Research Directions: Open Problems, Benchmarks

The history of language AI has been shaped by a small number of central questions that, once answered, enabled new generations of capability. The question of how to represent words led to distributional semantics and word embeddings. The question of how to handle long-range dependencies led to attention and transformers. The question of how to use unlabeled text at scale led to pretraining and the modern language model paradigm. Each answer brought new problems into focus, and the current frontier is no different.

As we covered throughout this book, modern language models have reached impressive capability levels. They can answer factual questions, write code, translate across dozens of languages, summarize long documents, and reason through multi-step problems. But these achievements sit alongside persistent and well-documented failure modes: models hallucinate facts they assert with confidence, they reason inconsistently when problems are rephrased, they memorize training data rather than learning generalizable rules, and their behavior on deployment can diverge from behavior during evaluation. These gaps are not incidental. They reflect deep unresolved questions about how language models work and what they need to do their jobs reliably.

This chapter maps the open problems, promising research directions, benchmark gaps, and community priorities that define the frontier as of 2026. The goal is not to predict what will happen next but to help you understand what the field is working on, why these problems are hard, and what progress looks like when it occurs. Understanding the shape of the frontier matters as much as understanding the techniques that got us here: the most consequential contributions in the next decade will come from people who know where the gaps are and why they exist.

The following problems represent areas where current methods fall short in ways that matter for real applications. They are not merely performance gaps that scale will close; they reflect architectural, data, or training limitations that require new ideas.

The distinction between "problems that scale will solve" and "problems that require new ideas" is important and often contested. The scaling hypothesis suggests that many apparent limitations of current models are simply the result of insufficient parameters, training data, or compute, and will disappear as these quantities increase. There is empirical support for this view in some domains: models that could not pass grade-school math tests at 7B parameters pass them at 70B, and models that struggled with basic coding tasks at one scale handle significantly harder tasks at a larger scale. But the same scaling that closes some gaps seems to leave others stubbornly open, and understanding which problems fall into which category is itself a major research question.

Large language models are trained to predict text. They are not trained to distinguish between things they know with evidence and things they are confabulating from patterns. The result is hallucination: the generation of plausible-sounding but incorrect information presented with the same surface confidence as accurate information.

Hallucination is not a single phenomenon. Researchers distinguish between several types, each with different characteristics and likely different causes. Factual hallucination produces claims that contradict verifiable facts, such as a model asserting that a historical figure was born in the wrong city or that a scientific paper reached conclusions it did not reach. Faithfulness hallucination produces summaries or answers that contradict the source document provided in context, even when the correct answer is directly stated in the passage the model was given. This second type is particularly troubling because it reveals that models do not always faithfully use context even when it is immediately available. Reasoning hallucination produces conclusions that do not follow from the stated premises, where the individual steps look plausible but the chain of inference contains a subtle error.

Each type likely has different causes and requires different interventions. Factual hallucination may stem from the model's inability to reliably distinguish high-confidence from low-confidence knowledge stored in its weights. Faithfulness hallucination may reflect attention mechanisms that do not always prioritize relevant context over memorized associations. Reasoning hallucination may reflect the absence of an explicit verification step during generation.

Retrieval-augmented generation (RAG), which we covered in detail in the RAG chapter, addresses the knowledge access problem by giving models external memory. But RAG does not fully solve hallucination: models still hallucinate even when the relevant passage is in context. In controlled experiments where researchers confirmed that the correct answer appeared verbatim in the provided context, significant fractions of model responses still contained inaccuracies. The core problem is that the model's generation process does not natively separate "I found this in the provided sources" from "I am generating from prior training." Making this distinction architectural rather than behavioral is an active research area.

A related challenge is calibration: models should express uncertainty in proportion to their actual accuracy. A model that hedges with "I am not certain, but I believe..." when it is likely to be wrong, and states facts without hedging when it is likely to be right, is substantially more useful than one that expresses uniform confidence in everything it says. Current models are often miscalibrated in both directions: they hedge on questions they should answer confidently, and express confidence on questions where they are frequently wrong. Calibration is particularly poor in domains far from the training distribution, in specialized technical domains, and on questions about recent events after the model's knowledge cutoff. The deep challenge here is that a model cannot directly inspect its own confidence; it can only learn to produce calibrated language from training signals that reward appropriate hedging.

Human learners generalize in a particular way: having learned the meaning of "compose a sonnet" and "translate into French," they can execute "translate this sonnet into French" without needing to see that specific combination during training. This compositional generalization, the ability to combine known operations and concepts in novel arrangements, is natural for humans but inconsistently achieved by language models.

Standard benchmarks often evaluate models on distributions similar to their training data. When researchers construct tests that require combining concepts in unfamiliar combinations, model performance drops substantially, even on models that achieve near-human scores on standard evaluations. The SCAN, COGS, and CFQ benchmarks were designed specifically to probe systematic generalization, and they reveal consistent weaknesses. A model might correctly follow the instruction "jump twice" and correctly follow "walk and look," but fail on "jump twice and look," even though this combination requires nothing beyond what the model has already demonstrated.

This failure connects to a deeper question about what language models learn. Do they acquire abstract rules about how concepts combine, the kind of structure that would allow arbitrary recombination of known elements? Or do they learn surface statistics that happen to generalize within the distribution of their training data, but break down when the combination of concepts is new? The evidence suggests the answer is closer to the latter than researchers would like. Models often perform as if they have memorized a very large set of (input, output) mappings and are interpolating between them, rather than applying compositional rules that would generalize to any input expressible in the grammar of the task.

