On November 12, 2026, researchers from Northeastern University, Parallax, the University of Copenhagen, and the University of Oxford released an arXiv preprint demonstrating that transformer large language models represent explicit, causal task goals in a sparse subset of internal attention heads. Led by Arnab Sen Sharma and David Bau, the study isolates how autoregressive models formulate multi-step plans and verbalize upcoming sub-goals prior to token emission, offering critical empirical evidence for mechanistic interpretability and safety auditing across advanced AI systems.
Preprint Overview: Tracing Internal Goal Representations
The preprint, titled ‘Goal Encoding in LLMs’ (arXiv:2611.17297), addresses a central unresolved debate in artificial intelligence: whether the goal-directed behavior displayed by language models reflects genuine internal intent representations or merely surface-level statistical mimicry. Grounding their inquiry in philosophical planning frameworks, the research team defines a goal as a functional internal state exhibiting three specific properties: pre-commitment prior to action execution, plasticity across diverse realization environments, and temporal persistence throughout execution.
As an unreviewed preprint hosted on the arXiv computer science repository, the manuscript has not yet undergone formal peer review. While scholarly communication workflows routinely examine preprint findings before journal publication, researchers and research offices should note that the paper’s experimental conclusions reflect author-reported evaluations rather than certified benchmarks. Readers interested in how preprints transition to archival literature can consult our overview of the peer-reviewed process.
Localization, Decodability, and Causal Steering
To evaluate goal formation empirically, the authors constructed a multi-step code-generation benchmark requiring models to solve multi-stage programming problems involving distinct sub-tasks. Analyzing open-weights models including Qwen-2.5-Coder-32B-Instruct alongside larger architectures, the investigators localized internal representations of upcoming sub-goals to a surprisingly sparse subset of attention heads—accounting for roughly 4.8% of heads in the 32-billion parameter model.
Using a probing technique designated GoalScope, the researchers demonstrated that activations at these identified heads can be directly decoded into natural language descriptions of the model’s intended next sub-task before code implementation begins. For example, during algorithmic problem-solving, internal states decodably signaled intents such as verifying whether an input string is a palindrome prior to generating the relevant logic.
Crucially, causal intervention experiments confirmed that these head activations do not merely correlate with generated tokens; they actively steer execution. When the researchers transplanted goal-head activations into distinct execution environments—such as transferring a sub-goal state from a Python context into a JavaScript or C++ prompt—the target model executed the transplanted operation using syntax and token structures appropriate for the new language. This demonstrates genuine functional plasticity, establishing that internal goal representations operate at an abstract task level rather than as rigid token predictions.
The Non-Recurrent Goal-Seeking (NRGS) Mechanism
A fundamental theoretical puzzle highlighted by the authors concerns temporal persistence: standard decoder-only transformer architectures lack recurrent hidden states to propagate goals across arbitrary sequence lengths. How, then, can a depth-bounded model sustain goal pursuit across hundreds of generated tokens without experiencing representational fade?
To resolve this paradox, the authors propose and validate a Non-Recurrent Goal-Seeking (NRGS) mechanism. Rather than maintaining an evolving state vector across token steps, the model dynamically computes goal pursuit at critical pivot tokens. Under this algorithm, the model repeatedly re-reads the overall task specification from the prompt tokens, compares that specification against the progress accumulated in the generated context, and re-derives the next pending sub-goal.
Targeted causal ablations and activation patching at pivot tokens confirmed this dynamic. When researchers disrupted the model’s ability to attend to prompt specifications or masked intermediate progress representations, goal execution broke down, whereas interventions between pivot steps left ongoing execution intact. Furthermore, the authors observed that the same goal heads verbalize intent in open-ended settings where sub-goals were never explicitly enumerated in the user prompt, suggesting a generalized internal mechanism for autonomous task decomposition.
Implications for AI Safety Auditing and Research Governance
The identification of decodable, steerable goal representations carries immediate implications for research administration, academic computing infrastructure, and AI safety auditing. In institutional settings where autonomous coding agents and general-purpose reasoning models assist in scientific research, verifying that an agent’s internal intent aligns with user instructions represents an essential compliance safeguard.
From a regulatory perspective, mechanistic interpretability techniques provide a technical foundation for auditing high-capability systems under evolving compliance frameworks, such as general-purpose AI obligations under the EU AI Act. As developers of frontier foundation models face heightened requirements for systemic risk evaluation, the ability to inspect an agent’s internal goal state prior to execution offers an auditing layer that external output filtering cannot provide.
For university research offices and data stewards managing institutional AI deployments, the research highlights the importance of transparent evaluation tooling. The authors have published their complete experimental codebase, intervention scripts, and multi-step evaluation datasets at goals.baulab.info, expanding the public ecosystem of AI and machine learning research outputs. Integrating internal goal monitoring into safety pipelines will be a focal area for institutional governance, reinforcing broader efforts documented across our frontier AI safety and governance resources.
Sources
Arnab Sen Sharma, Eric Todd, Sheridan Feucht, Andy Arditi, Rohit Gandikota, Natalie Shapira, Gabriele Sarti, Constanza Fierro, Lisa Schut, and David Bau. ‘Goal Encoding in LLMs.’ arXiv:2611.17297, November 12, 2026. Checked 2026-11-14.
Bau Lab, Northeastern University. ‘Goal Encoding in Large Language Models Project Repository and Benchmark.’ https://goals.baulab.info, November 2026. Checked 2026-11-14.
