Learning to Play Like Humans: A Framework for LLM Adaptation in Interactive Fiction Games

Jinming Zhang; Yunfei Long

REVIEW 4 major objections 6 minor 31 references

Learning to Play Like Humans: A Framework for LLM Adaptation in Interactive Fiction Games

T0 review · 4 major / 6 minor · reviewed 2026-08-15 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗ Review JSON

Pith's one-line read Building a live map, learning valid commands, and reflecting on scored events lets an LLM play interactive fiction far better than plain prompting, the paper reports.

desk verdict The framework is plausible and the Zork1 ablation is careful, but the headline results rest on an unfair scoring window and a Dragon train/test leak, and the 'first LLM system' claim is contradicted by the authors' own citation. read the letter →

arxiv 2505.12439 v1 pith:V2I6D6B4 submitted 2025-05-18 cs.CL

Jinming Zhang , Yunfei Long This is my paper · ORCID

classification cs.CL

keywords interactivefictionlargelanguagemodelstext-basedgamesknowledgegraphmapaction-spacelearningexperiencereflectionzero-shotdecision-making

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

This paper asks whether LLMs can learn to play interactive fiction the way human players do, and argues that the answer is yes: it introduces LPLH, a framework that adds three memory structures—a dynamic knowledge-graph map, a growing action space of valid verb-object commands, and an experience library of summarized successes and failures—to a zero-shot LLM's next-command generation. On nine text-adventure games, LPLH raises raw and maximum scores across four LLM backbones, with the largest reported jump going from 10/100 to 68/100 for a 7B model on Detective. Ablations on Zork1 show each module helps and the full combination works best, which the authors take as evidence that human-like comprehension, rather than score-centric reinforcement learning, is the productive route for text-world agents. If the framework is right, interactive fiction becomes a learning problem for LLMs, not just a benchmark for RL.

What carries the argument

The load-bearing mechanism is the LPLH decision loop, which couples three memory structures: a dynamic knowledge-graph map updated after every observation and action; an action space grown from validated verb-object pairs; and an experience library written by an LLM summarizer whenever the score changes. At each step, the current map, viable action-object pairings for the current location, and retrieved experiences are composed into a prompt, and a zero-shot LLM emits the next command. This turns gameplay into structured learning with external memory rather than policy optimization.

What would settle it

Re-run the base LLM under the same scoring protocol used for LPLH—averaging only the final three of ten epochs—and compare; if the base model's last-three average approaches LPLH's raw score, the reported gains are an artifact of how scores were averaged. Separately, retrain the three fine-tuned helper modules on games that exclude Dragon, re-run Dragon, and check whether the score drops, which would indicate memorization rather than generalizable map, action, and experience learning.

Watch

Extended reading notes

Core claim

The paper's central claim is that an LLM playing an interactive fiction game becomes substantially better when three human-like behaviors are simulated: maintaining an up-to-date map of locations and objects, remembering which verb-object commands actually work, and reflecting on scoring events to extract reusable lessons. These structures are fed to a zero-shot LLM that produces the next command, with retrieved experiences supplied through retrieval-augmented generation. Empirically, LPLH improves raw and maximum scores over every base LLM tested across the nine games; for example, a 7B open-weight model on Detective rises from 10 to 68 raw points, and a 14B variant on Spellbrkr reaches 41.7/60, beating the DBERT-DRRN RL baseline. Ablations on Zork1 (LPLH14B 39.7 raw, base 9.0) indicate the gains come from combining the map, action space, and experience modules, with a chain-of-thought variant adding a further small improvement. The authors conclude that modeling human-like comprehension and decision-making enhances LLM performance and yields more interpretable gameplay.

Load-bearing premise

The central claim depends on a fair comparison: the baseline's raw score is averaged over all ten epochs, while LPLH's raw score is averaged over only the final three, and the fine-tuned helper modules were trained on games that include Dragon, which also appears among the evaluation games.