Making compositional generalization reliable requires either fundamentally different architectures that encode compositional structure explicitly, training objectives that specifically reward compositional behavior, or data regimes that force models to learn composition rather than interpolation. None of these directions has yet produced a reliable solution that matches human-level systematic generalization across a broad range of task types.

As we discussed in the reasoning chapters, chain-of-thought prompting substantially improves model performance on problems that require multiple inference steps. But this improvement is brittle in ways that matter practically. Models fail on problems that require small changes to the reasoning chain, commit systematic errors in specific arithmetic and logical domains, and sometimes produce reasoning traces that appear sound but contain subtle errors that standard evaluation does not catch.

The brittleness of chain-of-thought reasoning is well-documented. A model that correctly solves a multi-step word problem may fail on the same problem with different surface features: different numbers, different character names, or a slightly different phrasing that changes the logical structure of the problem without changing its difficulty. This suggests that models are pattern-matching against surface features of common problem types rather than consistently applying the underlying reasoning procedure.

The deepest version of this problem is the gap between process and outcome. A model that produces a correct final answer via a flawed reasoning path is not reliably correct on related problems: the correct answer in that instance may have been reached by coincidence or by the model's statistical priors, not by the stated reasoning. Evaluating the correctness of the reasoning process as well as the final answer, requires either human evaluation, which is expensive and slow, or a separate verifier model, which introduces its own accuracy limitations. Process reward models, covered in the advanced training chapters, are one proposed solution, but they require reasoning-step-level supervision that is costly to obtain at scale.

A further complication is that models can learn to produce reasoning traces that satisfy verifiers without performing the reasoning those traces describe. Just as a student might write out the steps of a mathematical proof while following a memorized template rather than understanding why each step follows, a language model can learn to produce text that looks like valid reasoning without the computation reflecting valid reasoning steps. Distinguishing valid from simulated reasoning is one of the hardest problems in current interpretability research.

Modern language models support context windows of hundreds of thousands of tokens. But having a large context window does not guarantee faithful use of long contexts. Research has documented the "lost in the middle" phenomenon: information placed in the middle of a long context is retrieved less reliably than information at the beginning or end, even when the context fits comfortably within the model's window.

The magnitude of this effect is substantial. In controlled experiments where a single factual question could be answered from a specific passage within a long document, models showed significantly lower accuracy when the relevant passage was in the middle of the document compared to the start or end. The degradation was not gradual but showed a characteristic U-shape: high performance for early and late positions, substantially lower performance for middle positions, even with passages that the model could easily answer from if presented alone.

This suggests that models have learned to pay disproportionate attention to certain positions, a learned bias from pretraining on documents where key information often appears at the start or end. Many natural documents do have this structure: introductions summarize key points, conclusions recapitulate them, and supporting evidence sits in between. A model that has internalized this structure will be poorly suited to tasks that require extracting information from arbitrary positions in long inputs.

Addressing this requires either data curation changes during pretraining to include documents where key information appears in non-prominent positions, architectural changes to how position interacts with attention, or both. Some recent work on positional encodings specifically addresses the locality bias that causes models to disproportionately attend to nearby tokens, but general solutions that maintain faithful recall at arbitrary positions across very long contexts remain an open problem.

Long-context faithfulness also raises evaluation challenges. Standard benchmarks test short-context capabilities. Evaluating whether a model faithfully processes a 100,000-token document requires task designs that specifically test information from the middle of the document, against distractor information elsewhere in the same document. Needle-in-a-haystack evaluations, where a specific fact is embedded in a long irrelevant context and the model is asked to retrieve it, are a common approach, but they measure a narrow capability that may not generalize to more complex long-context reasoning tasks.

The benchmark evaluation problem is arguably meta-level, but it shapes every other research direction. A model that scores highly on a benchmark is not necessarily solving the underlying task; it may be exploiting patterns in how the benchmark was constructed, benefiting from data contamination, or optimizing toward the specific format of the evaluation without generalizing to related tasks.

Data contamination is a well-documented example. If a model's pretraining corpus contains examples from a benchmark's test set, the model's score on that benchmark reflects memorization rather than generalization. Given the size of modern pretraining datasets, which may contain hundreds of billions or trillions of tokens from diverse web sources, and the number of benchmarks in circulation, contamination is nearly impossible to rule out entirely without full dataset transparency. Even approximate contamination, where the model has seen paraphrased versions of test questions, can inflate scores in ways that are difficult to measure.

Benchmark saturation is a related issue. Once models achieve near-ceiling performance on a benchmark, it stops differentiating capabilities. The GLUE benchmark was saturated within two years of its introduction. Its successor SuperGLUE was saturated within another two. The field has shifted to progressively more difficult benchmarks, but the cycle repeats: today's hardest benchmarks become tomorrow's solved problems, and it is not always clear whether the improvement reflects general capability gains or optimization toward the benchmark's specific format. A model that achieves superhuman scores on MMLU has not demonstrated superhuman knowledge; it has demonstrated the ability to select correct answers in the multiple-choice format that MMLU uses, which may involve very different skills than producing correct answers without options provided.

