Skip to content
TheChatArchiveThe Chat Archive
GeminiGemini (model unspecified)· GoogleHugging Face ShareLM (browser-extension contributions)

please look for papers (conference or journals) that talk about how humans follow…

  1. User

    please look for papers (conference or journals) that talk about how humans follow specific task instructions, for example, whether the following sentence is grammatical correct or not

  2. Gemini Assistant

    Here's a research plan for that topic. If you need to update it, let me know! Human Instruction Following Research Research Websites(1) Search for academic papers (journals or conference proceedings) in cognitive science and human factors discussing the general cognitive processes involved in humans encoding and executing explicit task instructions. (2) Find research in psycholinguistics or cognitive neuroscience specifically examining how humans perform linguistic judgment tasks, such as assessing grammatical correctness, based on task instructions. (3) Investigate studies that analyze how the complexity, ambiguity, or framing of instructions affects human adherence, task performance, and decision-making in rule-based contexts. (4) Look for literature that explores the role of cognitive resources, such as working memory and attention, in the ability of humans to follow multi-step or conditional instructions. (5) Identify papers that contrast human performance in rule-based tasks when instructions are explicitly provided versus when rules are implicitly learned or inferred. (6) Search for comprehensive review papers or meta-analyses that synthesize current understanding and models of human instruction-following behavior across different experimental paradigms. (7) Examine research that analyzes common types of errors, deviations, or biases that arise when humans execute a repetitive task according to detailed, specific instructions. (8) Find theoretical frameworks or computational models that propose mechanisms to explain how humans encode, maintain, and execute specified task instructions in cognitive architecture.MoreAnalyze ResultsCreate ReportReady in a few mins Edit plan Start research

  3. User

    Start research

  4. Gemini Assistant

    I've completed your research. Feel free to ask me follow-up questions or request changes.Human Instruction Following ResearchCompleted21 Nov, 10:57Open

  5. User

    could focus more on psychology work and whether we see the same effects in language models?

  6. Gemini Assistant

    Here's a research plan for that topic. If you need to update it, let me know! LLM Instruction Following vs. Psychology Research Websites(1) Identify recent academic papers in cognitive psychology and behavioral science that isolate and detail specific human cognitive biases (e.g., framing, ambiguity effect, recency bias) during explicit instruction following. (2) Find comparative studies that explicitly analyze and contrast human and Large Language Model (LLM) adherence to task instructions, focusing on systematic deviations and error patterns. (3) Investigate research on LLM instruction following, focusing on how prompt complexity, length, and specificity affect model output and adherence to the desired format or goal. (4) Search for papers that map human cognitive constraints, such as working memory load or executive control failures, to analogous phenomena observed in LLM performance (e.g., context window limits, instruction tuning failure). (5) Specifically look for academic literature detailing cognitive biases in LLMs (e.g., priming, anchoring, order effects) that mirror human instruction-following biases. (6) Find theoretical frameworks or papers discussing the architectural similarities or differences between human cognitive models (like ACT-R/Soar) and transformer-based LLMs in terms of translating natural language instructions into executable steps.MoreAnalyze ResultsCreate ReportReady in a few mins Edit plan Start research

  7. User

    Start research

  8. Gemini Assistant

    I've completed your research. Feel free to ask me follow-up questions or request changes.LLM Instruction Following vs. PsychologyCompleted21 Nov, 11:06Open

  9. User

    is there evidence that instructions make humans look on task specific properties. for example when we ask whether given sentence is grammatical correct, humans compute more information about the grammatical correctness when reading the specific sentence

