Multi-Agent Transactive Memory

Ambuj Agrawal; Dishank Jain; Fernando Diaz; Negar Arabzadeh; To Eun Kim; Xuhong He

arxiv: 2606.19911 · v1 · pith:75EZKUQTnew · submitted 2026-06-18 · 💻 cs.AI · cs.CL· cs.IR

Multi-Agent Transactive Memory

To Eun Kim , Xuhong He , Dishank Jain , Ambuj Agrawal , Negar Arabzadeh , Fernando Diaz This is my paper

Pith reviewed 2026-06-26 17:37 UTC · model grok-4.3

classification 💻 cs.AI cs.CLcs.IR

keywords multi-agent systemstrajectory retrievalprocedural knowledgeknowledge sharinginteractive environmentsretrieval-augmented agentstransactive memory

0 comments

The pith

A shared repository lets agents retrieve and reuse trajectories produced by other agents to improve task performance and reduce steps.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper argues that agent trajectories encode reusable procedural knowledge that is typically wasted when each new agent starts fresh. By storing trajectories in a central repository that any agent can query, the system allows consumer agents to borrow solutions from producer agents. Experiments in ALFWorld and WebArena show higher success rates and fewer interaction steps when retrieval is used, even though agents come from separate instances and receive no joint training or coordination. A reader would care because the result suggests populations of agents can accumulate knowledge the way search engines accumulate human artifacts, rather than repeating the same discoveries.

Core claim

Multi-Agent Transactive Memory stores trajectories contributed by producer agents in a shared repository so that consumer agents can retrieve relevant ones, producing higher downstream task success and fewer interaction steps in interactive environments without coordination or joint training.

What carries the argument

The Multi-Agent Transactive Memory framework for population-level storage and retrieval of agent-generated trajectories.

If this is right

Producer agents contribute trajectories that later agents can apply to the same or similar tasks.
Task success rates rise in long-horizon interactive settings such as ALFWorld and WebArena.
Interaction length decreases without any requirement for agents to coordinate or train together.
The benefit holds across heterogeneous agent populations drawn from separate model families.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

Over time the repository could function as an accumulating library of solutions that grows with each new agent deployment.
Retrieval quality may depend on how well trajectory embeddings capture procedural similarity rather than surface features.
The same storage pattern could be tested on tasks outside simulated environments once suitable trajectory abstraction methods exist.

Load-bearing premise

Agent-generated trajectories contain reusable procedural knowledge that matches and transfers directly to consumer agents from different instances or models without extra adaptation or filtering.

What would settle it

A controlled test in which agents given retrieved trajectories show equal or lower success rates and equal or more steps than identical agents given no retrieval would falsify the performance claim.

Figures

Figures reproduced from arXiv: 2606.19911 by Ambuj Agrawal, Dishank Jain, Fernando Diaz, Negar Arabzadeh, To Eun Kim, Xuhong He.

**Figure 1.** Figure 1: Multi-Agent Transactive Memory (MATM). Traditional search serves humans retrieving humanauthored documents. RAG extends this to agents retrieving from human-generated corpora. MATM takes the next step by letting agents retrieve agent-generated artifacts such as interaction trajectories, which are atypical documents that differ fundamentally from humanwritten text. MATM can continually grow while servin… view at source ↗

**Figure 2.** Figure 2: Retrieval Advantage vs. Producer-Consumer [PITH_FULL_IMAGE:figures/full_fig_p007_2.png] view at source ↗

**Figure 3.** Figure 3: MATM memory scaling curves for ALFWorld (top) and WebArena (bottom). Success rate (left axis) and average steps per episode (right axis) as a function of index size. The dotted line marks SR of the noretrieval baseline. Results are averaged over five runs with different random seeds. to same-type trajectories is itself a source of degradation, excluding useful candidates that happen to cross task boundar… view at source ↗

**Figure 4.** Figure 4: The intent distribution across different web [PITH_FULL_IMAGE:figures/full_fig_p014_4.png] view at source ↗