The deeper problem is that we do not have a satisfying theory of what benchmarks measure. A benchmark score is a statistic about model behavior on a specific distribution of inputs, in a specific format, evaluated by a specific protocol. Generalizing from that statistic to claims about model capability requires assumptions about the relationship between benchmark performance and real-world task performance that are difficult to validate. Developing evaluation methodology that produces claims with clearly characterized scope and limitations is one of the most important open problems in the field.

Beyond identifying what does not work, the field is actively exploring approaches with demonstrated promise. These directions are not speculative; each has produced meaningful results and has identifiable paths to further progress. What distinguishes a promising direction from a speculative one is the presence of concrete empirical milestones, a plausible causal account of why the approach should work, and at least preliminary evidence that it scales or generalizes beyond the specific setting where it was first demonstrated.

One of the most active research areas is training models to check their own work. The intuition is simple: humans who solve complex problems benefit from checking their reasoning, not just producing answers. A mathematician who derives a result typically re-reads the derivation looking for errors before publishing it. A programmer who writes code runs it against test cases before deploying it. Models that can identify when their own reasoning has gone wrong and self-correct should produce more reliable outputs.

This direction has produced several concrete approaches with demonstrated effectiveness. Self-consistency decoding generates multiple reasoning paths for the same problem by sampling with temperature, then takes a majority vote over final answers. Correctness is measured by agreement rather than by ground-truth comparison, which makes this usable at inference time without labeled data. The key insight is that if a model generates many diverse paths and most of them reach the same answer, that answer is more likely to be correct than one generated by a single path, even if the individual paths are noisy. Self-consistency has shown consistent improvements across math, coding, and common-sense reasoning tasks, often with substantial gains for difficult problems where single-path chain-of-thought is most unreliable.

Iterative self-refinement takes a different approach: training models to produce critiques of their own outputs and then revise based on those critiques. A model generates an initial answer, then generates a critique pointing out potential errors or improvements, then generates a revised answer incorporating the critique. This cycle can be repeated multiple times. The challenge is that a model which generates a flawed answer is often unable to generate an accurate critique of that answer: the same bias or gap in knowledge that caused the error may prevent the model from identifying it. Nonetheless, even imperfect self-critique can catch a significant fraction of errors, particularly surface-level errors like factual inconsistencies within the generated text.

Constitutional AI, covered in the alignment chapters, uses self-critique to improve the harmlessness of model outputs. Rather than relying on human annotation to identify harmful outputs, the model is given a set of principles and asked to evaluate its own outputs against those principles. This approach extends the self-verification idea from correctness to values, and shows that models can learn to apply abstract principles to specific cases in useful ways.

The key research challenge is circularity: a model that makes a particular reasoning error will often make the same error when asked to evaluate its own reasoning, because the same representations that produced the error influence the evaluation. Breaking this circularity requires either external feedback from a different source, diverse sampling strategies that produce materially different reasoning paths from the same starting point, or training signals that specifically reward identifying one's own errors rather than just producing correct outputs.

Standard interpretability methods ask "what does this neuron respond to?" answering with high-activation examples that give a vague sense of what a neuron's function might be. Mechanistic interpretability asks a more ambitious question: what algorithm does this circuit implement? The goal is to understand, at the level of weights and activations, how specific model behaviors emerge from specific computational substructures, not just which inputs correlate with which activations.

Progress in this direction has been surprising and encouraging. Researchers have identified attention heads that implement induction, a behavior where the model attends to tokens that followed a given token in prior context, enabling in-context learning to retrieve and continue patterns. They have found name-mover heads that track which proper nouns appear in a prompt and route that information to the appropriate output positions. They have characterized circuit-level implementations of specific syntactic operations, where a small set of attention heads and MLP layers collectively perform a computation like indirect object identification. These findings demonstrate that interpretability at a mechanistic level is possible, not just aspirational.

The research challenge is scale. Current mechanistic interpretability results focus on small models, often models with millions rather than billions of parameters, and on simple, well-defined behaviors that can be cleanly isolated. Extending these techniques to frontier models requires new tools for handling the combinatorial explosion of possible circuits in a model with hundreds of billions of parameters. Any given behavior in a large model may be implemented across hundreds of attention heads and MLP layers, with redundancy and interference between components that makes clean circuit identification difficult.

Superposition is a particularly active subproblem in mechanistic interpretability. Models represent more features than they have dimensions by overlapping representations in high-dimensional space: multiple distinct concepts are encoded along directions that are not perfectly orthogonal, allowing a model with $d$ hidden dimensions to represent many more than $d$ features. This makes it difficult to identify which features a given neuron is responsible for, because each neuron participates in encoding many overlapping features. Sparse autoencoders, which learn dictionaries of sparse basis vectors that reconstruct the model's activations, have emerged as a promising tool for disentangling superposed representations, but the approach is computationally expensive and its results are difficult to validate.

Understanding superposition matters for interpretability and safety: if models encode dangerous capabilities in superposed representations, it may be possible to trigger those capabilities through carefully chosen inputs even when the model's outputs appear benign. Mechanistic interpretability provides the tools to look for such hidden representations before they manifest as harmful behaviors in deployment.

The quality of pretraining data matters more than was initially appreciated, and the field's understanding of what "quality" means has become more sophisticated over time. Research has shown that models trained on carefully curated, high-quality data achieve better performance per parameter and per training compute than models trained on larger but lower-quality datasets. Llama 3's training data curation, which involved extensive deduplication, quality filtering, and careful management of data mixtures, produced models that outperformed some models with substantially more total training compute. This finding motivates serious investment in data curation as both a research area and an engineering preprocessing step.

