Chronology of Multi-Agent Interactions for Provenance of Evolving Information

The gist The study introduces a chronological system for post hoc attribution of generative history from content alone, without reliance on internal memory states or external meta-information.

Provenance Tracking Concept

Provenance is defined as the chronological history of things, resonating with the fundamental pursuit to uncover origins, trace connections, and situate entities within the flow of space and time<ref:2504.12612#pg2>. The system proposes a chronological system for tracking provenance in the context of multi-agent collaboration<ref:2504.12612#pg4>. At its core lies the notion of symbolic chronicles, representing signed and time-stamped records, in a form analogous to the chain of custody in forensic science<ref:2504.12612#pg4>. This system operates through a feedback loop, whereby each generative timestep updates the chronicle of prior interactions and synchronises it with the synthetic content in the very act of generation<ref:2504.12612#pg4>.

Chronicle Definition and Codebook Construction

A chronicle is defined as a symbolic chain of length T (i.e. the number of generative steps), represented by x = [x1, x2,..., xT], where each symbol xt ∈ [0, 1,..., n] indicates the identity of the agent assigned at timestep t<ref:2504.12612#pg5>. Each unique chronicle x is associated with a corresponding binary codeword c(x) ∈ [0, 1] V, where V denotes the vocabulary of the underlying foundation model<ref:2504.12612#pg5>. The collection of all codewords defines the codebook C = c(x) X, encompassing all possible chronicle configurations<ref:2504.12612#pg5>. To keep the construction within a tractable combinatorial regime, a scalable codebook generation strategy is formulated that first constructs a base codebook in a reduced space of dimension d, resulting in a matrix of size X ×d<ref:2504.12612#pg6>. Each base codeword is then expanded to match the full vocabulary size V through repetition and structured padding<ref:2504.12612#pg6>.

Chronological System with Feedback Loop

At each timestep t, an antecedent chronicle x (t−1) is decoded and then updated into a subsequent chronicle x (t) to be encoded, as illustrated in Figure 3 and presented in Algorithm 1<ref:2504.12612#pg6>. The chronicler updates the antecedent chronicle by inserting the informed agent identity at position t, leaving the remaining entries as zeros<ref:2504.12612#pg6>. The up-to-date chronicle is passed to the encoder, which embeds it into the text during the generation process of the designated agent<ref:2504.12612#pg6>. Once the text is generated, it is returned to the decoder, which retrieves the embedded chronicle<ref:2504.12612#pg6>. This feedback loop recursively updates the chronicle across timesteps, thereby enabling the tracking of agentic provenance throughout the generative process<ref:2504.12612#pg6>.

Encoding and Decoding Procedures

At each timestep t, the encoder receives a prompt and an up-to-date chronicle x (t), along with access to a predefined codebook, as illustrated in Figure 4<ref:2504.12612#pg7>. The codeword corresponding to the given chronicle c(x(t)) is retrieved from the codebook<ref:2504.12612#pg7>. Given the prompt, the designated agent predicts an unnormalised probability distribution (i.e. logits) over the entire vocabulary<ref:2504.12612#pg8>. Each logit is then selectively biased to favour tokens marked in the codeword<ref:2504.12612#pg8>. The biased logits are then normalised via the softmax function to obtain a probability distribution pv = softmax(˜lv) = e l˜v PV i=1 e l˜i<ref:2504.12612#pg8>. A token is sampled from the probability distribution, and the text generation process continues until an end criterion is met<ref:2504.12612#pg8>. The decoder estimates the embedded chronicle by comparing the lexical statistics computed over all possible codewords<ref:2504.12612#pg9>. The most probable chronicle xˆ is then estimated by selecting the one with maximal frequency<ref:2504.12612#pg10>.

Evaluation and Trade-offs

The accuracy of the recovered provenance, measured as ACC = Pt i=1 I(ˆxi = xi) t, captures the phenomenon of error propagation, wherein an incorrectly decoded symbol is carried forward into subsequent chronicle updates and encoding steps<ref:2504.12612#pg9>. It was observed that the accuracy tended to decrease as the timestep increased, reflecting the challenge of preserving provenance integrity over extended generation horizons<ref:2504.12612#pg9>. The experimental results reveal an inherent trade-off between the reliability of chronological attribution and the quality of text generation, governed by the bias strength parameter<ref:2504.12612#pg10>. Increasing the bias strength amplifies the distinguishability of statistical patterns, thereby improving the accuracy of chronological attribution<ref:2504.12612#pg10>. However, this comes at the cost of constraining the token sampling process, leading to higher perplexity and reduced linguistic fluency<ref:2504.12612#pg10>.

