Skip to content
Menu
Agenda

Research agenda

Interpretability has begun to open artificial networks, and psychology and neuroscience have long read minds from outside. Both now ask what a system represents, uses, reports, keeps and holds steady. Far Akin's programme is to answer those questions on both sides under one standard of evidence, so that a claim that a model and a mind are alike can be tested rather than asserted.

Updated
Version
1.0
Generated cyanotype print of a two-masted sailing ship on a calm sea, with a finely detailed rocky coastline to the right and open, unmarked water ahead.Generated photograph of a long white plaster sea-chart relief: a detailed coast on the left fades into blank plaster on the right, crossed by a line of vermilion pins.Generated night-survey render of parallel survey strips over a dark sea floor: two complete, one half run with a small ship at its end, and faint planned lines ahead.
Plate IThe coast behind is charted. The water ahead is not.Charted coast, blank sea. The route is planned point by point and stops where the chart stops.Lines run, a line under way and lines planned. An agenda is a survey plan.Generated image in a cyanotype treatment. Far Akin art direction.Generated image. Far Akin art direction, Datum look.Generated image. Far Akin art direction, Swath look.

The question

Minds and models share one condition: neither can be read directly for what it holds. A person's inner life is known through what they do, what they report and what changes when they are disturbed. A model's hidden state can be recorded in full, yet what it represents is not written on it. Psychology, neuroscience and cybernetics spent more than a century building ways to read an interior from outside. Interpretability is building them again for artificial networks.

Far Akin's long-horizon programme is to bring the two together: theory and instruments that put the same questions to both kinds of system under one standard of evidence. What does a system represent, what does it use, what can it report, what does it keep, and what does it hold steady when disturbed? A claim that a model and a mind are alike, or unlike, should be something that can be tested rather than asserted. Each inquiry on this site is a step towards that comparative science of minds.

The first inquiries start where the need is most immediate. Models of worlds, of people and of a model's own interior are already used to predict and decide.

The problem

Learned models of worlds and of people are moving from demonstrations into use. A world model learns how an environment responds to action. An agent has been trained entirely inside a world model's own generated dream and its policy carried back to the real environment[1]. A single algorithm that improves its behaviour by imagining future scenarios has been applied across more than 150 tasks with one configuration[2]. Generative environments learned from unlabelled internet video can be prompted into worlds that respond to actions frame by frame[3]. In robotics, world models are increasingly used to judge policies: they simulate the consequences of a policy's actions and return a success or safety verdict[4].

Models of people are following the same path. Language models conditioned on the backstories of real survey participants have been proposed as proxies for human subpopulations in social science[5], and agents built for 1,052 people from interviews, surveys or both predicted their held-out survey answers at 82% to 86% of the people's own two-week consistency[6]. Such results make simulated respondents attractive for work that would otherwise need people.

The difficulty is that the evidence for these models is gathered in one place and the use happens in another. Visual quality ranked two driving world models the opposite way from action-following[4], and physical understanding in video generators was unrelated to visual realism[7]. Generative models that pass existing diagnostics for world models can hold world models far less coherent than they appear, and fail on related but subtly different tasks[8]. On the human side, answers generated by ChatGPT that matched average feeling-thermometer scores for sociopolitical groups closely showed less variation than real respondents, regression estimates that differed from the real survey, sensitivity to small changes in wording, and different results from the same prompt over three months[9]. The opinions language models reflect diverge substantially from those of many demographic groups, and the gap persists when the model is steered towards a group[10].

None of this makes such models useless. The lab's working hypothesis is that their fidelity is local: a model may look complete everywhere and be trustworthy only near the places where it was checked. A depth chart is like this: exact at each sounding, inferred everywhere between, with no mark to show where the inference fails. The problem the lab works on is that models of worlds and of people are starting to be used to predict and decide before anyone can say which parts of them were measured and which were filled in.

Older disciplines met parts of this problem. Simulation engineering separates verification, validation and accreditation, and distinguishes conceptual model validity, operational validity and data validity[11]⁠[12]. Psychological measurement built the theory of construct validity and the method of convergent and discriminant validation[13]⁠[14]. Learned world models and simulated people inherit both problems, and the older tools do not carry over unchanged. Classical validation assumes a trusted simulator testing an untrusted policy; a learned simulator is itself an unverified artefact[4]. A simulated respondent may have read, during training, the survey it is asked to answer.