**Figure 5.** Figure 5: Retrieval Advantage vs. Producer-Consumer [PITH_FULL_IMAGE:figures/full_fig_p016_5.png] view at source ↗

read the original abstract

The decentralized deployment of LLM agents with diverse capabilities across diverse tasks motivates infrastructure for knowledge sharing across heterogeneous agent populations. Just as search engines index human-generated artifacts to support human problem solving, retrieval systems can organize agent-generated artifacts for reuse across agent populations. We extend retrieval-augmented generation - which demonstrates the value of human-authored artifacts to individual agents - to retrieval of agent-generated artifacts supporting a population of agents. In particular, agent trajectories encode reusable procedural knowledge, yet these artifacts are typically discarded after a single use or retained only by the producing agent, forcing newly instantiated agents to repeatedly rediscover existing solutions. We propose Multi-Agent Transactive Memory (MATM), a framework for population-level storage and retrieval of agent-generated trajectories, where producer agents contribute trajectories to a shared repository and consumer agents retrieve them to improve task execution. We focus on interactive environments (ALFWorld and WebArena), where trajectories are long and encode especially rich procedural structure. Our experiments demonstrate that retrieving trajectories from MATM improves downstream task performance and reduces interaction steps without coordination or joint training. These results position MATM as a design pattern for population-level experience sharing in open agent ecosystems.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Desk Editor's Note private letter to a colleague

MATM is a clean extension of RAG to agent trajectories that addresses redundant rediscovery in open agent populations, though the size of the gains remains unclear without numbers.

read the letter

The core point is that this paper shows how to store agent trajectories in a shared repository so other agents can retrieve them and skip re-solving the same procedural steps. It works without any coordination or joint training, tested in ALFWorld and WebArena.

What is new is the population-level framing: instead of each agent discarding its trajectory after one run, the system treats those trajectories as reusable artifacts for heterogeneous agents. The choice of long-horizon interactive environments is sensible because those trajectories carry the richest structure. The design keeps things simple and decentralized, which fits the stated goal of open ecosystems.

The paper does a solid job naming the practical problem of repeated computation across agent instances and offering a retrieval-based fix that builds directly on existing RAG methods. The claim that this reduces interaction steps is consistent with how retrieval usually helps.

The soft spots are mostly around evidence. The abstract states performance gains but supplies no quantitative results, baselines, or retrieval details, so it is difficult to judge how large or reliable the improvement is. The key assumption—that trajectories encode transferable procedural knowledge across different model families—makes sense in principle but would benefit from more direct checks on when retrieval actually helps versus when it adds noise. Minor issues include the lack of error bars or ablation on the retrieval implementation itself.

This is for researchers working on agent infrastructure and multi-agent deployment who care about reducing redundant work. A reader looking for a straightforward design pattern rather than a theoretical advance would find it useful. It deserves a serious referee because the idea is testable on standard benchmarks and the central empirical claim can be checked with the right controls.

Referee Report

2 major / 2 minor

Summary. The manuscript proposes Multi-Agent Transactive Memory (MATM), a framework extending retrieval-augmented generation to agent-generated trajectories. Producer agents store trajectories encoding procedural knowledge in a shared repository; consumer agents retrieve them to improve execution in interactive environments (ALFWorld, WebArena). The central empirical claim is that retrieval improves downstream task performance and reduces interaction steps without coordination or joint training across heterogeneous agents.

Significance. If the empirical results hold with proper controls, MATM offers a practical, coordination-free design pattern for experience reuse across agent populations and model families. This extends RAG concepts to multi-agent settings and could support scalable open ecosystems where newly instantiated agents avoid rediscovering solutions.

major comments (2)

[Abstract, Experiments] Abstract and Experiments section: the manuscript states that retrieving trajectories improves performance and reduces steps on ALFWorld and WebArena, yet provides no quantitative metrics, error bars, baseline comparisons, retrieval implementation details, or statistical tests. Without these, the magnitude and reliability of the claimed gains cannot be assessed and post-hoc selection of trajectories remains possible.
[Methods] Methods/Implementation: the weakest assumption—that agent trajectories encode reusable procedural knowledge transferable across different agent instances or model families without adaptation or filtering—is not tested via ablation on cross-model transfer or negative examples; this is load-bearing for the population-level claim but lacks supporting controls.

minor comments (2)

[Framework] Notation for the shared repository and retrieval process is introduced informally; a formal definition or pseudocode would improve clarity.
[Abstract] The abstract claims 'without coordination or joint training' but does not explicitly contrast against coordinated or jointly trained baselines in the reported experiments.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive feedback, which identifies key areas where the empirical claims can be presented more rigorously. We address each major comment below.

read point-by-point responses

Referee: [Abstract, Experiments] Abstract and Experiments section: the manuscript states that retrieving trajectories improves performance and reduces steps on ALFWorld and WebArena, yet provides no quantitative metrics, error bars, baseline comparisons, retrieval implementation details, or statistical tests. Without these, the magnitude and reliability of the claimed gains cannot be assessed and post-hoc selection of trajectories remains possible.