Editorial extensions

If this is right

Open-weight LLMs using LPLH can match or exceed the maximum scores of RL baselines on some games without reward shaping, as seen on Omniquest and Balances.
The module ablations imply that the value of the framework is additive: the full Zork1 configuration (39.7 raw) outperforms KG-map only (11.0), experience only (25.6), and action-space only (26.6).
Because the action LLM is prompted zero-shot, the framework is model-agnostic: changing only the backbone changes scores continuously, as shown across open-weight and proprietary models.
LPLH remains bounded by the same puzzle bottlenecks as RL agents, so its gains are specific to exploration, memory, and command selection rather than to solving obscure domain-specific puzzles.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

Editorial inference: the map-action-experience recipe should transfer to other partially observable language tasks, such as interactive instruction following or web navigation, where humans likewise build a mental model, learn a command vocabulary, and reuse lessons from success and failure.
Editorial inference: the contribution of the experience module likely depends on retrieval quality; changing the number of retrieved experiences or adding recency weighting could alter raw scores more than further prompt engineering of the summarizer.
Editorial inference: because the fine-tuned helper modules were trained on Dragon, Karn, and Night, and Dragon is itself one of the evaluation games, a held-out evaluation is needed to separate generalizable module learning from memorization of specific games.

Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Desk Editor's Note

read the letter

The headline results don't hold as reported. Table 1 averages LPLH raw scores over the final three epochs but baselines over all ten; the table note says so explicitly. That is not a subtle artifact, it is a structurally unfair comparison, and the 6.8x Detective gain is uninterpretable until both sides are scored on the same window. The second load-bearing issue is the Dragon leak: Appendix A says the fine-tuned fm was trained on 'Dragon, Karn, and Night' and calls them 'not in our test game,' yet Dragon appears in Table 1. That is a direct train/test contamination, even if fm is a small auxiliary model. Third, the novelty claim is simply wrong: the paper says it is the first system to leverage LLMs for IF games while citing Tsai et al. 2023, which played Zork1 with ChatGPT/GPT-4. The authors even describe that as the only prior LLM work a few lines later. These are not minor blemishes; they undercut the central empirical and novelty claims.

What the paper does well: the three-module design (dynamic KG map, action-space learning, experience memory) is coherent, and the Zork1 ablation in Table 2 is more careful. It compares component combinations against each other and against a plain baseline, and the pattern—each component helps, the full system helps most—is plausible. The error analysis of puzzle bottlenecks is honest and useful. The learning-curve interpretation is suggestive but overreaches: calling the curves 'human-like' without any human data is interpretive, not evidence.

Other soft spots: 'training-free' is misleading because fm is fine-tuned on Qwen2.5-1.5B. The paper also has typos that suggest haste, including 'LHLP' for 'LPLH' in Section 4.3.

The framework is worth a serious look, but the paper as submitted does not support its claims. The fix is clear: report matched raw scores (same epoch window for both), rerun Dragon without the contaminated fm or remove Dragon from evaluation, and correct the novelty claim. If those are addressed, the ablation and framework could make a useful contribution to LLM-agent research for text-based games.

The right move is to send it to peer review with a clear request for a matched evaluation and a reworked novelty statement. A serious referee can push on those points; desk rejection would lose a plausible engineering idea over fixable flaws.

Referee Report

4 major / 6 minor

Summary. The manuscript proposes LPLH, a modular framework for adapting LLM agents to interactive fiction games. It combines (1) a dynamically updated knowledge-graph map, (2) an action-space learner that stores validated verb–object pairs, and (3) an experience library that summarizes scoring events and retrieves them via RAG; the main agent acts in a zero-shot manner using these structures. The authors evaluate on nine Jericho games with four LLM backbones against RL and LLM baselines, report large raw-score improvements (e.g., Qwen-7B on Detective 68 vs 10), an ablation on Zork1, and case studies of learning from failure and success. They conclude that LPLH delivers human-like, interpretable learning and is the first system to leverage LLMs for IF game play.

Significance. The framework is clearly described, modular, and easy to reproduce from the appendix; the authors provide prompt templates, hyperparameters, and an ablation study with standard deviations, which is a strength. If the headline comparisons were valid, LPLH would be a useful step toward memory-augmented LLM agents in text-based games, and the cognitive-science framing, while qualitative, is a reasonable design inspiration. However, the current empirical support is compromised by an asymmetric scoring protocol and a train/evaluation overlap for the auxiliary model, so the claimed improvements are not yet interpretable as framework gains.

major comments (4)