The aim

The lab's aim is to explain how natural and artificial systems come to hold what they hold, and to test those explanations. Its first inquiries apply that aim where reliability is in question: the internal state and memory of language models, where a model's knowledge should live, world models and simulated people. For these, the aim is reliability that is measured rather than asserted, and instruments built to measure it. A model whose faithful region is unknown cannot responsibly carry a decision, and an instrument that has never met a model able to fail it shows little.

For each model, a design states the use it is checked for, the axis along which test conditions move away from the conditions it was built on, the comparison it has to beat, and the result that would count against it. A grade is then a statement of the form: for this use, within this distance, this model is akin to what it models, with this uncertainty. Outside that region the grade is silent, and says so.

The name states the test. Far Akin: far apart, possibly akin. A world model and a world, a simulated person and a person, a specialist model and the expertise it stands in for are far apart in kind. Whether they are akin in the respects a given use depends on is a question for evidence. Where a model is akin to what it models, the work is to show it. Where the resemblance breaks, the break is a result: it marks where the model stops being reliable.

The wider map

The four programmes below hold the lab's current inquiries. The questions behind them reach further. The lab asks how biological and artificial systems acquire internal organisation, knowledge and agency, and what that explains about intelligent behaviour. Its deepest interest is internal cognition: what a model's hidden state holds, what a person processes without reporting it, and how either can be known from outside. Mechanistic interpretability is central to that work. It is one method among several, not the whole of it.

Subjects are what the lab studies. They fall into six groups:

  • Learning systems and mechanisms: machine-learning foundations, latent space and internal cognition, and memory and continual learning.
  • Minds, experience and adaptation: cognitive science and psychology, consciousness science, computational neuroscience, and cybernetics.
  • People, worlds and bodies: simulated people, world models and generated worlds, characters and agency, and robotics and embodied intelligence.
  • Perception, generation and meaning: perception and aesthetics, generative media, and language and meaning.
  • Judgement, conversation and evidence: how people form preferences and decisions, interviewing and qualitative research, and measurement.
  • Teaching and designed environments: adaptive teaching, human-AI interaction and the design of spaces and objects.

Lenses explain the subjects. They include mathematics and first-principles analysis, physics and dynamical systems, information theory, learning theory, causal and probabilistic inference, and philosophy of science. Psychology, neuroscience and cybernetics serve as lenses as well as subjects. Methods investigate a question: mathematical analysis, model construction, controlled experiments and interventions, human and neural evidence, and replication. Inquiries are bounded studies, each testing one question. Applications come after a mechanism is shown, not before.

Consciousness, subjective experience and biological neural systems are objects of inquiry here, not premises. A model that behaves like a person, a feature that can be decoded from its activations or a persona that reads as real does not show a shared brain mechanism, fidelity to a person or experience of any kind. Where the lab compares a model with a mind or a brain, it states the differences as well as the similarities.

Four programmes and why they belong together

Each of the four programmes asks the same question of a different stand-in: a model's own internal state and memory, the place where a model's knowledge lives, a model of a world, and a model of a person. In each, a model is checked near where it was built and used further away, and the open question is how far the resemblance carries. The designs aim to share methods where the question allows: held-out conditions placed at a stated distance from the building conditions, comparators matched in cost and information, and stop rules written before data are read.

The programmes also depend on one another. A simulated person who must remain the same person through an interview poses a memory problem. Telling whether a simulated respondent estimates or recalls requires evidence about what the model has absorbed. A model built for a specific problem has to settle where its knowledge lives, in its weights or in a store it consults, which the inquiry on specialised models tests directly and the memory inquiries inform.

Worlds and action

This programme asks whether a learned model of a world keeps what that world keeps: objects that persist out of view, actions that cannot be undone, obligations passed between agents, and the consequences of actions not taken.

A sequence model trained only on game moves formed an internal representation of the board that interventions could use to steer its output[15]. A hand-written world as a test for world models treats a written world as ground truth, types it into an event ledger and generates questions of five kinds: revisit, occlusion, irreversibility, handoff and counterfactual. It asks whether those questions separate a competent world model from a deliberately weak one where a fidelity score does not, and whether the grade transfers to a world written by someone else.

What a learned policy holds still asks the question from inside an agent. It disturbs a frozen robot policy in simulation, an open vision-language-action model such as OpenVLA[16], along specific scene directions to find the few variables the policy actively defends. It then tests whether that set predicts which unseen disturbances the policy survives, better than a probe that decodes the policy's internal state. One inquiry grades a world model directly; the other reads one from an agent's behaviour.