Robustness Against Adversarial Attacks

Under high bias strength (δ = 3), a significant improvement in adversarial robustness was observed, as accuracy remained above 90% even when up to 60% of words were substituted<ref:2504.12612#pg10>. These trends confirm that a stronger sampling bias can, to some extent, counteract aggressive lexical substitution by concentrating the sampling distribution on codeword-associated tokens<ref:2504.12612#pg10>. The degradation observed under low bias strength is due to the disruption of alignment between the generated tokens and the codeword-defined vocabulary subset<ref:2504.12612#pg10>.

Future Directions

The combinatorial growth governed by chronicle length and agent population reflects scaling challenges, suggesting the need for efficient systems that support longer or variable-length chronicles<ref:2504.12612#pg11>. Incorporating application-specific priors into the methodological design may enable adaptive reduction of the chronicle space, yielding a more scalable and context-aware framework<ref:2504.12612#pg11>. Future research may explore chronological provenance in more complex cyber ecosystems, including multimodal integration and human-inthe-loop generation<ref:2504.12612#pg11>. The paper concludes that the problem of tracking multi-agent provenance amidst the continual act of shared creation was investigated in this study<ref:2504.12612#pg10>. The history of agent contributions is not explicitly recorded as metadata but embedded within the generated content itself through the process of sampling lexical tokens<ref:2504.12612#pg10>.

Acknowledgement

This work was supported in part by the Japan Society for the Promotion of Science (JSPS) under KAKENHI Grants JP21H04907 and JP24H00732, and in part by the Japan Science and Technology Agency (JST) under the CREST Grants JPMJCR20D3 and JPMJCR2562. The paper's findings provide a chronological system for post hoc forensic attribution from generated text alone<ref:2504.12612#pg10>. It highlights the tension between forensic traceability and generation quality governed by the bias strength parameter<ref:2504.12612#pg10>. The system's robustness is enhanced by stronger sampling bias against lexical substitution attacks<ref:2504.12612#pg10>. The study confirms that the combinatorial growth of chronicle space poses a significant scaling challenge for practical deployment<ref:2504.12612#pg10>. The research points towards the need for efficient systems supporting longer or variable-length chronicles<ref:2504.12612#pg10>. The paper suggests that future work should explore chronological provenance in more complex cyber ecosystems<ref:2504.12612#pg10>. The system's reliance on lexical bias to embed history within text itself is a key methodological contribution<ref:2504.12612#pg10>. It demonstrates that provenance can be verified by testing for the presence or absence of a given watermark<ref:2504.

1. Bold Chronicle Tracking System Implementation: The system can track multi-agent provenance across the temporal dimension of generation by embedding a chronicle into text via a biased language generation process conditioned on the associated subset, allowing for post hoc attribution of generative history from content alone, without reliance on internal memory states or external meta-information.

2. Bold Codebook Scalability Strategy: To manage combinatorial explosion, the system can utilize a scalable codebook construction where a base codebook is generated in a reduced space and then expanded to match the full vocabulary size through repetition and structured padding, ensuring that sufficiently many distinct base codewords exist for random sampling.

3. Bold Robustness via Lexical Bias Tuning: The system's chronological accuracy can be maintained even under adversarial attacks by increasing the bias strength, as this allows the system to counteract lexical substitution by concentrating the sampling distribution on codeword-associated tokens, thereby forming more robust statistical features.

4. Bold Trade-off Management: The AI system can be configured to balance forensic traceability and generation quality by tuning the bias strength parameter, acknowledging that a stronger bias restricts the vocabulary available during the generation process which leads to a trade-off with perplexity.

Provenance is the chronological history of things, resonating with the fundamental pursuit to uncover origins, trace connections, and situate entities within the flow of space and time. As artificial intelligence advances towards autonomous agents capable of interactive collaboration on complex tasks, the provenance of generated content becomes entangled in the interplay of collective creation, where contributions are continuously revised, extended or overwritten. In a multi-agent generative chain, content undergoes successive transformations, often leaving little, if any, trace of prior contributions. In this study, we investigate the problem of tracking multi-agent provenance across the temporal dimension of generation. We propose a chronological system for post hoc attribution of generative history from content alone, without reliance on internal memory states or external meta-information. At its core lies the notion of symbolic chronicles, representing signed and time-stamped records, in a form analogous to the chain of custody in forensic science. The system operates through a feedback loop, whereby each generative timestep updates the chronicle of prior interactions and synchronises it with the synthetic content in the very act of generation. This research seeks to develop an accountable form of collaborative artificial intelligence within evolving cyber ecosystems.