  10. Gemini Assistant

    Yes, there is clear evidence within psycholinguistics research that the explicit instructions given for a task like a Grammaticality Judgment Task (GJT) fundamentally change the type and depth of cognitive processing applied to the sentence stimulus.When humans are instructed to determine whether a given sentence is "grammatically correct," they are forced to shift from relying on their natural, implicit linguistic knowledge (competence) to engaging their conscious, rule-based metalinguistic ability.Here is how instructions affect task-specific processing:1. Shift from Intuition to Conscious AnalysisThe primary impact of the instruction is the shift in cognitive strategy. The goal of a GJT is typically to measure a native speaker's unconscious linguistic competence. However, when explicit instructions or rules regarding the relevant grammatical structures are provided, participants often utilize this information as an "efficient shortcut".   Instead of accessing their deeply ingrained linguistic intuition, they rely on the explicit application of the learned rules to generate the judgment. This process involves the conscious reflection upon and analysis of the language's formal structure. In essence, the instruction primes the system to treat the sentence not as natural language to be comprehended, but as a formal problem to be solved using a specific set of task rules.   2. Enhanced Computational Load for Grammatical FeaturesThe task instruction, by specifying the target property ("grammatical correctness"), compels the cognitive system to allocate resources to processes associated with analysis of linguistic knowledge and control of linguistic processing.Neuroscientific studies indicate that when sentences are grammatically demanding or complex, brain regions associated with increased processing load, such as the left inferior frontal gyrus (IFG), exhibit enhanced activation. This suggests that the act of judging a sentence for correctness, particularly when the grammatical features are challenging, requires increased computational effort focused directly on those features.Furthermore, research on instruction type suggests that the instructional condition itself modulates the resultant depth of processing. In one study, providing a direct metalinguistic explanation (explicit instruction) did not lead to the predicted cognitive shifts, possibly because it encouraged a more superficial, rule-checking approach rather than the deep, meaning-focused processing that was desired.Therefore, the instruction does make humans compute more information about the task-specific properties, but this computation is often a resource-intensive, metalinguistic performance—a process of deliberately applying an explicit, task-defined rule set —rather than the passive measurement of implicit linguistic competence. The concern in psycholinguistics is that this instructional bias acts as a strategic confound, compromising the validity of the measure.

  11. User

    is there evidence that instructions make humans look on task specific properties. please focus not on linguistic correct but rather on semantic task such as coherent reasoning, world knowledge, or theory of mind

  12. Gemini Assistant

    This is an excellent distinction. While the previous discussion focused on how instructions activate conscious, metalinguistic rule-checking for linguistic tasks, the research on semantic, reasoning, and social tasks demonstrates that instructions function as a powerful top-down control mechanism that specifically focuses cognitive resources on the required conceptual domain.When the task is shifted from "Is this sentence legal?" to "Does this situation make sense?" or "What is the agent thinking?", the instructions dictate which specific task properties—such as coherence or world knowledge—must be prioritized and computed.Here is the evidence that task instructions modulate the depth and specificity of semantic, world knowledge, and social reasoning processes:1. Modulation of Semantic and World Knowledge ProcessingIn complex tasks requiring comprehension and reasoning, instructions establish a proactive control state within the cognitive system. This state enables an "early selection" process where goal-relevant information is actively maintained and used to bias the system in a goal-driven manner.Controlling the Depth of Semantic Processing: The task environment, dictated by the instruction, can significantly modulate how deeply semantic information is processed. Research has found that subliminal semantic processing effects (the unconscious activation of word meaning) are significantly reduced when individuals are engaged in a high-demand, hard primary task compared to an easy one. This suggests that high cognitive load, often imposed by complex instructions or tasks, consumes attentional resources necessary for deep semantic processing, attenuating the automatic computation of meaning. Conversely, if the instruction is simple and requires a high-level semantic decision, cognitive resources are freed up and dedicated to that goal.Top-Down Attention for Task Fulfillment: When researchers explicitly instruct subjects to direct their attention internally or externally, this manipulation immediately necessitates top-down attentional control to fulfill the specific task demands. For tasks requiring world knowledge or coherent reasoning, the instruction forces the system to retrieve and integrate chunks of factual knowledge that align with the required output criteria.2. Instructing Theory of Mind (ToM) ActivationTheory of Mind (ToM) is the ability to interpret the behavior of others by inferring their mental states, such as beliefs, goals, or desires. Instructions targeting this domain influence performance by guiding attention to specific features related to mental state tracking.Spontaneous vs. Instructed Belief Tracking: Some evidence suggests that humans possess a spontaneous ToM mechanism; for example, a study using eye-tracking showed that even when one group of participants was given no instructions and another was instructed to track the mere position of a ball, all groups' eye-movement patterns were consistent with spontaneous belief analysis. This implies a baseline tendency for humans to compute mental states automatically.Functional Dissociation: However, other research confirms that task instructions modulate the specialized brain regions involved in ToM. In the False-Belief Task (FBT), the instructions defining the specific task phase (e.g., belief-formation versus outcome) cause a functional dissociation in neural mechanisms. Regions like the temporoparietal junction (TPJ) show heightened responses during the phase specifically focused on coding beliefs, distinguishing this process from mere tracking of non-mental state information. The instruction, therefore, directs attentional resources to the task-specific computation of "belief" rather than general stimulus tracking.3. Allocation of Deliberate Reasoning EffortFor tasks requiring complex reasoning (deductive, inductive, or abductive), the instruction serves to define the difficulty and scope of the problem, directly influencing the amount of mental effort expended.Adaptive Cognitive Effort: Human performance scales naturally with task difficulty; people apply greater cognitive effort (the extent of mental resources required) to solve complex tasks. Instructions effectively set the perceived complexity of the task. If an instruction implies a high degree of difficulty or uncertainty, the individual allocates more deliberate reasoning effort to solve it.Influence of Complexity: Studies show that when tasks are deemed highly complex, they lead to significantly longer duration and greater self-reported cognitive effort compared to simple tasks. Therefore, the instruction, by defining the task as demanding coherent reasoning or extensive world knowledge retrieval, ensures that the system dedicates maximum resources to computing those complex, task-specific properties.