Models of people

This programme asks when a simulated person or population can stand in for real people, for which uses, and how that can be checked before anyone relies on it.

Generative agents that record, reflect on and retrieve their own experience produce believable behaviour in a sandbox[17]. Language models replicated three of four classic human-subject studies and showed a hyper-accuracy distortion in the fourth[18]. Against this, synthetic samples show too little variance and unstable estimates[9]. Interventions in simulated experiments can shift the latent attributes of the simulated population, so that treatment and control no longer describe the same people[19]. Simulated customers tend to be swayed by a sales agent's tone and drift from their persona specifications[20].

An interviewer that changes the person it studies takes up one form of that drift. A planner that rehearses interview questions on a simulated person can come to prefer questions that win by changing the person instead of learning about them. The design holds the person's fixed traits, permitted learning and revealed answers in separate channels, and scores each planner by how many hidden attributes it recovers within a fixed question budget. Echo or estimate? asks whether a model playing a survey respondent lands closer to a survey's published results the more it was exposed to that survey, beyond what each item's average answer explains. If it does, part of the agreement with published numbers is recall, and a validity check built on published surveys measures something other than simulation.

Representation and memory

This programme asks what a model holds inside and over time: which internal readings reflect what the model uses, what it retains, and whether it can foresee when what it retains will fail.

Interpretability has learned to distrust its own instruments. A linear probe shows what can be decoded from a layer's activations[21], but a flexible probe can learn the task itself, which control tasks were designed to expose[22]. A review of probing sets out its methodological shortcomings[23], and amnesic probing showed that behavioural conclusions cannot be drawn from probing alone: probing accuracy was not correlated with a property's importance to the task[24]. Interchange interventions test whether representations play the causal role of variables in a hypothesised model[25], and even a subspace intervention that changes behaviour as intended can do so through a dormant parallel pathway, giving an illusion of interpretability[26]. Recent work reports that representations a model is poised to verbalise, found with a technique called the Jacobian lens, show the functional properties of a global workspace[27]. Can the instruments for a model's workspace be trusted? asks whether three tests of such a workspace give repeatable readings on existing open-weight models, and whether they rank models alike, before anything is built on them.

Memory is where interiors meet time. Memory systems for language models manage tiers of context in the manner of an operating system[28]. A long-term memory benchmark, LongMemEval, tests information extraction, multi-session and temporal reasoning, knowledge updates and abstention, and reports a 30% drop in accuracy for commercial chat assistants and long-context models over sustained interactions[29]. Editing feed-forward weights can change a stored factual association[30], yet facts that should change as a consequence often do not[31]⁠[32]. When one fact changes measures how propagation to dependent facts falls, and damage to unrelated facts rises, as the number of dependents grows, for each editing method at a stated budget. Can a memory foresee what will break it? asks whether a trained model can predict which of two future disruptions will leave a stored memory unusable, better than a model given the same information without training on outcomes. Language models can partly predict whether they know an answer[33]; this inquiry asks for a forecast about the future of a memory instead. What a memory should keep apart asks which structure an agent's memory needs beyond retrieval, from typed records and temporal validity up to a light ontology. Each layer is kept only if it beats the simpler memory below it, at matched cost, on the failures it is meant to address.

Where knowledge lives

This programme asks when a model built for one domain does better than a general model given the same domain material, and where knowledge for a domain should live: in weights, in a store the model retrieves from, or in validators that check its output.

Continued pretraining on domain text improves performance on domain tasks[34]. Retrieval joins a model's parametric memory to an external store and was proposed partly to address provenance and updating[35]; in one comparison of ways to inject knowledge, retrieval consistently outperformed unsupervised fine-tuning[36]. Graph-structured retrieval addresses questions about a whole corpus[37], and ontology-driven self-training uses an existing ontology to organise a model's domain knowledge[38]. In the weights or beside them starts from a sceptical default: retrieval, validators and repair around a frontier model may match a trained specialist, so the specialist must show value beyond them. On one family of software-architecture judgement tasks, with fixtures to be published before any arm runs, it compares a frontier model alone, with plain or structured retrieval, and with structured retrieval plus validators and repair, against small models trained on plain text or on the domain's structure, and a frontier model that consults the structure-trained one.