Several research questions remain open. How do we define and measure data quality at scale, without requiring human annotation of hundreds of billions of documents? Early quality heuristics relied on simple proxies like document length, token count, and similarity to high-quality reference corpora like Wikipedia. More recent approaches use classifier models to predict quality, but these classifiers need training data, which creates a circularity: the best source of training data for a quality classifier is a human judgment of quality, which is exactly what we are trying to replace.

What kinds of data diversity matter for what kinds of downstream capabilities? A model trained exclusively on web text may have better language fluency but worse mathematical reasoning than a model trained on a mixture that includes textbooks, code, and scientific papers. The right mixture is task-dependent and cannot be determined from first principles; it requires empirical experimentation that is expensive at the scale of frontier model training.

How do we detect and handle near-duplicates and contaminated examples in datasets that are too large to inspect manually? Near-duplication in training data has been shown to cause models to memorize examples more deeply and generalize worse. Deduplication at scale requires approximate methods like MinHash locality-sensitive hashing, but the right deduplication threshold, too aggressive and you remove distinct documents that happen to be similar, too lenient and memorization persists, is difficult to calibrate.

Data mixture research examines how the proportions of different data types affect model capabilities. The findings suggest that the mix matters substantially for specific capabilities but that optimal mixes depend on evaluation criteria, making it difficult to define a single best mixture that serves all use cases. A model optimized for coding tasks needs a different mixture than one optimized for multilingual translation, and a model intended for general use must balance these competing demands. Dynamic data mixtures that change during training, stressing different data types at different stages, are an active research direction.

As we covered in the inference optimization and model compression chapters, serving large language models at scale is expensive. Training frontier models costs tens to hundreds of millions of dollars in compute alone, and inference costs at scale are a significant fraction of the operational costs of model-based products. Research in this area has two primary objectives: reducing the compute required to train capable models, and reducing the compute and memory required to run them.

On the training side, mixture of experts architectures activate only a subset of model parameters for each input, giving a way to increase total model capacity without proportionally increasing per-token compute. The Mixtral model family demonstrated that sparse activation could match dense model quality at significantly lower training costs: a model with a large total parameter count but sparse activation can outperform a smaller dense model trained with the same compute budget. Open research questions include how to train expert routing stably, how to prevent experts from collapsing so that only a few experts receive the bulk of the training signal, and how to enable effective expert specialization without explicit supervision about which expert should handle which type of input.

On the inference side, speculative decoding uses a small draft model to propose token sequences that are then verified by the large model. When the draft is usually right, the verification step is fast because it parallelizes across the proposed sequence, achieving throughput close to the small model while maintaining the quality of the large model. Research challenges include training effective draft models that approximate the large model's distribution well, handling distribution shift between draft and target models as the models evolve, and extending speculative decoding to settings where the draft and target models have different tokenizers or architectures.

Quantization research continues to push the limits of how aggressively model weights and activations can be compressed without quality loss. 4-bit and even 2-bit quantization schemes that preserve model quality on key benchmarks have been demonstrated for inference, dramatically reducing the memory requirements for deploying large models. The mechanisms by which quantization affects model behavior in edge cases are not fully understood, however. A model that is quantized to 4 bits may maintain average benchmark scores while showing degraded performance on specific types of inputs, and identifying which inputs are affected and why requires tools from interpretability as well as efficiency research.

The alignment problem, as we discussed in the RLHF chapter, involves training models to behave in ways that are helpful, harmless, and honest across the full distribution of inputs and contexts they will encounter. This sounds simple as a statement of goals but is technically difficult: the space of possible inputs is vast, human preferences over model behaviors are not always consistent or well-defined, and models can learn to game the training signal rather than satisfying the underlying intent.

Current alignment methods rely primarily on reinforcement learning from human feedback, which has known limitations. Human feedback is expensive to collect at the scale needed to cover a broad distribution of inputs and edge cases. Annotators disagree on subjective judgments, particularly for questions involving values, cultural context, or specialized knowledge. And models can learn to optimize for the reward model that summarizes human feedback rather than for the underlying human preferences that the reward model is trying to capture. This reward hacking manifests as model outputs that score highly on automated evaluation but fail in ways that human reviewers would reject.

Research is exploring alternatives and extensions. Constitutional AI replaces per-example human feedback with general principles, reducing annotation cost while improving consistency. The model is given a constitution, a set of principles about what good behavior looks like, and trained to evaluate its own outputs against those principles and revise them accordingly. This approach requires careful design of the constitution and does not eliminate the need for some human oversight, but it scales better than per-example feedback. Direct preference optimization (DPO) eliminates the reward model entirely, training directly on preference pairs with a simplified objective that is more stable and computationally efficient than RLHF while producing similar results on standard alignment benchmarks.

A harder version of the alignment problem is value learning under distributional shift: aligning model behavior on the distribution of evaluation prompts and on the long tail of unusual inputs, adversarial inputs, and situations the designers did not anticipate. Constitutional approaches partially address this by encoding values at a level of abstraction above specific examples, which allows principles to generalize to novel situations. But the gap between stated values and implemented behavior remains a significant challenge. A model can be trained to produce outputs that express correct values without having internalized those values in a way that generalizes robustly. Understanding when stated values correspond to stable behavioral dispositions and when they are surface patterns learned from value-expressing training data is a core question in alignment research.

