Human-like Temporal Abstraction

• Human-like temporal abstraction is a framework for modeling how humans represent time through hierarchical event segmentation, deictic reference points, and logarithmic compression.
• It utilizes computational techniques such as deictic temporal frameworks, log-linear distance metrics, and hierarchical encoders to predict behavior and efficiently plan actions.
• The approach integrates rapid atomic events with extended goal planning, enabling more robust real-time predictions and long-horizon strategies in AI systems.

Human-like temporal abstraction refers to the cognitive and algorithmic ability to represent, reason, and predict across multiple granularities and distances in time, mirroring the ways in which humans understand temporal relations, segment activities, and plan over both short and extended temporal horizons. Across contemporary research, temporal abstraction encompasses hierarchical event structures, deictic reference-point reasoning, and efficient latent-space planning through options or temporally extended skills. These capabilities underpin not just robust event anticipation and planning but also reflect core aspects of human temporal cognition such as reference point dependence, multi-clock reasoning, and logarithmically compressed timelines.

1. Theoretical Foundations: Temporal Frames, Hierarchies, and Deixis

Human temporal abstraction is anchored in multiple, interrelated theoretical constructs:

• Temporal Frame of Reference (t-FoR): Drawing from cognitive linguistics and temporal metaphors, a t-FoR specifies how a target event (TE) is located relative to a reference point (RP) and an origo (O), mirroring spatial reference frames. Human cognition distinguishes between an absolute, fixed RP (B-series) and a deictic, shifting RP (A-series), where “now” traverses the timeline and orients all temporal judgments (Zhang et al., 19 Oct 2025).
• Deictic Temporal Cognition: Human temporal metaphors—Moving-Ego (“We are approaching Tuesday”), Moving-Time (“Tuesday is approaching”)—emphasize the central role of a subjective “now” which defines psychological distance and similarity to temporal events. The deictic perspective induces a compressive, asymmetric profile in perceived temporal similarity: present > near past > near future > remote (Zhang et al., 19 Oct 2025).
• Hierarchical Temporal Abstraction: Human activities are natively structured as event hierarchies $\bigl\{A^1, A^2,\dots,A^L\bigr\}$, with each level $A^l$ offering a coarser or finer segmentation of the activity stream. Lower levels tick rapidly over atomic steps; higher levels represent intentions or tasks stretched over long spans. This nested partitioning supports robust planning and dampens error propagation (Morais et al., 2020).
• Temporal Compression Laws: Empirical evidence and modeling indicate that humans, and by convergence, LLMs, naturally compress perceived temporal distance logarithmically as points recede from a subjective reference—reflecting the Weber-Fechner law for time (Li et al., 21 Jul 2025).

2. Computational Frameworks and Metrics for Temporal Abstraction

Contemporary approaches formalize temporal abstraction using a range of mathematical and algorithmic techniques:

• Similarity Judgment Tasks under Deictic t-FoR: The TUuD framework operationalizes human-like temporal abstraction in LLMs through similarity ratings $s_{\rm LLM}\in[0.00,1.00]$ between a moving “now” ($t_{\rm now}$) and a set of target events ($t_x$). Temporal similarity curves are evaluated as a function of signed distance $\mathrm{dist} = t_x - t_{\rm now}$; a human-like model exhibits a peaked profile at $\mathrm{dist}=0$ decaying with $|\mathrm{dist}|$ (Zhang et al., 19 Oct 2025).
• Reference-Log-Linear Distances and Temporal Compression: In “The Other Mind,” perceived distance $d_{\text{ref}}(i,j)$ is modeled as $|\log(|i-R|+1)-\log(|j-R|+1)|$, where $R$ is the subjective present, capturing the logarithmic compression empirically found in both human and LLM similarity judgments. Goodness-of-fit $R^2$ quantifies the model's adherence to this compressed temporal landscape (Li et al., 21 Jul 2025).
• Multi-Resolution Behavioral Prediction: Combining high-resolution temporal indicators (recent goal-directed movements) and low-resolution features (coarse, long-term state-history) yields superior real-time behavior prediction, reflecting the cognitive integration of fast-changing detail with accumulated context (Zhang et al., 2022).
• Hierarchical Encoder–Refresher–Anticipator (HERA): In hierarchical action prediction, HERA models encode and roll out multi-level event sequences, with message passing between levels and explicit “refreshers” bridging incomplete events at test time. Predictive objectives combine label cross-entropy and duration mean-squared error across all levels (Morais et al., 2020).

3. Mechanistic and Representational Levels of Analysis

Mechanistic investigations reveal how human-like temporal abstraction emerges across layers and representations:

• Neuron-Level Specialization: A small fraction ($\sim$0.7–1.7%) of feed-forward network neurons in large LLMs are highly selective for temporal tokens, particularly “year” representations. These neurons exhibit minimal activation at the subjective reference point and encode a logarithmic response profile—mapping precisely onto empirical compression curves (Li et al., 21 Jul 2025).
• Layer-Wise Abstraction: Early model layers encode numeric magnitude (log-linear distance), whereas deeper layers reconstruct temporally referenced abstraction (reference-log distance). This hierarchical construction is consistent with a process that transitions from primitive counting to fully anchored temporal reasoning as depth increases (Li et al., 21 Jul 2025).
• Corpus-Level Temporal Structure: Semantic embedding spaces trained on large corpora inherently reflect a non-linear, compressed distribution of temporal items (e.g., years), with dense clustering of remote events and spreading near the present. These biases furnish models with raw material for emergent temporal abstraction (Li et al., 21 Jul 2025).