From question to experiment

Each inquiry follows one path: a question, the rival explanations that could answer it, a controlled experiment that can tell them apart, a bounded finding, and the next question that finding raises. Five of the current inquiries show the path, and each page records its stage.

  • Internal representations. Can the instruments for a model's workspace be trusted? asks whether proposed readings of a language model's workspace repeat. One rival is a stable internal structure and the other is an artefact of how the lens was fitted. Repeated refits of the lens on published open-weight models, scored across seeds, test whether the readings repeat. Only if they do, a conditional second stage compares a trained model with a random-rotation control. Possible use: instruments that can carry claims about what a model holds.
  • Memory and revision. When one fact changes asks whether, as the number of facts that depend on one changed fact grows, editing methods update the dependents correctly and avoid wrong changes to unrelated facts. A ladder of real superseded facts, ordered by their number of dependents, tests this. Possible use: models that can be kept current.
  • World models and action. A hand-written world as a test for world models asks whether questions generated from one arc of a hand-written world separate a competent world model from a deliberately weak one in a way a fidelity score does not. Possible use: tests for simulators that judge robot policies.
  • Control and embodiment. What a learned policy holds still asks whether a frozen robot policy resists disturbance along specific scene directions, and whether that resistance predicts which held-out disturbances it survives better than a probe on its internal state does. Possible use: a description of a policy's limits in terms of what it holds still.
  • People and measurement. Echo or estimate? asks whether a language model playing a survey respondent lands closer to a survey's published results the more it was exposed to that survey. Possible use: a check on when simulated respondents are evidence about people.

A finding at the end of any of these is bounded by its design: the models, tasks and worlds it uses. It becomes a claim about minds, brains or people only through a further test built for that claim.

Working hypotheses

The lab works from the hypotheses below. Each could be wrong, and the inquiry that tests it states in advance what would show that.

  1. Fidelity is local, and its edge can be located. Agreement between a model and what it models falls with distance from the conditions it was checked under, along axes named beforehand, such as time out of view, unseen disturbances, surveys published after a training cut-off or the number of facts that depend on a changed one. If agreement does not fall along the named axes, the hypothesis fails.
  2. Surface fidelity does not predict fitness for use. Visual or average-level agreement ranks models differently from structured tests of what a use depends on[4]⁠[7]. The world-model questions would count against this in their domain if they ranked models the same way a fidelity score does.
  3. What a policy defends is not what a probe decodes. Resistance to disturbance predicts survival under unseen disturbances better than decoding does. If a decoding probe predicts as well, the claim fails.
  4. Part of the agreement between simulated and real respondents is recall. Agreement with published survey results rises with exposure to those surveys. A gap inside a pre-registered equivalence margin would count against this.
  5. A simulated person stays the same person only if built to. Holding fixed traits, permitted learning and revealed answers apart stops a planner from exploiting drift. A persona-aligned comparator that matched the separated design would falsify it.
  6. Instruments for a model's workspace need testing before use. Some proposed workspace tests will not reach the reliability bar. If all three do, the hypothesis is wrong and the tests can carry later claims.
  7. Memory failure is partly forecastable. A trained predictor forecasts which disruption will break a memory better than an equal-information model does.
  8. Richer memory has to earn its cost. Each richer memory layer stays only where it wins on the failure it targets.
  9. Edits break with fan-out. For a given editing method and budget there is a number of dependent facts above which an edit stops propagating. If propagation holds at every tested fan-out, the claim fails for that method.
  10. A specialist has to beat retrieval. Retrieval and validation close much of the gap between a general and a specialised model. A specialist that clearly beats the best retrieval configuration on held-out tasks would count against this.

Instruments

The programmes need instruments as much as findings. The inquiries specify the instruments below, and each specification is on the instruments page with its state: specified, being built, tested or released.