Authors: We agree that the abstract summarizes results qualitatively and that the Experiments section requires more explicit reporting to substantiate the claims. In the revised manuscript we will update the abstract with specific quantitative metrics (success rates, step reductions), include error bars, baseline comparisons, retrieval implementation details, and statistical tests. This will also clarify how trajectories were selected to mitigate concerns about post-hoc selection. revision: yes
Referee: [Methods] Methods/Implementation: the weakest assumption—that agent trajectories encode reusable procedural knowledge transferable across different agent instances or model families without adaptation or filtering—is not tested via ablation on cross-model transfer or negative examples; this is load-bearing for the population-level claim but lacks supporting controls.

Authors: We acknowledge that the current experiments do not include explicit ablations for cross-model transfer or negative examples, which would strengthen support for the transferability assumption. We will add these controls (e.g., cross-model trajectory reuse and negative-example filtering) to the Methods and Experiments sections in the revision. revision: yes

Circularity Check

0 steps flagged

No significant circularity

full rationale

The paper proposes an empirical retrieval framework (MATM) for agent trajectories and validates it via experiments on ALFWorld and WebArena. No derivation chain, equations, or first-principles predictions exist that could reduce to inputs by construction. Claims rest on reported performance gains from retrieval, which are externally falsifiable via the described benchmarks. No self-citation load-bearing steps or ansatz smuggling are present in the provided text.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 1 invented entities

The central claim depends on the untested premise that trajectories contain transferable procedural structure and that a retrieval system can surface relevant ones; no free parameters, axioms, or invented entities are quantified in the abstract.

axioms (1)

domain assumption Agent trajectories encode reusable procedural knowledge that transfers across agents and task instances.
Invoked in the abstract when stating that trajectories are typically discarded yet contain reusable knowledge.

invented entities (1)

Multi-Agent Transactive Memory shared repository no independent evidence
purpose: Central storage for producer trajectories that consumer agents query.
New system component introduced to enable population-level sharing; no independent evidence provided in abstract.

pith-pipeline@v0.9.1-grok · 5739 in / 1311 out tokens · 22527 ms · 2026-06-26T17:37:50.771028+00:00 · methodology

discussion (0)

Reference graph

Works this paper leans on

40 extracted references · 4 canonical work pages · 1 internal anchor

[1]

InThe Eleventh International Confer- ence on Learning Representations

Quantifying memorization across neural lan- guage models. InThe Eleventh International Confer- ence on Learning Representations. Kai-Wei Chang, Akshay Krishnamurthy, Alekh Agar- wal, Hal Daumé III, and John Langford. 2015. Learn- ing to search better than your teacher. InInterna- tional Conference on Machine Learning, pages 2058–

2015
[2]

Zijian Chen, Xueguang Ma, Shengyao Zhuang, Jimmy Lin, Akari Asai, and Victor Zhong

PMLR. Zijian Chen, Xueguang Ma, Shengyao Zhuang, Jimmy Lin, Akari Asai, and Victor Zhong. 2026. Agentir: Reasoning-aware retrival for deep research agents. arXiv preprint arXiv:2603.04384. Marc-Alexandre Côté, Ákos Kádár, Xingdi Yuan, Ben Kybartas, Tavian Barnes, Emery Fine, James Moore, Ruo Yu Tao, Matthew Hausknecht, Layla El Asri, Mahmoud Adada, Wendy ...

work page arXiv 2026
[3]

Textworld: A learning environment for text- based games.CoRR, abs/1806.11532. 9 T. Cover and P. Hart. 1967. Nearest neighbor pattern classification.IEEE Transactions on Information Theory, 13(1):21–27. Yufan Dang, Chen Qian, Xueheng Luo, Jingru Fan, Zihao Xie, Ruijie Shi, Weize Chen, Cheng Yang, Xiaoyin Che, Ye Tian, Xuantang Xiong, Lei Han, Zhiyuan Liu, ...

work page arXiv 1967
[4]

