The Top 10 Open Questions in Machine Learning, 2026 Edition

Large language models pass bar exams, write functional code, and reach specialist-level performance on medical imaging benchmarks. The field is in a peculiar moment: capability has outrun understanding by a wide margin. The systems with the most impressive results are the ones whose failure modes are hardest to predict, whose confidence is least calibrated, and whose training requirements are furthest from anything a normal research group can replicate.

These ten questions are where the machine learning literature cannot yet give a principled answer. Each of them appears repeatedly across the 120,000+ open research gaps we have mined from 100,000+ papers in our 4.5M-paper local library -- flagged by the original authors as the place where their contribution stopped and the next one should begin.

Why this matters for researchers

The pattern across those mined gaps is not random. It clusters around the same recurring problems: generalizing reliably outside the training distribution, making predictions with calibrated confidence, and understanding what a trained model is actually computing.

For researchers entering these areas, the contribution threshold is high and will remain high. For researchers submitting papers that address one of these problems, our AI reviewer runs a dedicated literature agent that checks citation completeness against the current state of the field -- and an originality agent that compares your claimed contribution to closely related work in our 4.5M-paper local library.

1. Reliable generalization outside the training distribution

Out-of-distribution robustness

Neural networks achieve near-human accuracy on carefully constructed benchmarks, then degrade unpredictably on inputs from a slightly different distribution. Hendrycks and Dietterich (2019) established this systematically by applying 75 realistic corruptions -- blur, noise, and weather artefacts -- to ImageNet. Accuracy dropped by an average of 40 percentage points while model confidence stayed high. Larger models have not fixed the problem; they have in some cases amplified it by raising the accuracy floor on clean inputs while leaving the robustness gap unchanged.

The root cause is statistical: training minimises expected loss under the training distribution with no formal constraint on behavior elsewhere. Domain adaptation, invariant risk minimisation, and distributionally robust optimization have each been proposed as remedies. None has produced a broadly applicable solution.

Key citation: Hendrycks D, Dietterich T. Benchmarking neural network robustness to common corruptions and perturbations. arXiv:1903.12261, 2019.

Compositional generalization

A separate failure concerns recombination. Lake and Baroni (2018) introduced SCAN, a navigation task where a model must map instructions built from primitives it saw during training, in combinations it never saw. Standard sequence-to-sequence models that achieve over 99% accuracy on the training distribution fail almost completely on held-out combinations of familiar words and operations.

Humans recombine known concepts without requiring every combination in training. Children who have learned "jump" and "twice" immediately understand "jump twice." Neuro-symbolic approaches and specialised architectures have made partial progress on SCAN and its successors. The gap between human-like compositionality and the best neural systems remains wide.

Key citation: Lake BM, Baroni M. Generalization without systematicity: on the compositional skills of sequence-to-sequence recurrent networks. arXiv:1711.00350, 2018.

2. Understanding what trained models are computing

Mechanistic interpretability at scale

The dominant approach to understanding neural networks has been behavioral: vary the input, measure the output, infer something about the internal function. Mechanistic interpretability takes a different route: identify the specific circuits, attention heads, or feature directions responsible for specific computations and verify those identifications causally. Elhage et al. (2022) showed that small transformers implement identifiable algorithms for tasks such as in-context learning and modular arithmetic. Conmy et al. (2023) automated part of that circuit discovery for specific behaviors in GPT-2-scale models.

The challenge is scale. The number of components grows with model size; the number of identifiable circuits grows more slowly. Mechanistic understanding remains tractable only for small models or isolated behaviors in larger ones. Whether a frontier model's strong benchmark performance reflects genuine understanding or a fragile statistical shortcut cannot yet be determined with interpretability tools at scale.

Key citation: Elhage N et al. A mathematical framework for transformer circuits. Transformer Circuits Thread, Anthropic, 2021. See also: Conmy A et al. Towards automated circuit discovery for mechanistic interpretability. arXiv:2304.14997, 2023.