A research direction that has grown rapidly alongside capability improvements is the development of agentic systems: models that can take actions in the world, use external tools, plan over extended time horizons, and coordinate with other models or systems to accomplish complex goals. An agent that can search the web, read documents, write and execute code, and revise its approach based on intermediate results represents a qualitatively different kind of system than a model that simply generates a response to a prompt.

Agentic systems introduce new research challenges beyond those of standard language model training. Planning under uncertainty requires models to reason about the consequences of actions they have not yet taken and to make decisions when information is incomplete. Tool use requires models to learn when to call external tools versus when to rely on internal knowledge, and to correctly interpret tool outputs in context. Error recovery requires models to recognize when a plan has failed and to develop alternative approaches without human intervention. These capabilities require training signals and evaluation environments that are substantially more complex than standard language model benchmarks.

Multi-agent settings, where multiple language models interact with each other and with humans to accomplish shared goals, introduce additional complexity around coordination, communication, and the alignment of individual model behaviors with collective outcomes. Research questions in this area include how to assign credit for outcomes in multi-agent systems, how to prevent behaviors that are individually rational but collectively harmful, and how to ensure that emergent behaviors in multi-agent systems remain aligned with human preferences.

The evaluation system for language AI has significant gaps that affect which capabilities receive research attention and which remain unmeasured. Understanding these gaps is as important for a researcher as understanding the benchmarks themselves, because evaluation choices shape what gets optimized and what remains invisible.

The problem of evaluation is not merely technical. It is partly sociological: the benchmarks that get used are those that are easy to run, easy to compare across papers, and well-established enough that reviewers recognize them. This creates inertia toward existing benchmarks even after they become saturated, and a bias against new evaluation methodologies that require specialized infrastructure or careful experimental design.

Most existing benchmarks evaluate final answers, not reasoning processes. A model that produces the right answer by the wrong reasoning, or by exploiting dataset biases, is indistinguishable on a standard benchmark from one that understood the problem. Developing evaluations that probe understanding at a finer grain than correct or incorrect remains methodologically difficult and receives less research attention than capability benchmarks.

One approach is counterfactual evaluation: present a model with a question, observe its answer, then modify the question in a way that should change the correct answer without changing surface features, and check whether the model's answer changes accordingly. If a model answers correctly on the original but incorrectly on the counterfactual, it likely relied on surface cues rather than understanding the underlying structure. This approach has been used effectively to demonstrate that models' impressive performance on natural language inference benchmarks is partly explained by spurious correlations in the benchmark rather than entailment reasoning.

Another approach is process evaluation: ask models to explain their reasoning step by step, then separately evaluate whether each step is logically valid, whether the chain of steps leads to the stated conclusion, and whether the conclusion is correct. This provides richer signal than binary correct or incorrect: a model that always produces a correct final answer but through flawed reasoning is a different kind of failure from one that produces correct reasoning but an incorrect conclusion. Process evaluation requires more evaluation infrastructure, including either trained human annotators or reliable automated step-validity checkers, but the richer signal it provides is valuable for guiding improvements in reasoning capability.

Adversarial evaluation, where the evaluation set is constructed specifically to challenge models' surface-level pattern matching, is another approach to probing understanding. Adversarial examples that preserve semantic content while changing surface features, such as paraphrasing a question in ways that should not change the answer, can reveal when models are relying on superficial patterns. The challenge is that constructing effective adversarial examples requires knowing in advance what shortcuts the model is using, which is itself a non-trivial research problem.

Current benchmarks mostly evaluate single-turn interactions or short multi-turn conversations. Real-world use cases increasingly involve long-horizon tasks: planning a research project over multiple days, debugging a complex codebase with many interdependencies, managing a multi-step workflow that requires coordination between tools, data sources, and human collaborators. Standard benchmarks do not capture performance on these tasks, and models that perform impressively on standard benchmarks can fail badly on extended real-world usage.

The SWE-bench benchmark, which evaluates whether models can resolve GitHub issues in open-source Python repositories, is one step toward realistic long-horizon evaluation. The tasks require understanding a codebase that may span thousands of lines and many files, identifying the relevant change needed to fix a reported issue, implementing that change correctly without breaking other functionality, and passing the existing test suite. This is substantially more realistic than standard code generation benchmarks that ask models to implement isolated functions from a docstring. Performance gaps between models are often larger on SWE-bench than on simpler coding tests, suggesting it captures something about long-horizon reasoning that shorter tasks do not.

Developing evaluation environments for long-horizon tasks is itself a research challenge. Such tasks require reproducible execution environments where model actions have consistent effects, well-defined success criteria that do not require human judgment to evaluate, and evaluation procedures that are fair across different approaches. Automated evaluation is difficult because success on complex tasks often requires qualitative assessment: a model might technically pass a test suite while implementing a solution that is brittle, insecure, or poorly structured.

Interactive evaluation environments, where models must complete tasks through sequential actions with intermediate feedback, represent a particularly important frontier. Tasks like research planning, experimental design, and complex data analysis require models to update their approach based on intermediate results, which cannot be evaluated from a single generated output. Standardizing evaluation protocols for interactive tasks, and building the infrastructure needed to run them reproducibly at scale, is an important area of community investment.

The majority of benchmarks are in English. The majority of language AI research optimizes for English performance. This creates a systematic gap: capabilities that are well-measured and well-optimized for English may not generalize to other languages, particularly those with different morphological structures, writing systems, or lower representation in pretraining data. From the perspective of the billions of people who primarily communicate in languages other than English, the state of language AI is substantially worse than English-language benchmark scores suggest.