  • World ledger and question generator. A typed record of a written world's entities, obligations and caused events, a generator and adjudicator for the five classes of question, and a report that states the domain of use a model is graded for, with error bars.
  • Disturbance sweep for learned policies. Independent scene directions, graded disturbance mid-episode, a matched-effort control, a specificity score and fixed held-out conditions for testing survival.
  • Workspace measurement battery. Three tests (report-use agreement, ignition steepness and monitorability), frozen with calibration constants, test-retest figures and reference code, with a random-rotation control.
  • Interview rollout bench. A simulated interviewee whose fixed traits, permitted learning and revealed answers are kept apart, with a scorer the planner cannot see and a leakage check.
  • Exposure-graded survey benchmark. Construct-matched public survey instruments dated against training cut-offs, and a recall probe validated on documents known to be inside and outside training data.
  • Counterfactual memory bench. A test that copies a memory state, takes a forecast before either of two futures runs and scores it against exact outcomes; if the forecast passes, a controller that spends a fixed budget protecting the memory it flags.
  • Supersession benchmark. Changed facts and their dependents from a licensed public record, stratified by number of dependents, scored for propagation against unwanted change on matched controls.
  • Revisable memory runtime. A layered agent memory, tested as a cost-normalised ladder from full context to a light ontology, in which each layer stays only if it earns its cost.
  • Domain judgement model. A small model trained on a domain's structure, tested on one task family against the comparison conditions above under one scoring protocol, and kept only if it adds value over them.

The intention is to release each of these with documentation, code where it can be released, and its known limits, so that others can grade models the lab has not.

Systems are one kind of output among several. The lab counts theories and mathematical results, methods, instruments, datasets and papers as contributions in their own right. Applications such as simulated respondents, robot evaluation or teaching tools may follow from a shown mechanism. They are not the purpose of the work.

How the lab works

Work on a question starts from first principles. A design defines the phenomenon, its state, the rules by which it changes and the constraints on it, then derives what each rival explanation predicts. Mathematics can show what the data could identify at all. An elementary illustration, not a new result: if a model's output is a linear readout of its internal state, the output alone cannot fix that state's coordinates, because any invertible re-coordinatisation of the state, paired with a matching change to the readout, gives the same behaviour. That is one reason the lab pairs behavioural tests with interventions, and it does not treat a probe that decodes information as showing that the system uses it.

Human and biological evidence enters as published data and as constraints on what a model of a mind must explain. The lab runs no studies with people at this stage. It plans to test each instrument on planted cases whose answer is known before using it on a real question, and to check each key analysis with a separate implementation before reporting it.

Every inquiry is written as a design before it runs. The design states the question, the approach, the comparison to beat, what would count as an answer and what would end the attempt. Stop rules carry numbers where the design allows: a test-retest correlation of at least 0.8 for each workspace test, a minimum count of real cases with three or more dependents for the supersession benchmark, and pass and stop thresholds on a Brier-score gain for the memory forecast. These values will be fixed in a dated pre-registration for each design before any data is read, because pre-registration keeps the testing of predictions apart from explanations found afterwards[39].

Each design also states what it would not show. A good grade on one written world would not show that the grade carries to a world written by someone else. A recovery gain on synthetic profiles would not show that a planner works with real people. A policy result in simulation would not show what happens on a physical robot. These limits are written before the result, so that a result cannot quietly widen them.

Outcomes are classed in advance as pass, stop or inconclusive, and will be reported as classed. A failed test will be reported with what the failure rules out. Every page has a link for reporting an error, and corrections are listed with their dates on the updates page.

The lab plans to publish designs first, as now, then results with their limits, failed checks and comparison conditions, then code and data where licences allow. Material the lab may not redistribute will be described so that others can obtain it.

Where the work starts

Two inquiries are planned to run first. Both use open-weight language models, and one uses a synthetic corpus the lab writes itself. Each first stage settles something later work depends on and ends in a decision about whether to go on.

Can the instruments for a model's workspace be trusted? comes first. Its first stage needs only inference on open-weight models that already exist and no new data. A second stage, if it runs, trains small models. A test that misses the reliability bar is not used for any later claim, and that finding is itself a result.

Can a memory foresee what will break it? runs alongside it. Its data is a synthetic corpus of rule-and-exception families with exact labels from a rule engine. It compares a trained predictor with an equal-information model and two ablations on a small open-weight model. An early check stops the build if too few families can be verified and sealed or if the two futures rarely differ, and the held-out comparison has pass, stop and inconclusive outcomes written in advance.

Each of the other inquiries names on its page what it needs before it can start.

Status

This agenda sets out aims, designs and hypotheses, not findings. The inquiries are published as written designs, and results will be reported on their pages, with their limits, as each one runs.