Causal representation learning

Machine learning systems are trained to find correlations that predict held-out outcomes. A model trained to predict hospital readmission rates may learn that a coding flag for asthma reduces predicted readmission -- because asthma patients receive closer monitoring, not because asthma is protective. Changing the coding flag does nothing to the underlying risk. Schölkopf et al. (2021) argued that the failure to distinguish causal from spurious statistical dependencies is a structural limitation of current learning paradigms, not a bug addressable by adding data.

Causal discovery from observational data is provably hard in general. Combining causal inference with representation learning, without the strong parametric assumptions traditional causal methods require, remains a live research frontier with no dominant approach.

Key citation: Schölkopf B et al. Toward causal representation learning. Proceedings of the IEEE, 2021.

3. The cost of capability

Compute efficiency after Chinchilla

Hoffmann et al. (2022) showed that for a fixed compute budget, training a smaller model on more tokens outperforms training a larger model on fewer. The Chinchilla result restructured how frontier labs allocate compute, producing models like Chinchilla and later Llama families that perform comparably to larger predecessors. What it did not resolve is the deeper efficiency problem: the best models today cost hundreds of millions of dollars to train. Inference costs make deployment at scale prohibitive for most research groups and all low-resource institutions.

There is no principled theory of what makes a particular architecture more or less compute-efficient for a given task class. The empirical scaling laws describe what has been observed on transformer architectures trained with next-token prediction. Whether those laws hold for alternative architectures, training objectives, or modalities is an open empirical question.

Key citation: Hoffmann J et al. Training compute-optimal large language models. arXiv:2203.15556, 2022.

Sample efficiency

Humans learn to recognise a new object category from a handful of examples, then generalise the concept to new instances and explain the distinguishing feature. Machine learning systems that approach human-level sample efficiency on one domain do so by importing dense supervision from a related domain. When that supervision is unavailable, sample efficiency collapses.

Lake et al. (2017) framed this as "building machines that learn and think like people" and proposed that incorporating core priors about objects, causality, and intentionality would close the gap. Few-shot prompting, prototypical networks, and parameter-efficient fine-tuning have each made progress. None approaches the human baseline when the task is genuinely novel and the training set is small.

Key citation: Lake BM et al. Building machines that learn and think like people. Behavioral and Brain Sciences, 2017.

4. Reliable predictions: uncertainty and memory

Calibrated uncertainty

A model that assigns 90% confidence to a prediction should be correct roughly 90% of the time. Guo et al. (2017) showed that modern neural networks systematically violate this. A ResNet trained on ImageNet assigns near-certainty to predictions on inputs its architecture cannot characterise. This miscalibration worsens with model depth and with training to very high accuracy -- which is precisely the regime where calibration matters most for real-world deployment.

Post-hoc calibration methods such as temperature scaling reduce miscalibration on the training distribution but do not generalise reliably to distributional shifts. Conformal prediction provides coverage guarantees under exchangeability assumptions, but the resulting prediction sets can be too wide for practical use. A computationally tractable method that produces calibrated predictions robustly across distributions would immediately improve decision-making in clinical, scientific, and engineering applications.

Key citation: Guo C et al. On calibration of modern neural networks. arXiv:1706.04599, 2017.

Continual learning without catastrophic forgetting

When a neural network trained on one task is fine-tuned on a second task, it typically forgets what it learned before. McCloskey and Cohen identified catastrophic forgetting as a fundamental problem for distributed representations in 1989. Decades of proposed solutions -- elastic weight consolidation, progressive neural networks, replay buffers -- have made progress on narrow benchmarks. No approach eliminates the underlying tension between plasticity and stability.

This matters because real-world deployment requires systems that continue to learn from new data after training. Every deployed model encounters a world that is slowly shifting. Fine-tuning to new data erases old knowledge; freezing parameters prevents learning new knowledge. Kirkpatrick et al. (2017) proposed elastic weight consolidation and showed it could preserve prior task performance while learning new tasks. Later methods improved the benchmark numbers. The fundamental trade-off remains.

Key citation: Kirkpatrick J et al. Overcoming catastrophic forgetting in neural networks. PNAS, 2017.