Multilingual benchmarks like XCOPA, XNLI, and TyDiQA exist but have smaller coverage and lower diversity than their English counterparts. They tend to cover typologically similar languages with larger digital footprints, leaving languages with fundamentally different structures, including polysynthetic languages, languages with non-Latin scripts, and languages with limited web presence, without meaningful benchmark coverage. More importantly, high performance on existing multilingual benchmarks is often achieved by models that were primarily trained on English and rely on cross-lingual transfer, rather than models with multilingual understanding of each language's specific structures and cultural context.

Low-resource language evaluation is particularly underdeveloped. Many languages spoken by tens of millions of people have essentially no benchmark coverage, which means it is unclear how current models perform on them or how to measure improvement. The research community has begun to address this through community-driven benchmark construction projects, such as Masakhane for African languages and AmericasNLP for indigenous languages of the Americas. These projects face fundamental challenges: the linguistic expertise, annotator availability, and data resources needed to construct high-quality benchmarks are much harder to assemble for low-resource languages than for well-resourced ones.

This is an area where the gap between what is measured and what matters is particularly large. A model that performs poorly on low-resource languages may still be reported as achieving excellent results on standard multilingual benchmarks, because those benchmarks weight languages by coverage rather than by speaker population or need. Developing evaluation infrastructure that gives appropriate weight to low-resource languages is a community-level challenge that requires sustained investment over years.

Models that perform well on test sets drawn from the same distribution as training data often fail when evaluated on inputs from different distributions. This distribution sensitivity is well documented but poorly measured by standard evaluation: most benchmarks do not include out-of-distribution test sets, so robustness is typically measured only on specific datasets designed for that purpose, which are themselves limited in scope.

Constructing good robustness evaluations requires defining what constitutes a meaningful distribution shift for a given task. For question answering, meaningful shifts might include questions about recent events after the model's training cutoff, questions in domains underrepresented in training, questions with adversarially modified surface features but the same underlying answer, or questions where common world-knowledge shortcuts do not apply. Different shifts stress-test different aspects of model behavior, and a model that is robust to one type of shift may be brittle to another.

Temporal distribution shift is a particularly important case. Models have training cutoffs beyond which they have no direct knowledge of world events. But the problem is more subtle than simply not knowing recent facts: models may confidently answer questions about topics that have changed significantly since their training, applying outdated knowledge without any signal to the user that the information may be stale. Developing evaluation methodology for temporal robustness, and training models that are appropriately uncertain about potentially outdated information, is an active research area with clear practical importance.

Research direction is shaped by technical opportunities as well as community structures, incentive systems, and resource constraints. Understanding these dynamics helps explain why some problems receive more attention than others, and why certain technically important areas remain underinvested despite their potential impact.

The concentration of compute, data, and talent at a small number of large organizations creates a structural tension in the research ecosystem. Frontier model development requires resources that only a handful of organizations can provide, while academic research institutions have access to models and infrastructure that lag behind the frontier. This shapes what questions are investigated, what results are published, and what the broader research community can study and build on.

The dominance of closed, proprietary models in the frontier of capability created a significant research access problem: researchers at academic institutions could evaluate frontier models through APIs but could not study their internals. Interpretability, fine-tuning research, mechanistic analysis, and reproducibility all require access to model weights and activations, not just API outputs. This access gap meant that a large fraction of the most important research questions about how frontier models work could only be investigated by researchers at the organizations that trained them.

The release of LLaMA and its successors dramatically changed this field. Open-weight models at 7B to 70B parameter scales became the baseline for a large fraction of research, allowing work on fine-tuning, alignment, interpretability, and efficient inference that would have been impossible with API-only access. The Mistral, Qwen, and Gemma families expanded the range of open-weight models available for research, giving diverse architectures and training methodologies that researchers could study and build on. The ecosystem around open-weight models, including efficient inference engines like llama.cpp, fine-tuning frameworks like Hugging Face's TRL and LLaMA Factory, and evaluation frameworks like lm-eval, has become the primary infrastructure for academic language model research.

The gap between open-weight frontier models and closed frontier models remains significant in terms of capability. Most research using open-weight models is working with models that lag behind the current commercial frontier by some margin. Understanding how findings from smaller, accessible models transfer to larger models that remain proprietary is a persistent methodological challenge. A finding about interpretability or alignment that holds at 7B parameters may or may not hold at 70B or 700B; testing this requires access that most researchers do not have.

The community debate about model release, whether to release weights openly, through limited licenses, or not at all, involves material trade-offs that the research community has not fully resolved. Open release enables research, accelerates community-driven improvements, and democratizes access. It also enables fine-tuning for malicious purposes and removes the control that model developers have over downstream use. The norms around open release are evolving, with different organizations taking different positions and the policy environment changing as governments develop AI regulation frameworks.

The pace of publication in language AI is high enough that reproduction of results before publication, and detailed experimental methodology in papers, are both under sustained pressure. Many papers report state-of-the-art results on specific benchmarks under conditions that are difficult to reproduce: specific prompting strategies that are not fully documented, proprietary evaluation infrastructure with implementation details unavailable, training compute that academic labs cannot replicate, or evaluation conditions that differ from the standard benchmark protocols in ways that are not always clearly disclosed.

