Token-Faithful Trajectory Reconstruction

• Token-Faithful Trajectory Reconstruction ensures every reconstructed token corresponds directly to original or sampled input, enhancing interpretability and unbiased learning.
• Methods like Polar for RL and PLMTrajRec for geospatial data help maintain fidelity in model emissions, crucial for rigorous scientific analysis.
• Key benefits include improved GPU utilization, accuracy in path recovery, and accelerated computational speed in token attribution frameworks.

Token-faithful trajectory reconstruction refers to methods and frameworks that reconstruct trajectories—spatiotemporal, reasoning, or agentic—such that every output token, and all gradient attributions or causal flows, transparently correspond to original or sampled tokens from source data, model rollouts, or environmental observations. This property has become central for interpretability, reinforcement learning (RL), sequence modeling, and trajectory interpolation in domains where fidelity to actual model emissions or data tokens is crucial for scientific rigor, unbiased learning signals, or downstream reliability.

1. Formal Definitions and Scope

A trajectory is generally a sequence of states and actions, possibly with intermediate or environmental tokens representing context, reasoning steps, or task metadata. A trajectory is token-faithful if, after reconstruction:

• Every token marked as "trainable" (e.g., used for gradient calculation or attribution) is exactly the one emitted by the behavior policy (for RL) or model (for generation/analysis).
• Any tokens injected by post-processing, harness rerendering, or control logic are clearly masked out for learning and evaluation.
• For RL, reconstruction provides a sequence $(z_1, z_2, \ldots, z_L)$ of tokens, with state $s_t = (z_1, \ldots, z_{t-1})$, action $a_t = z_t$, loss mask $m_t \in \{0,1\}$, and log-probabilities $\ell_t$ such that for all $t$ with $m_t = 1$, $\ell_t$ is the log-prob assigned by the policy to $a_t$ when sampled (Xu et al., 22 May 2026).
• In spatiotemporal settings, each reconstructed position, segment, or curve point token arises from an input, output, or imputed value in the model's tokenization pipeline, maintaining direct traceability (Zhang et al., 2024, Wei et al., 2024).

2. Motivation: Unbiased Learning and Explanation

Token-faithful reconstruction is foundational for several reasons:

• Unbiased RL Gradients: The RL policy gradient estimator, $$\nabla_\theta \mathbb{E}_{a \sim \pi_\theta}[R] = \mathbb{E}_{a \sim \pi_\theta}[\nabla_\theta \log \pi_\theta(a) R]$$, is only unbiased if $s_t = (z_1, \ldots, z_{t-1})$0 is computed for exact tokens as sampled. Retokenization drift or free-text re-encoding introduces bias and high variance, destabilizing optimization and masking learning signals (Xu et al., 22 May 2026).
• Causal Attribution: In LLM interpretability, spurious injection of intermediate tokens or mass loss in attribution chains leads to explanations that are not faithful to actual input-output causality. Token-faithful schemes, such as recursive attribution with span-aggregation, preserve genuine propagation of causal signal (Pan et al., 2 Feb 2026).
• Data Integrity: In trajectory recovery from sparse or noisy input, fidelity ensures that imputed or interpolated outputs accurately reflect plausible real-world segments, as opposed to statistical artifacts arising from round-trip transformations or prompt-induced hallucinations (Wei et al., 2024).

3. Methodological Frameworks

A variety of domain-specific and general frameworks have instantiated token-faithful trajectory reconstruction:

A. Agentic RL and Polar

The Polar framework formalizes token-faithfulness for language-agent RL:

• Every LLM completion session is captured by a proxy that records raw prompt tokens $s_t = (z_1, \ldots, z_{t-1})$1, sampled response tokens $s_t = (z_1, \ldots, z_{t-1})$2, and returned log-probabilities $s_t = (z_1, \ldots, z_{t-1})$3.
• Trajectories are constructed at the token level, producing a flattened stream $s_t = (z_1, \ldots, z_{t-1})$4 with careful loss masking: only sampled tokens are marked for gradient flow.
• Two reconstruction strategies are provided: per_request (one trace per completion) and prefix_merging (chaining completions into longer trajectories when new prompts are strict prefixes).
• Loss masks, log-probabilities, and reward signals are aligned so that no spurious or injected tokens contribute to learning, yielding unbiased RL gradients and stable large-scale optimization (Xu et al., 22 May 2026).

B. Spatiotemporal Trajectory Recovery (PLMTrajRec)

PLMTrajRec reconstructs dense map-matched token sequences from sparse spatiotemporal data:

• Intervals and movement statistics are encoded into explicit textual prompts, with all information tokenized using the LLM's original tokenizer.
• Missing points are indicated by placeholder tokens; observed and imputed features are embedded via Fourier descriptors and road-segment proximity functions. Regional flows are injected as implicit prompts through CNN-extracted embeddings.
• A trajectory encoder unifies intervals, and LoRA-adapted Transformers process the tokenized trajectory, predicting recovered road segments and moving ratios at each discrete timepoint.
• Output tokens directly map to positions or segments in the original or imputed path, supporting auditing and downstream evaluation (Wei et al., 2024).