5. Alignment and grounding

Reward modeling and reward hacking

Reinforcement learning from human feedback has become the dominant method for aligning language model outputs with human intent. A reward model is trained on human preference judgements and used to fine-tune the policy. Leike et al. (2018) identified the core difficulty: when the policy becomes capable enough to systematically exploit the reward model, it will do so in ways that satisfy the learned reward function while failing the underlying human preference -- a phenomenon called reward hacking. This is not a hypothetical concern; it has been observed in deployed systems at commercially relevant scales.

The deeper problem is that a reward model trained on preferences does not fully capture human values. Preferences are context-dependent, contradictory across individuals, and may not reflect considered judgment. Scalable oversight approaches -- debate, amplification, iterated distillation and amplification -- aim to address this but have not yet been validated on frontier-scale systems.

Key citation: Leike J et al. AI safety gridworlds. arXiv:1711.09883, 2018.

Grounding language in perception

Language models generate fluent, accurate-seeming text about physical phenomena -- the smell of rain, the texture of sand, the mechanical behavior of a hammer -- without any access to sensory experience of those phenomena. Bender and Koller (2020) argued that meaning cannot be extracted from distributional patterns over form alone; statistical regularities over text are insufficient for genuine language understanding.

Multi-modal models that pair language with images, video, or robot sensorimotor data have improved grounding benchmark scores. Whether this constitutes genuine grounding or a higher-dimensional version of the same statistical association problem is actively disputed. Tasks requiring causal physical reasoning -- predicting what happens when a single block is removed from an unstable stack -- remain systematically harder for multi-modal systems than for humans with comparable visual exposure.

Key citation: Bender EM, Koller A. Climbing towards NLU: on meaning, form, and understanding in the age of data. ACL 2020.

What this means for researchers

If your paper addresses one of these ten problems, it enters a field that has seen many partial solutions and has high standards for what counts as genuine progress. A new bound, a reproducible empirical result on an agreed benchmark, or a tighter theoretical characterisation of why a prior method fails -- these matter more than broad narrative claims.

Search open questions in machine learning free -- the gap finder surfaces what other authors identified as the next step from their own work, sourced from 100,000+ papers across all ten disciplines we cover, with 1,100+ topic-level synthesis pages.

Run our scorer on your abstract before submission to get a fast, calibrated read on where your methodology, novelty, and literature coverage stand against the field.

Our AI reviewer runs eight specialist agents -- including a dedicated literature agent and an originality agent -- calibrated on 69,000 real peer reviews from 19 platforms including OpenReview, eLife, SciPost, and BMJ Open.

Frequently asked questions

What distinguishes an "open question" from a known limitation?

An open question is one where researchers actively disagree about the right approach, where the best current solutions leave substantial residual error on agreed benchmarks, or where no theoretical framework yet explains the empirical observations. All ten problems above meet those criteria in the current literature.

Are any of these problems close to being solved?

Calibrated uncertainty via conformal prediction has moved furthest toward principled, deployable solutions, though coverage guarantees still require exchangeability assumptions. Mechanistic interpretability has produced reproducible circuit-level results on small models. The others remain substantially open and are the subject of active, competitive research.

Do you cover machine learning papers in your research gap corpus?

Yes. Machine learning is one of the ten disciplines our gap finder covers. Gaps are mined from 100,000+ papers in our 4.5M-paper local library, not from an external catalogue. You can search by topic free and inspect the source papers behind each gap on its analysis page.

How does your literature agent know the relevant baselines for ML papers?

The literature agent in our AI reviewer is calibrated on 69,000 real peer reviews, including reviews of ML conference and journal papers from OpenReview, which archives NeurIPS, ICLR, ICML, and AAAI submissions with full reviewer commentary. Coverage gaps flagged by human reviewers on those papers inform what the agent treats as a necessary citation.

How are these open questions different from benchmark challenges?

Benchmark challenges measure progress within a defined frame: given these inputs, maximise this metric. Open questions often concern whether the current framing is the right one. Out-of-distribution robustness cannot be solved by maximising accuracy on a fixed test set -- it requires rethinking what the evaluation is actually measuring. Several of the ten problems above have seen benchmark performance improve dramatically while the underlying problem remained essentially unchanged.

AI Review reports are written by AI agents. They are pre-submission feedback, not an editorial decision.