The incentive structure of the field rewards claiming state-of-the-art results, which encourages optimization of benchmark performance rather than thorough methodology documentation. A paper that carefully documents why its results may not generalize, or that quantifies the variance across multiple experimental runs, is often less competitive in the review process than a paper that simply reports the best numbers, even if the former provides more reliable scientific information.

The community has developed partial responses to this. The BigScience, EleutherAI, and LAION collaborations have produced both open models and open training frameworks that enable reproduction at meaningful scales, and have established norms around documentation and transparency that larger organizations sometimes adopt. Evaluation frameworks like the Language Model Evaluation Harness (lm-eval) standardize benchmark evaluation, making it easier to compare results across papers and harder for researchers to inadvertently use non-standard evaluation conditions. Cards for models and datasets, including model cards, dataset cards, and evaluation cards, document the conditions under which results are valid and the limitations of the artifacts.

Improving reproducibility norms is itself a community-level research problem. It requires identifying which evaluation choices are most consequential for result comparability, developing infrastructure that makes standard evaluation easy and deviation from standard evaluation explicit, and creating incentive structures that reward rigorous methodology rather than just benchmark performance. Conferences and journals in the field are increasingly requiring model release, code release, and evaluation transparency, but enforcement is inconsistent and the gap between stated norms and actual practice remains large.

As we discussed in the safety and security chapter, the research community is increasingly engaging with questions about the long-term behavior of increasingly capable models. These questions motivate research into alignment, interpretability, and evaluation for dangerous capabilities, and they have begun to influence the norms around model release, evaluation disclosure, and deployment practice.

The community is not uniform in its views on the importance or character of AI safety risks. Some researchers focus primarily on near-term harms: biased outputs that disadvantage specific groups, misuse of model capabilities for disinformation or fraud, privacy violations through memorization of training data, and economic displacement from automation. These concerns are grounded in immediate, measurable harms and are addressed through techniques including bias evaluation, differential privacy, content moderation, and appropriate deployment restrictions.

Others focus on longer-term concerns about the behavior of models significantly more capable than current systems. These concerns motivate research into deception, where models might learn to behave differently when they believe they are being evaluated, goal generalization, where models pursue goals in training that generalize to harmful objectives in deployment, and autonomous operation, where models take actions in the world with potential for large-scale impact. The empirical basis for these concerns is less developed than for near-term harms, which makes them more controversial within the research community.

Both sets of concerns motivate concrete research with substantial methodological overlap. Evaluation for dangerous capabilities, interpretability methods that can detect misaligned representations, and training procedures that produce more controllable behaviors are relevant to both near-term and longer-term safety questions. The debate is primarily about priorities and resource allocation rather than about whether safety research is valuable.

Community norms around capability evaluations are evolving in response to these concerns. Major laboratories have adopted responsible scaling policies that tie deployment and training decisions to evaluations of dangerous capabilities. These policies specify thresholds above which certain deployment restrictions apply and commit organizations to conducting evaluations before proceeding with training runs that might cross capability thresholds. Academic researchers are developing shared evaluation methodologies for capabilities like autonomous operation, deception, and cyberoffense that could serve as a common framework across organizations, reducing the risk that capability evaluations are gamed or that different organizations use incompatible measurement approaches.

Whether you are starting research in this field or trying to direct existing research toward areas of highest impact, the field described above suggests several practical observations about where effort is most likely to produce useful results.

The problems most likely to yield to new ideas are those with clear evaluation criteria and tractable sub-problems that can be studied independently. Hallucination is hard to measure in aggregate but factual accuracy on specific domains is measurable, and improving it on even narrow domains provides useful signal about what interventions work. Systematic generalization is hard to study in large, opaque models but can be studied in controlled settings that isolate the relevant variables. Identifying the tractable sub-problem within a large open problem is often the most important creative act in starting a new research direction.

The problems most likely to require community-scale efforts are those involving data. Improving multilingual representation, constructing long-horizon evaluation environments, building robustness evaluation suites, and creating evaluation infrastructure for low-resource languages all require more data work, annotation work, and infrastructure work than any single research group can accomplish. These are areas where community collaboration, standardization, and sustained infrastructure investment pay disproportionate dividends relative to individual research contributions.

The problems most likely to be solved by scaling alone are narrower than they appear from the outside. Scaling has reliably improved average performance on standard benchmarks, but it has not consistently closed the gap on systematic generalization, calibration, or long-context faithfulness. Researchers who believe that scale is the primary driver of progress should check their intuitions against these specific failure cases, which have persisted across multiple generations of scaling. The persistence of specific failure modes across orders of magnitude of parameter count and compute is often the best evidence that those failure modes have architectural or training causes that are not addressed by scale.

The problems most underinvested relative to their impact, based on the gap between their importance and the research attention they receive, are evaluation methodology, multilingual capability, and mechanistic interpretability. Evaluation methodology is the bottleneck for measuring progress on everything else; without better evaluation, it is impossible to know whether proposed solutions to open problems work. Multilingual capability affects the majority of the world's population and is systematically underserved by current research. Mechanistic interpretability is the most promising path to understanding why models fail when they fail, which is necessary for designing principled improvements rather than empirical patches.

To make the abstract discussion concrete, let us work through a practical analysis of how model capabilities relate to benchmark scores. We will construct a synthetic but realistic comparison of model performance across different evaluation categories, illustrating the kinds of patterns researchers look for when assessing capability profiles.