References

  1. Ha, D., Schmidhuber, J. (2018). World Models. arXiv preprint 1803.10122.
  2. Hafner, D., Pasukonis, J., Ba, J., Lillicrap, T. (2025). Mastering diverse control tasks through world models. Nature 640(8059), 647–653.
  3. Bruce, J., Dennis, M., Edwards, A., Parker-Holder, J., Shi, Y., Hughes, E., et al. (2024). Genie: Generative Interactive Environments. arXiv preprint 2402.15391.
  4. Oefinger, C., Schäfer, F. R., Moller, K., Piccinini, M., Betz, J. (2026). Validate the Dream Before You Trust Its Verdict: Admissibility for World-Model Simulators. arXiv preprint 2607.07196.
  5. Argyle, L. P., Busby, E. C., Fulda, N., Gubler, J. R., Rytting, C., Wingate, D. (2023). Out of One, Many: Using Language Models to Simulate Human Samples. Political Analysis 31(3), 337–351.
  6. Park, J. S., Zou, C. Q., Kamphorst, J., Egan, N., Shaw, A., Hill, B. M., et al. (2024). LLM Agents Grounded in Self-Reports Enable General-Purpose Simulation of Individuals. arXiv preprint 2411.10109. Earlier title: Generative Agent Simulations of 1,000 People.
  7. Motamed, S., Culp, L., Swersky, K., Jaini, P., Geirhos, R. (2025). Do generative video models understand physical principles? arXiv preprint 2501.09038.
  8. Vafa, K., Chen, J. Y., Rambachan, A., Kleinberg, J., Mullainathan, S. (2024). Evaluating the World Model Implicit in a Generative Model. arXiv preprint 2406.03689.
  9. Bisbee, J., Clinton, J. D., Dorff, C., Kenkel, B., Larson, J. M. (2024). Synthetic Replacements for Human Survey Data? The Perils of Large Language Models. Political Analysis 32(4), 401–416.
  10. Santurkar, S., Durmus, E., Ladhak, F., Lee, C., Liang, P., Hashimoto, T. (2023). Whose Opinions Do Language Models Reflect? arXiv preprint 2303.17548.
  11. Balci, O. (1998). Verification, validation, and accreditation. Proceedings of the 1998 Winter Simulation Conference, 41–48.
  12. Sargent, R. G. (2008). Verification and validation of simulation models. Proceedings of the 2008 Winter Simulation Conference, 157–169.
  13. Cronbach, L. J., Meehl, P. E. (1955). Construct validity in psychological tests. Psychological Bulletin 52(4), 281–302.
  14. Campbell, D. T., Fiske, D. W. (1959). Convergent and discriminant validation by the multitrait-multimethod matrix. Psychological Bulletin 56(2), 81–105.
  15. Li, K., Hopkins, A. K., Bau, D., Viégas, F., Pfister, H., Wattenberg, M. (2022). Emergent World Representations: Exploring a Sequence Model Trained on a Synthetic Task. arXiv preprint 2210.13382.
  16. Kim, M. J., Pertsch, K., Karamcheti, S., Xiao, T., Balakrishna, A., Nair, S., et al. (2024). OpenVLA: An Open-Source Vision-Language-Action Model. arXiv preprint 2406.09246.
  17. Park, J. S., O'Brien, J. C., Cai, C. J., Morris, M. R., Liang, P., Bernstein, M. S. (2023). Generative Agents: Interactive Simulacra of Human Behavior. Proceedings of the 36th Annual ACM Symposium on User Interface Software and Technology (UIST 2023), 1–22.
  18. Aher, G., Arriaga, R. I., Kalai, A. T. (2023). Using Large Language Models to Simulate Multiple Humans and Replicate Human Subject Studies. International Conference on Machine Learning (ICML 2023).
  19. Lin, V., Yun, T., Matarić, M., Canny, J., Gretton, A., D'Amour, A. (2026). The Illusion of Intervention: Your LLM-Simulated Experiment is an Observational Study. arXiv preprint 2605.20767.
  20. Pruksachatkun, Y., Wan, Y., Chen, X., Chang, K.-W., Wu, C.-S. (2026). CustomerSim: Benchmarking and Aligning Multimodal Language Models as Retail User Simulators. arXiv preprint 2605.08334.
  21. Alain, G., Bengio, Y. (2016). Understanding intermediate layers using linear classifier probes. arXiv preprint 1610.01644.