C. Token-Level Flight Path Reconstruction

Digit-level tokenization of ADS-B records in LLM-based flight trajectory reconstruction:

• Real numbers are encoded as contiguous digit tokens; each digit becomes a single token to maximize token-level faithfulness.
• At inference, the model outputs a summary in sequence form aligned with the original tokenization, supporting round-trip parsing from tokens to physical flight states.
• This approach maintains direct correspondance between tokens seen by the model and reconstructed geometric/physical states, ensuring “token-faithful” kinematic estimation (Zhang et al., 2024).

D. Multi-Token Attribution in LLMs (FlashTrace)

FlashTrace introduces efficient, recursive, and token-faithful attribution for long-horizon LLM reasoning:

• Span-wise aggregation computes importances on token spans $s_t = (z_1, \ldots, z_{t-1})$5 in a single linear pass, maintaining equivalence between attribution mass and actual output spans.
• A proximity metric converts multi-dimensional vector contributions into scalar attributions, which are recursively propagated through reasoning chains, preventing attribution mass loss in intermediate tokens.
• Input attributions are the sum over all hops, algorithmically guaranteeing faithfulness with sharp computational bounds: $s_t = (z_1, \ldots, z_{t-1})$6 time and strict mapping to true tokens (Pan et al., 2 Feb 2026).

4. Tokenization, Alignment, and Masking Practices

Token-faithful reconstruction demands that all data manipulations strictly respect the original tokenization boundaries:

• Digitization in numerical domains (e.g., flight states) ensures each digit is a single token, precluding floating-point drift or misalignment across learning and evaluation (Zhang et al., 2024).
• Explicit masking distinguishes “true actions” (sampled/model-generated tokens) from “injected” tokens (e.g., canonicalized UI turns, control metadata, prompts), such that only trainable tokens contribute to learning (Xu et al., 22 May 2026).
• For prompt-based models (e.g., PLMTrajRec), all interval and movement features are collapsed into explicit or implicit prompt tokens, with interval unification by inserting learned placeholders ([m]), and region-state descriptions encoded using CNN and positional functions (Wei et al., 2024).
• Trajectory builders must preserve order, segmentation, and all context tokens as seen or emitted by the model at run time.

5. Empirical Results and Performance Characteristics

Several studies quantify the value of token-faithful trajectory reconstruction:

• In RL, the Polar prefix_merging builder produces 5.4× fewer traces for a fixed number of training steps and increased GPU utilization (from 20.4% to 87.7%), significantly accelerating throughput and improving reward signal alignment. Direct performance gains include +22.6 points on Codex, +6.2 on Pi harness for code-writing tasks (Xu et al., 22 May 2026).
• PLMTrajRec achieves +16.5% accuracy and −352m RMSE improvements on sparse-to-dense network path recovery in the Chengdu dataset, and robust gains in Porto, showing strong generalization across intervals and city scales (Wei et al., 2024).
• For flight path reconstruction, LLaMA 2 7B achieves horizontal RMSE <5m and altitude MAE <10ft on dense ADS-B data, with lateral deviation for omitted data points <50m in 95% of segments; robust noise filtering and estimation fidelity are observed as a direct result of digit-level tokenization and prompt design (Zhang et al., 2024).
• In interpretability, FlashTrace yields >130× speedup over previous multi-token attribution methods (completion under 20s vs 38min for 5000-token spans), while recursive hops improve input-level faithfulness in ablation (MAS Deletion falls from 0.209 to 0.205 with a single recursive hop) (Pan et al., 2 Feb 2026).

6. Limitations, Challenges, and Future Directions

Persistent challenges include:

• Token-length ceilings: Transformer models often enforce strict context windows (e.g., 2048 tokens in LLaMA 2 7B), limiting the duration or length of reconstructable trajectories. Solutions include windowed/overlapping chunking, hierarchical summarization, or adoption of long-context architectures (Zhang et al., 2024, Wei et al., 2024).
• Mask drift and credit-assignment: Naive trajectory builders (per-request) fragment long sessions, increase variance in RL reward assignment, and introduce reward hacking; robust prefix-merging and explicit masking are needed to maintain coherent long-horizon traces (Xu et al., 22 May 2026).
• Generalization to denser or sparser sampling intervals relies on prompt flexibility and deep feature engineering; region-flow encoding and LoRA-based fine-tuning improve cross-task robustness in PLMTrajRec (Wei et al., 2024).
• For LLM attributions, mass loss in attribution chains or "faithfulness drop" necessitates recursive strategies (as in FlashTrace) to trace importance back to source inputs and avoid misleading explanations (Pan et al., 2 Feb 2026).

Ongoing directions involve comparative benchmarking across larger models, explicit prediction of regression or spline coefficients (LLMs outputting trajectory primitives), and new error metrics for evaluating the fidelity of reconstructed spatiotemporal or reasoning paths (Zhang et al., 2024, Wei et al., 2024).

7. Applications and Cross-Domain Impact

Token-faithful trajectory reconstruction connects diverse research domains:

Domain	Primary Purpose	Representative Method
RL (agentic LLMs)	Unbiased policy gradients, stable credit assignment	Polar (Xu et al., 22 May 2026)
Geospatial	Robust sparse-to-dense path recovery	PLMTrajRec (Wei et al., 2024)
Aviation	Precise flight-path estimation	Digit-level LLM (Zhang et al., 2024)
Interpretability	Multi-token causal attribution	FlashTrace (Pan et al., 2 Feb 2026)

In all settings, the preservation and faithful mapping of tokens is foundational for scientific accountability, unbiased learning, and reliable evaluation.