Artificial Intelligence, 33(1):1–64

Soar: An architecture for general intelligence. Artificial Intelligence, 33(1):1–64. Patrick Lewis, Ethan Perez, Aleksandra Piktus, Fabio Petroni, Vladimir Karpukhin, Naman Goyal, Hein- rich Küttler, Mike Lewis, Wen-tau Yih, Tim Rock- täschel, and 1 others. 2020. Retrieval-augmented gen- eration for knowledge-intensive nlp tasks.Advances in neural informa...

work page arXiv 2020
[5]

Multi-Agent Collaboration Mechanisms: A Survey of LLMs

CLIN: A continually learning language agent for rapid task adaptation and generalization. InFirst Conference on Language Modeling. Model Context Protocol. 2026. Model context pro- tocol. https://modelcontextprotocol.io. Ac- cessed: 2026-05-26. Niklas Muennighoff, Zitong Yang, Weijia Shi, Xi- ang Lisa Li, Li Fei-Fei, Hannaneh Hajishirzi, Luke Zettlemoyer, ...

work page internal anchor Pith review Pith/arXiv arXiv 2026
[6]

goal": "<goal_text>

Supercorrect: Advancing small LLM rea- soning with thought template distillation and self- correction. InThe Thirteenth International Confer- ence on Learning Representations. Shunyu Yao, Dian Yu, Jeffrey Zhao, Izhak Shafran, Thomas L. Griffiths, Yuan Cao, and Karthik R Narasimhan. 2023. Tree of thoughts: Deliberate problem solving with large language mod...

2023
[7]

Your action MUST be character-for-character identical to one item in admissible_actions
[8]

Do NOT modify, abbreviate, or paraphrase actions
[9]

Do NOT use actions from retrieved trajectories unless they appear in current admissible_actions
[10]

look" or navigation ONE-SHOT EXAMPLE: Task: clean some apple and put it in sidetable. Turn 1: Observation: The fridge 1 is closed. Response: {

If confused, pick a safe exploratory action like "look" or navigation ONE-SHOT EXAMPLE: Task: clean some apple and put it in sidetable. Turn 1: Observation: The fridge 1 is closed. Response: {"reasoning": "Let’s think step by step. I should open fridge 1 to see what’s inside.", "action": "open fridge 1"} 18 Turn 2: Observation: You open the fridge 1. The ...
[11]

reasoning

<action_2> ... RESPONSE FORMAT: You MUST respond with valid JSON in this exact format: {"reasoning": "Let’s think step by step. [your detailed reasoning]", "action": "exact action from admissible_actions"} Where: - reasoning: MUST start with ‘Let’s think step by step.’ Then explain your thought process, what you observe, and why this action is best - acti...
[13]

Do not include any text before or after the JSON object. M.3 ALFWorld Trajectory-Augmented Prompt SYSTEM Same as ALFWorld Baseline (no-retrieval) System Prompt USER GOAL: <goal_text> CURRENT STEP: <current_step> / <max_steps> — RECENT HISTORY (Previous Steps - For Reference Only) — <recent_history_str> — End of Recent History — — RETRIEVED TRAJECTORY GUID...
[14]

The sequence of actions that led to success
[15]

The observations and their progression
[16]

How to adapt this strategy to the current situation and your goal
[17]

What steps might be different or similar in your current context — End of Trajectory Guidance — >>> CURRENT OBSERVATION (Focus on This - Current State): <observation> <<< End of Current Observation ADMISSIBLE ACTIONS (<N> total):
[18]

reasoning

<action_2> ... RESPONSE FORMAT: You MUST respond with valid JSON in this exact format: {"reasoning": "Let’s think step by step. [your detailed reasoning]", "action": "exact action from admissible_actions"} Where: - reasoning: MUST start with ‘Let’s think step by step.’ Then explain your thought process, what you observe, and why this action is best - acti...
[20]

M.4 WebArena Baseline (no-retrieval) Prompt SYSTEM You are an autonomous intelligent agent tasked with navigating a web browser

Do not include any text before or after the JSON object. M.4 WebArena Baseline (no-retrieval) Prompt SYSTEM You are an autonomous intelligent agent tasked with navigating a web browser. You will be given web-based tasks. These tasks will be accomplished through the use of specific actions you can issue. Here’s the information you’ll have: The user’s objec...
[21]

You should only issue an action that is valid given the current observation
[22]

You should only issue one action at a time
[23]

For each observation, think step by step about what you see and what action to take next
[24]