  22. Hewitt, J., Liang, P. (2019). Designing and Interpreting Probes with Control Tasks. Proceedings of EMNLP-IJCNLP 2019, 2733–2743.
  23. Belinkov, Y. (2022). Probing Classifiers: Promises, Shortcomings, and Advances. Computational Linguistics 48(1), 207–219.
  24. Elazar, Y., Ravfogel, S., Jacovi, A., Goldberg, Y. (2021). Amnesic Probing: Behavioral Explanation with Amnesic Counterfactuals. Transactions of the Association for Computational Linguistics 9, 160–175.
  25. Geiger, A., Lu, H., Icard, T., Potts, C. (2021). Causal Abstractions of Neural Networks. Advances in Neural Information Processing Systems 34 (NeurIPS 2021).
  26. Makelov, A., Lange, G., Nanda, N. (2023). Is This the Subspace You Are Looking for? An Interpretability Illusion for Subspace Activation Patching. arXiv preprint 2311.17030.
  27. Gurnee, W., Sofroniew, N., Pearce, A., Piotrowski, M., Kauvar, I., Chen, R., et al. (2026). Verbalizable Representations Form a Global Workspace in Language Models. arXiv preprint 2607.15495.
  28. Packer, C., Wooders, S., Lin, K., Fang, V., Patil, S. G., Stoica, I., et al. (2023). MemGPT: Towards LLMs as Operating Systems. arXiv preprint 2310.08560.
  29. Wu, D., Wang, H., Yu, W., Zhang, Y., Chang, K.-W., Yu, D. (2025). LongMemEval: Benchmarking Chat Assistants on Long-Term Interactive Memory. International Conference on Learning Representations (ICLR 2025).
  30. Meng, K., Bau, D., Andonian, A., Belinkov, Y. (2022). Locating and Editing Factual Associations in GPT. Advances in Neural Information Processing Systems 35 (NeurIPS 2022).
  31. Cohen, R., Biran, E., Yoran, O., Globerson, A., Geva, M. (2024). Evaluating the Ripple Effects of Knowledge Editing in Language Models. Transactions of the Association for Computational Linguistics 12, 283–298.
  32. Zhong, Z., Wu, Z., Manning, C. D., Potts, C., Chen, D. (2023). MQuAKE: Assessing Knowledge Editing in Language Models via Multi-Hop Questions. Proceedings of EMNLP 2023, 15686–15702.
  33. Kadavath, S., Conerly, T., Askell, A., Henighan, T., Drain, D., Perez, E., et al. (2022). Language Models (Mostly) Know What They Know. arXiv preprint 2207.05221.
  34. Gururangan, S., Marasović, A., Swayamdipta, S., Lo, K., Beltagy, I., Downey, D., et al. (2020). Don't Stop Pretraining: Adapt Language Models to Domains and Tasks. Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics (ACL 2020), 8342–8360.
  35. Lewis, P., Perez, E., Piktus, A., Petroni, F., Karpukhin, V., Goyal, N., et al. (2020). Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks. Advances in Neural Information Processing Systems (NeurIPS 2020).
  36. Ovadia, O., Brief, M., Mishaeli, M., Elisha, O. (2024). Fine-Tuning or Retrieval? Comparing Knowledge Injection in LLMs. Proceedings of EMNLP 2024, 237–250.
  37. Edge, D., Trinh, H., Cheng, N., Bradley, J., Chao, A., Mody, A., et al. (2024). From Local to Global: A Graph RAG Approach to Query-Focused Summarization. arXiv preprint 2404.16130.
  38. Liu, Z., Gan, C., Wang, J., Zhang, Y., Bo, Z., Sun, M., et al. (2025). OntoTune: Ontology-Driven Self-training for Aligning Large Language Models. Proceedings of the ACM Web Conference 2025, 119–133.
  39. Nosek, B. A., Ebersole, C. R., DeHaven, A. C., Mellor, D. T. (2018). The preregistration revolution. Proceedings of the National Academy of Sciences 115(11), 2600–2606.

Cite as

Far Akin Research Labs (2026). Research agenda. Version 1.0. Far Akin Research Labs. https://farakin.vercel.app/agenda

BibTeX
@misc{farakin2026-agenda,
  title = {{Research agenda}},
  author = {Far Akin Research Labs},
  howpublished = {\url{https://farakin.vercel.app/agenda}},
  year = {2026},
  month = nov,
  note = {Version 1.0},
  url = {https://farakin.vercel.app/agenda}
}

Noticed an error or a missing source? Send a correction