4. Hierarchical and Option-Based Temporal Abstractions in Planning

Hierarchical models and option-based frameworks further instantiate human temporal abstraction in both language and control:

• Multi-Clock Hierarchies for Action Prediction: Human-like temporal abstraction is achieved by partitioning actions into hierarchies—slower “clocks” for high-level goals, faster ones for fine detail. Models such as HERA avoid the error accumulation and inflexibility of single-scale predictors by enabling long-range structure to influence near-term planning, enforcing dependency constraints between levels, and segmenting predictions over multiple abstraction granularities (Morais et al., 2020).
• Option-Induced Abstract MDPs and Variational Homomorphisms: Temporally extended actions (“options”) are formalized within Option-Induced Hierarchical MDPs and learned as latent embeddings via the Variational Markovian Option Critic (VMOC). The theoretical machinery of continuous MDP homomorphisms (pair $(f,g)$) guarantees that planning in the abstracted latent space preserves optimality and enables efficient composition and reuse of temporally abstract skills. Supervised fine-tuning on human chain-of-thought data initializes these options as compressed reasoning units (Li et al., 22 Jul 2025).
• Latent Option Reasoning for Efficient Inference: Rather than requiring explicit enumeration of every decision step, models can “think” in a succinct sequence of options (typically $\sim$6 discrete abstractions per problem), providing both interpretability and computational efficiency compared to explicit chain-of-thought generation, with empirical performance matching or surpassing baselines on language and control tasks (Li et al., 22 Jul 2025).

5. Empirical Results and Quantitative Profiles

Extensive empirical investigations provide quantitative and qualitative signatures of human-like temporal abstraction:

• LLM Temporal Similarity Profiles: All evaluated LLMs in TUuD produce peaked similarity at the present ($s_{\text{avg}}(0) \approx 1.00$) with monotonic decay: $s_{\text{avg}}(\pm 10)\approx 0.75 - 0.85$, $s_{\text{avg}}(\pm 50)\approx 0.45 - 0.55$, and rapid breakdown past $\pm 50$ years, reflecting the limits of local temporal abstraction (Zhang et al., 19 Oct 2025).
• Compression Law Fit: The reference-log-linear model outperforms alternatives for year-year similarity across scaling experiments ($R^2_{\text{ref}} = 0.63$ for Qwen-14B, $0.58$ for Llama-70B; control tasks with numbers alone are instead captured by simple log-linear distance) (Li et al., 21 Jul 2025).
• Hierarchical Prediction Accuracy: In action anticipation, HERA with hierarchical architecture, bidirectional messaging, and refresh modules outperforms single-level RNNs and other baselines, especially in long-horizon or fine-scale metric settings ($[email protected]$ evaluation) (Morais et al., 2020).
• Behavioral Prediction with Multi-Resolution Features: Integrating high- and low-resolution temporal cues in the joint learning framework improves predictive accuracy, with best models achieving $0.72$ (easy), $0.71$ (medium), and $0.68$ (hard) accuracy for goal prediction, and consistently high performance on victim type—substantially higher than single-resolution models (Zhang et al., 2022).

6. Limitations, Open Questions, and Directions for Advancement

Despite progress, current models manifest several fundamental limitations relative to human temporal abstraction:

• Locality and Instability Beyond Near-Term: LLMs’ performance degrades beyond the $\pm 50$ year window; distant events are handled inconsistently and may even exhibit spurious similarity “bumps,” indicating a lack of robust global timeline modeling (Zhang et al., 19 Oct 2025).
• Shallow Pattern Exploitation: Reliance on explicit context tokens (e.g., timestamps) suggests that even advanced models may deploy shallow pattern-matching, lacking an internalized and persistent representation akin to the human mental timeline (Zhang et al., 19 Oct 2025).
• Alignment Risks and “Alien Cognition”: Emergent LLM cognition may parallel, but is never isomorphic to, human temporal abstraction; non-anthropomorphic, unanticipated cognitive structures can arise from fundamentally different architectures and training regimes. Oversight in alignment must probe these internal world constructions, not only outputs (Li et al., 21 Jul 2025).
• Scalable Abstraction Discovery: While current methods fix hierarchy depth or abstract action libraries, automated discovery, scaling to deeper or more irregular hierarchies, and continuous abstraction remain active directions (Morais et al., 2020).

Potential advances include explicit temporal embedding or ordinal pretraining, hybrid symbolic–neural architectures with dynamically updating “now,” and curricula demanding multi-horizon planning or simulation-based experience for more robust and generalizable temporal reasoning (Zhang et al., 19 Oct 2025).

---

Key references:

• “Temporal Understanding under Deictic Frame of Reference” (Zhang et al., 19 Oct 2025)
• “The Other Mind: How LLMs Exhibit Human Temporal Cognition” (Li et al., 21 Jul 2025)
• “Learning to Abstract and Predict Human Actions” (Morais et al., 2020)
• “Using Features at Multiple Temporal and Spatial Resolutions to Predict Human Behavior in Real Time” (Zhang et al., 2022)
• “Learning Temporal Abstractions via Variational Homomorphisms in Option-Induced Abstract MDPs” (Li et al., 22 Jul 2025)