Example 1: OBSERVATION:

Your action MUST be one of the valid action formats listed above Few-shot examples and closing sentence. Example 1: OBSERVATION:
[25]

link ’HP CB782A#ABA 640 Inkjet Fax Machine (Renewed)’
[26]

StaticText ’$279.49’
[27]

button ’Add to Cart’
[28]

button ’Add to Wish List’
[29]

reasoning

button ’Add to Compare’ URL: http://onestopmarket.com/office-products/office-electronics.html OBJECTIVE: What is the price of HP CB782A#ABA 640 Inkjet Fax Machine (Renewed)? PREVIOUS ACTION: None Response: {"reasoning": "Let’s think step by step. I can see the product HP CB782A#ABA 640 Inkjet Fax Machine (Renewed) is displayed on the page. The price is cl...
[30]

textbox ’Search’ focused: True required: False
[31]

reasoning

button ’Search’ URL: http://onestopmarket.com OBJECTIVE: Search for USB-C to USB-B cable on the site. PREVIOUS ACTION: None Response: {"reasoning": "Let’s think step by step. I need to search for USB-C to USB-B cable on this site. I can see a search textbox with id [164] that is already focused. I should type my search query into this textbox and press en...
[32]

heading ’Certified Refurbished Kindle Paperwhite’
[33]

StaticText ’by Amazon’
[34]

StaticText ’Price: $79.99’
[35]

reasoning

button ’Add to Cart’ URL: http://onestopmarket.com/dp/B00ZV 9RDKK OBJECTIVE: Show me the details of the Kindle Paperwhite. PREVIOUS ACTION: click [582] Response: {"reasoning": "Let’s think step by step. I can see the Kindle Paperwhite product page with the heading, price ($79.99), and an Add to Cart button. The objective is to show details of the product....
[37]

Do not include any text before or after the JSON object. M.5 WebArena Trajectory-Augmented Prompt SYSTEM Same as WebArena Baseline (no-retrieval) System Prompt USER GOAL: <goal_text> CURRENT STEP: <current_step> / <max_steps> — RECENT HISTORY (Previous Steps - For Reference Only) — <recent_history_str> — End of Recent History — — RETRIEVED TRAJECTORY GUID...
[38]

The sequence of actions taken 23
[39]

How the agent navigated through the website
[40]

What elements were clicked and in what order
[41]

reasoning

When the task was completed — End of Trajectory Guidance — >>> CURRENT OBSERVATION (Focus on This - Current State): <observation> <<< End of Current Observation CURRENT URL: <url> RESPONSE FORMAT: You MUST respond with valid JSON in this exact format: {"reasoning": "Let’s think step by step. [your detailed reasoning]", "action": "action to be taken"} Wher...
[42]

Your reasoning MUST begin with ‘Let’s think step by step.’
[43]

N Dataset License • ALFWorld: MIT License • WebArena: Apache License 2.0 O Computational Budget For retrieval from the MATM index, we use one NVIDIA L40S GPU

Do not include any text before or after the JSON object. N Dataset License • ALFWorld: MIT License • WebArena: Apache License 2.0 O Computational Budget For retrieval from the MATM index, we use one NVIDIA L40S GPU. For LLM inference in experiments, the OpenRouter API was used. The total cost was approximately 2,000 USD. P Use of AI Assistants AI assistan...

[1] [1]

InThe Eleventh International Confer- ence on Learning Representations

Quantifying memorization across neural lan- guage models. InThe Eleventh International Confer- ence on Learning Representations. Kai-Wei Chang, Akshay Krishnamurthy, Alekh Agar- wal, Hal Daumé III, and John Langford. 2015. Learn- ing to search better than your teacher. InInterna- tional Conference on Machine Learning, pages 2058–

2015

[2] [2]

Zijian Chen, Xueguang Ma, Shengyao Zhuang, Jimmy Lin, Akari Asai, and Victor Zhong

PMLR. Zijian Chen, Xueguang Ma, Shengyao Zhuang, Jimmy Lin, Akari Asai, and Victor Zhong. 2026. Agentir: Reasoning-aware retrival for deep research agents. arXiv preprint arXiv:2603.04384. Marc-Alexandre Côté, Ákos Kádár, Xingdi Yuan, Ben Kybartas, Tavian Barnes, Emery Fine, James Moore, Ruo Yu Tao, Matthew Hausknecht, Layla El Asri, Mahmoud Adada, Wendy ...