  13. User

    is there evidence that instructions make humans look on task specific properties. please focus not on linguistic correct but rather on semantic task such as coherent reasoning, world knowledge, or theory of mind

  14. Gemini Assistant

    That is an essential clarification, as the mechanism by which instructions influence semantic and reasoning tasks is related to—but distinct from—the metalinguistic rule-checking activated by grammatical tasks.For tasks involving coherent reasoning, world knowledge retrieval, and social cognition (Theory of Mind), the instruction acts as a top-down control signal that specifically dedicates high-level cognitive resources toward computing the required conceptual properties.Here is the evidence demonstrating how task instructions prioritize these specific, non-linguistic computations:1. Activating and Directing World Knowledge Retrieval (Proactive Control)Instructions fundamentally function by establishing a proactive control state that biases the cognitive system toward goal-relevant information. For a semantic or knowledge-based task, this means the instruction dictates which conceptual domain the human must retrieve and maintain.Goal-Driven Bias: Proactive control enables a form of "early selection," where goal-relevant information—such as necessary world knowledge or contextual facts—is actively maintained in a sustained manner, optimally biasing cognitive systems in a goal-driven way. If a task requires reasoning about an economic scenario, the instruction to focus on "economic resilience" or "technological leadership" forces the retrieval of related domain-specific knowledge chunks rather than generic facts.   Top-Down Attention: Instructions that manipulate the external or internal focus of attention compel the system to utilize top-down attentional control to meet the specific requirements of the task. The system is instructed to prioritize retrieving and integrating the necessary world knowledge to produce a coherent output.2. Modulating the Depth of Semantic and Reasoning ProcessingThe demands set by the task instructions directly influence how much and how deeply semantic information is processed.Instructional Complexity and Effort: For tasks demanding coherent reasoning, the complexity defined by the instruction is a direct determinant of the cognitive effort allocated. Studies comparing task execution show that consistently complex tasks lead to significantly longer duration and greater self-reported cognitive effort than simple ones. Thus, instructing a person to perform a "complex reasoning task" guarantees the computation of task-specific information will be more resource-intensive and thorough.   Attenuation of Automatic Processing: Conversely, when instructions impose a very difficult primary task (high cognitive load), subliminal semantic processing effects (the unconscious, automatic computation of meaning) are significantly reduced. This implies that when deliberate, reasoning-heavy instructions are given, the computational resources are deliberately channeled away from automatic semantic processing and toward the conscious, strategic processing required by the instruction.3. Dedicated Processing for Theory of Mind (ToM)For social cognition tasks, the instruction focuses processing on the specific mental state to be analyzed, demonstrating the precise tuning of cognitive systems.Functional Dissociation in Neural Networks: Research using the False-Belief Task (FBT), a standard test of Theory of Mind (ToM), shows that explicit instructions functionally dissociate neural mechanisms associated with mental state tracking.Coding Beliefs: Specifically, the temporoparietal junction (TPJ), a core region in the ToM network, shows heightened responses during the phase of the task explicitly focused on coding beliefs, distinguishing this process from general non-mental state tracking. Even though some ToM processes may occur spontaneously regardless of instruction, the instruction defining the task (e.g., focusing on the agent's belief formation or the outcome of that belief) forces the dedication of resources to the explicit computation of the task-relevant mental state.

Responses were produced in different, uncontrolled contexts (system prompts, settings, prior turns). This is not a benchmark.