertical conferences lies not only in publication, but in the close-knit professional communities they nurture:

Researchers can find truly knowledgeable peers and engage in substantive dialogue
Newcomers can grow in a relatively supportive environment and receive effective feedback
Knowledge transmission and methodological evolution within a field have a stable carrier
Cross-institutional, cross-generational collaborative relationships can be sustained over time

Beyond this, vertical communities also serve an important "boundary-guarding" function — they maintain awareness of a field's core questions and prevent research directions from drifting excessively toward whatever is currently trending. The FG community's sustained attention to the ethics of face analysis, and the ICASSP community's commitment to signal interpretability, are testaments to the independent academic value of vertical communities.

When everything gets drawn into a few mega flagship conferences, this sense of intimate community is invisibly diluted. At a conference with tens of thousands of attendees, it actually becomes harder to find your own "tribe." Growth in scale comes, to some extent, at the cost of community cohesion.

Nostalgia, or a Real Structural Problem?

Returning to the original poster's question: is this just nostalgia, or has the AI academic ecosystem genuinely gone wrong?

Both elements are present, but the latter carries more weight.

On the positive side, the concentration of the academic ecosystem is not entirely without reason. The Matthew effect of flagship conferences partly reflects an underlying trend toward disciplinary convergence — as deep learning became a universal methodology, the boundaries between vision, speech, and signal processing grew increasingly blurry, and researchers naturally gravitated toward platforms that could reach the widest audience. The rise of arXiv has also genuinely accelerated the immediate transmission of knowledge and lowered information barriers.

But from a structural risk perspective, the costs of over-concentration cannot be ignored. The overloading of the review system, the fragmentation of result dissemination, and the decline of vertical communities are all continuously undermining the health and diversity of the academic ecosystem. When the entire field's evaluation system bets on a handful of conferences, the system's fragility increases accordingly.

Possible Paths Forward

Faced with this dilemma, the academic community is already exploring several responses:

Improving tiering mechanisms. Some top conferences have begun introducing more flexible acceptance categories, or forming layered systems with workshops and journals, to provide appropriate outlets for work at different stages of maturity.

Repositioning vertical conferences. Rather than competing head-to-head with flagship conferences on submission volume, specialized conferences might strengthen their community character and the value they offer for deep exchange — becoming a "spiritual home" for researchers in particular fields.

Driving peer review reform. The OpenReview platform has promoted open review practices, making review comments publicly visible. The "Desk Reject" system allows area chairs to directly reject submissions that clearly fail to meet standards. Some conferences have begun piloting "Rolling Review" mechanisms, decoupling the submission cycle from conference dates to smooth out review load. All of these approaches aim to ease the structural pressure on review systems.

Acknowledging the reality of preprint culture. Since arXiv has become the de facto primary dissemination channel, the academic community may need to develop more mature mechanisms for evaluating and recognizing research that has not gone through traditional peer review.

Conclusion

The Reddit user's question gave voice to a sentiment shared by many AI researchers. The concentration of the academic conference ecosystem is a direct product of this field's explosive growth — it has brought efficiency gains, but also a loss of diversity.

The nostalgia for mid-sized, tightly-knit vertical conferences is not merely an emotional sentiment. It is a genuine call for a healthier, more diverse, and more sustainable research ecosystem. How to find a new balance between concentration and decentralization, between efficiency and diversity, is a core challenge that the entire AI academic community will need to face for the long term.