work page arXiv 2026

[3] [3]

Textworld: A learning environment for text- based games.CoRR, abs/1806.11532. 9 T. Cover and P. Hart. 1967. Nearest neighbor pattern classification.IEEE Transactions on Information Theory, 13(1):21–27. Yufan Dang, Chen Qian, Xueheng Luo, Jingru Fan, Zihao Xie, Ruijie Shi, Weize Chen, Cheng Yang, Xiaoyin Che, Ye Tian, Xuantang Xiong, Lei Han, Zhiyuan Liu, ...

work page arXiv 1967

[4] [4]

Artificial Intelligence, 33(1):1–64

Soar: An architecture for general intelligence. Artificial Intelligence, 33(1):1–64. Patrick Lewis, Ethan Perez, Aleksandra Piktus, Fabio Petroni, Vladimir Karpukhin, Naman Goyal, Hein- rich Küttler, Mike Lewis, Wen-tau Yih, Tim Rock- täschel, and 1 others. 2020. Retrieval-augmented gen- eration for knowledge-intensive nlp tasks.Advances in neural informa...

work page arXiv 2020

[5] [5]

Multi-Agent Collaboration Mechanisms: A Survey of LLMs

CLIN: A continually learning language agent for rapid task adaptation and generalization. InFirst Conference on Language Modeling. Model Context Protocol. 2026. Model context pro- tocol. https://modelcontextprotocol.io. Ac- cessed: 2026-05-26. Niklas Muennighoff, Zitong Yang, Weijia Shi, Xi- ang Lisa Li, Li Fei-Fei, Hannaneh Hajishirzi, Luke Zettlemoyer, ...

work page internal anchor Pith review Pith/arXiv arXiv 2026

[6] [6]

goal": "<goal_text>

Supercorrect: Advancing small LLM rea- soning with thought template distillation and self- correction. InThe Thirteenth International Confer- ence on Learning Representations. Shunyu Yao, Dian Yu, Jeffrey Zhao, Izhak Shafran, Thomas L. Griffiths, Yuan Cao, and Karthik R Narasimhan. 2023. Tree of thoughts: Deliberate problem solving with large language mod...

2023

[7] [7]

Your action MUST be character-for-character identical to one item in admissible_actions

[8] [8]

Do NOT modify, abbreviate, or paraphrase actions

[9] [9]

Do NOT use actions from retrieved trajectories unless they appear in current admissible_actions

[10] [10]

look" or navigation ONE-SHOT EXAMPLE: Task: clean some apple and put it in sidetable. Turn 1: Observation: The fridge 1 is closed. Response: {

If confused, pick a safe exploratory action like "look" or navigation ONE-SHOT EXAMPLE: Task: clean some apple and put it in sidetable. Turn 1: Observation: The fridge 1 is closed. Response: {"reasoning": "Let’s think step by step. I should open fridge 1 to see what’s inside.", "action": "open fridge 1"} 18 Turn 2: Observation: You open the fridge 1. The ...

[11] [11]

reasoning

<action_2> ... RESPONSE FORMAT: You MUST respond with valid JSON in this exact format: {"reasoning": "Let’s think step by step. [your detailed reasoning]", "action": "exact action from admissible_actions"} Where: - reasoning: MUST start with ‘Let’s think step by step.’ Then explain your thought process, what you observe, and why this action is best - acti...

[12] [13]

Do not include any text before or after the JSON object. M.3 ALFWorld Trajectory-Augmented Prompt SYSTEM Same as ALFWorld Baseline (no-retrieval) System Prompt USER GOAL: <goal_text> CURRENT STEP: <current_step> / <max_steps> — RECENT HISTORY (Previous Steps - For Reference Only) — <recent_history_str> — End of Recent History — — RETRIEVED TRAJECTORY GUID...

[13] [14]

The sequence of actions that led to success

[14] [15]

The observations and their progression

[15] [16]

How to adapt this strategy to the current situation and your goal

[16] [17]

What steps might be different or similar in your current context — End of Trajectory Guidance — >>> CURRENT OBSERVATION (Focus on This - Current State): <observation> <<< End of Current Observation ADMISSIBLE ACTIONS (<N> total):

[17] [18]

reasoning

<action_2> ... RESPONSE FORMAT: You MUST respond with valid JSON in this exact format: {"reasoning": "Let’s think step by step. [your detailed reasoning]", "action": "exact action from admissible_actions"} Where: - reasoning: MUST start with ‘Let’s think step by step.’ Then explain your thought process, what you observe, and why this action is best - acti...