We start by modeling performance across several benchmark dimensions for hypothetical model families at different scales.

The research attention trends reveal a structural feature of the field: attention flows toward tractable problems with clear evaluation criteria and visible payoffs. Scaling and pretraining, fine-tuning, and efficiency work all have these features. Interpretability, evaluation methodology, and multilingual research, which arguably have higher impact per researcher-hour but slower and harder-to-measure payoffs, grow more slowly. These dynamics are not solely a product of researchers making bad choices; they reflect the incentive structures of academic publication, industry deployment needs, and the practical difficulty of constructing the evaluation infrastructure that harder problems require. Understanding these dynamics can help you make deliberate choices about where to direct your own research attention.

This chapter has described the current research frontier, but any such description has a short shelf life. The field moves quickly, and the problems that seem hardest today may yield to a new architectural insight, a new training objective, or a new evaluation framework developed next month. The inverse is also true: what appears to be a solved problem can reveal new dimensions of failure when probed carefully or applied to new domains.

There are also problems this chapter has not covered in depth. Embodied language learning, where models develop language understanding through interaction with physical environments rather than text alone, remains a rich research area largely separate from the text-based models covered in this book. Language grounded in physical interaction may generalize differently and have different failure modes than text-only models. Neuroscience-inspired approaches to language modeling draw on cognitive science and linguistics in ways that the statistics-forward mainstream of the field has not fully engaged; understanding how the brain processes language may suggest inductive biases that improve sample efficiency or generalization. The interaction between language models and formal verification, where the goal is mathematically certified correctness rather than statistically probable correctness, is another frontier with high potential value for applications in software verification, mathematics, and safety-critical systems.

The gap between what models can do on benchmarks and what they can be trusted to do in deployment remains the most important practical frontier. Models that perform impressively in controlled evaluation settings continue to fail in unpredictable ways when deployed to real users with real needs and real adversarial intentions. Closing this gap requires better evaluation, better training, and better understanding of why current models fail where they fail. The tools for this understanding are being developed now, and their maturation will shape the next generation of capable and reliable language AI systems.

The research directions described in this chapter are not independent; progress on any one of them tends to enable progress on others. Mechanistic interpretability provides tools for understanding where reasoning fails, which informs better training objectives for improved reasoning. Better calibration helps distinguish confident correct answers from confident errors, which improves the usefulness of self-verification. Improved evaluation identifies the specific failure modes that fine-tuning and alignment research should target. The field's challenges are interconnected in ways that mean sustained progress requires work across all of these areas rather than only the ones that are easiest to make progress on in the short term.

As we conclude this handbook, the field is at an open moment. The foundation of modern language AI is robust: transformers, attention, pretraining, and scaling laws are settled knowledge that forms the base on which new work builds. What is built on that base over the next decade is an open question, shaped by the research agenda described in this chapter and by discoveries not yet made. The ideas in this book, from n-gram language models through self-attention to alignment, represent the tools available to the research community today. What you do with them, and what new tools you develop, is the open problem that matters most.

This chapter surveyed the open problems, promising research directions, benchmark gaps, and community priorities that define the current frontier of language AI research.

Open problems include:

• Reliable factual grounding: Models hallucinate confidently, distinctions between evidence-based knowledge and pattern-based confabulation are not architecturally encoded, and calibration is poor especially in out-of-distribution domains.
• Systematic generalization: Models struggle to compose known concepts in novel arrangements, suggesting they learn surface statistics rather than abstract compositional rules.
• Consistent multi-step reasoning: Chain-of-thought reasoning is brittle to rephrasing, and distinguishing valid from simulated reasoning remains technically difficult.
• Long-context faithfulness: The "lost in the middle" phenomenon reveals that large context windows do not guarantee faithful use of long contexts at all positions.
• Robust evaluation: Data contamination, benchmark saturation, and the lack of a principled theory of what benchmark scores measure undermine the reliability of standard evaluations.

Promising research directions include:

• Self-verification: Self-consistency decoding, iterative refinement, and constitutional AI show that models can check and improve their own outputs, though circularity remains a challenge.
• Mechanistic interpretability: Circuit-level analysis has identified specific computational substructures behind specific behaviors, though scale is a significant barrier.
• Better data curation: Data quality matters more than raw quantity, and research into quality measurement, deduplication, and mixture optimization is yielding real improvements.
• Efficient training and inference: Mixture-of-experts, speculative decoding, and aggressive quantization are making frontier-quality models more accessible and affordable.
• Alignment research: DPO, constitutional AI, and debate-based evaluation are extending the alignment toolkit beyond standard RLHF.
• Agentic systems: Long-horizon planning, tool use, and multi-agent coordination represent a qualitatively new capability frontier with its own open research questions.

Benchmark gaps exist in testing understanding versus pattern matching, evaluating long-horizon tasks, covering multilingual and low-resource languages, and measuring robustness to distribution shift.

Community dynamics shape which problems receive attention: open-weight model releases have expanded research access, reproducibility remains challenging, and the community is actively developing evaluation norms around safety capabilities.

The most underinvested areas relative to their impact are evaluation methodology, multilingual capability, and mechanistic interpretability. The foundation laid throughout this book equips you to engage with these open problems directly. The next step is to pick a problem that interests you, find the benchmark or evaluation environment that measures progress on it, and start building.

Ready to test your understanding? Take this quick quiz to reinforce what you've learned about Research Directions in Language AI.