[18] [20]

M.4 WebArena Baseline (no-retrieval) Prompt SYSTEM You are an autonomous intelligent agent tasked with navigating a web browser

Do not include any text before or after the JSON object. M.4 WebArena Baseline (no-retrieval) Prompt SYSTEM You are an autonomous intelligent agent tasked with navigating a web browser. You will be given web-based tasks. These tasks will be accomplished through the use of specific actions you can issue. Here’s the information you’ll have: The user’s objec...

[19] [21]

You should only issue an action that is valid given the current observation

[20] [22]

You should only issue one action at a time

[21] [23]

For each observation, think step by step about what you see and what action to take next

[22] [24]

Example 1: OBSERVATION:

Your action MUST be one of the valid action formats listed above Few-shot examples and closing sentence. Example 1: OBSERVATION:

[23] [25]

link ’HP CB782A#ABA 640 Inkjet Fax Machine (Renewed)’

[24] [26]

StaticText ’$279.49’

[25] [27]

button ’Add to Cart’

[26] [28]

button ’Add to Wish List’

[27] [29]

reasoning

button ’Add to Compare’ URL: http://onestopmarket.com/office-products/office-electronics.html OBJECTIVE: What is the price of HP CB782A#ABA 640 Inkjet Fax Machine (Renewed)? PREVIOUS ACTION: None Response: {"reasoning": "Let’s think step by step. I can see the product HP CB782A#ABA 640 Inkjet Fax Machine (Renewed) is displayed on the page. The price is cl...

[28] [30]

textbox ’Search’ focused: True required: False

[29] [31]

reasoning

button ’Search’ URL: http://onestopmarket.com OBJECTIVE: Search for USB-C to USB-B cable on the site. PREVIOUS ACTION: None Response: {"reasoning": "Let’s think step by step. I need to search for USB-C to USB-B cable on this site. I can see a search textbox with id [164] that is already focused. I should type my search query into this textbox and press en...

[30] [32]

heading ’Certified Refurbished Kindle Paperwhite’

[31] [33]

StaticText ’by Amazon’

[32] [34]

StaticText ’Price: $79.99’

[33] [35]

reasoning

button ’Add to Cart’ URL: http://onestopmarket.com/dp/B00ZV 9RDKK OBJECTIVE: Show me the details of the Kindle Paperwhite. PREVIOUS ACTION: click [582] Response: {"reasoning": "Let’s think step by step. I can see the Kindle Paperwhite product page with the heading, price ($79.99), and an Add to Cart button. The objective is to show details of the product....

[34] [37]

Do not include any text before or after the JSON object. M.5 WebArena Trajectory-Augmented Prompt SYSTEM Same as WebArena Baseline (no-retrieval) System Prompt USER GOAL: <goal_text> CURRENT STEP: <current_step> / <max_steps> — RECENT HISTORY (Previous Steps - For Reference Only) — <recent_history_str> — End of Recent History — — RETRIEVED TRAJECTORY GUID...

[35] [38]

The sequence of actions taken 23

[36] [39]

How the agent navigated through the website

[37] [40]

What elements were clicked and in what order

[38] [41]

reasoning

When the task was completed — End of Trajectory Guidance — >>> CURRENT OBSERVATION (Focus on This - Current State): <observation> <<< End of Current Observation CURRENT URL: <url> RESPONSE FORMAT: You MUST respond with valid JSON in this exact format: {"reasoning": "Let’s think step by step. [your detailed reasoning]", "action": "action to be taken"} Wher...

[39] [42]

Your reasoning MUST begin with ‘Let’s think step by step.’

[40] [43]

N Dataset License • ALFWorld: MIT License • WebArena: Apache License 2.0 O Computational Budget For retrieval from the MATM index, we use one NVIDIA L40S GPU

Do not include any text before or after the JSON object. N Dataset License • ALFWorld: MIT License • WebArena: Apache License 2.0 O Computational Budget For retrieval from the MATM index, we use one NVIDIA L40S GPU. For LLM inference in experiments, the OpenRouter API was used. The total cost was approximately 2,000 USD. P Use of AI Assistants AI assistan...