[Section 4.3, Table 1] The raw score for the base LLM averages all ten epochs, while the raw score for LPLH averages only the final three epochs ('learning outcomes'). Since LPLH accumulates KG-map, action-space, and experience-library content across epochs, its late-epoch scores reflect a learned policy; the base agent has no cross-epoch memory, so its all-epoch average is not the appropriate baseline for post-learning performance. The Section 5.2 headline examples (e.g., Qwen-7B on Detective: 68/100 vs 10/100, called a 6.8x improvement) therefore confound the framework's effect with the scoring-window choice. Please provide matched comparisons: for both conditions, report the final-three-epoch average, the final-epoch score, or the all-ten-epoch average, together with per-epoch curves or deviations.
[Appendix A, Table 1] Appendix A states that the fine-tuned fm is trained on 'Dragon,' 'Karn,' and 'Night' and describes these as 'not in our test game,' yet Table 1 includes Dragon among the evaluation games. Because fm is used inside LPLH for action validation, relation extraction, and verb/object decomposition, the LPLH results on Dragon are in-domain for that auxiliary model and are not comparable to the base results without fm. Remove Dragon from the evaluation or retrain fm without Dragon and rerun the affected table entries.
[Section 5.1, Figure 3] The paper interprets score, vocabulary, and room-exploration curves as evidence that LPLH 'exhibits... learning dynamics that parallel human cognitive strategies' (Section 5.1). The only human reference is a single expert score line; no human vocabulary-growth, room-exploration, or score trajectories are collected. The qualitative alignment with the cited cognitive-science literature is plausible but untested. Either collect human play data or revise the abstract and conclusion claims from 'human-like' to 'cognitively inspired' or similar.
[Sections 1, 3.5, 4.3] The framework is called 'training-free' and said to 'eliminate the need for fine-tuning' (Sections 1 and 3.5), but Section 4.3 fine-tunes Qwen2.5-1.5B with LoRA for three auxiliary tasks. This is an internal inconsistency in a central framing claim; please clarify that the main action-generation LLM is zero-shot while the auxiliary fm is fine-tuned.

minor comments (6)

[Section 4.3] The acronym 'LHLP' should be 'LPLH'.
[Section 3.3] The heading 'LPHP' should be 'LPLH'.
[Section 3.5] The heading 'Zore-shot' should be 'Zero-shot'.
[Introduction and Conclusion] The claim that LPLH is the first LLM-based system for IF games is contradicted by the paper's own review of Tsai et al. (2023), who evaluated ChatGPT and GPT-4 on Zork1; please temper the novelty statement.
[Table 1] Reporting standard deviations or confidence intervals, as in Table 2, would help assess whether the observed differences are meaningful.
[Appendix A] The phrase 'In both games' follows a list of three games and should read 'In these games'.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the framework is empirical and its modules are not definitionally tied to the reported scores.

full rationale

This paper contains no derivation chain of the kind that could be circular. LPLH is an engineering system; its modules (KG-map, action space, experience library) are defined independently of the evaluation metric, and Equations (1)-(5) specify module computations rather than entailing the score improvements. No fitted parameter is renamed as a prediction, no uniqueness theorem is imported from the authors' prior work, and no load-bearing claim rests on a self-citation; the only self-citation (Zhang and Long 2025) is a general reference on LLM reasoning and plays no role in the experiments. The evaluation asymmetries noted by the reader—LPLH raw scores averaged over the last three epochs while the baseline averages all ten, and the fine-tuned fm being trained on Dragon, Karn, and Night while Dragon also appears in Table 1—are legitimate concerns about experimental fairness and potential data contamination, but they are not instances of a result being equivalent to its input by construction. No equation in the paper reduces to its own assumptions, and the central claim that the framework improves LLM performance is checked against game scores rather than asserted from the framework's definitions. Therefore the appropriate finding is no significant circularity.

Assumptions & free parameters 2 free parameters · 3 assumptions · 0 invented entities

The framework relies on two fitted components: the fine-tuned fm models and the post-hoc choice of which epochs count as learning outcomes. No new physical or conceptual entities are postulated. The core assumptions are about the validity of Jericho scores as metrics and the generalizability of the trained helper models.

free parameters (2)

fine-tuned fm weights = not released
The action validation, relation extraction, and verb/object decomposition models are fine-tuned on GPT-4o-generated annotations from Dragon, Karn, and Night. Their accuracy (90%, 15% error, 98%) directly affects the quality of the KG map and action space.
epoch window for learning outcomes = last 3 of 10 epochs
The reported LPLH raw score is the average of the final three epochs, while baseline raw score is the average of all ten. This ex-post selection favors LPLH and is an ad hoc decision.

assumptions (3)

domain assumption Jericho point scores are a reliable proxy for gameplay competence
The paper uses Jericho game scores as the primary evaluation metric without validating that scores correlate with human-like behavior or narrative understanding.
domain assumption The fine-tuned fm models generalize from three training games to unseen games
The fm models are trained on Dragon, Karn, and Night, then used across all nine evaluation games; the paper does not test generalization on held-out games beyond a Zork1 accuracy check.
domain assumption LLM agents do not exploit memorized game content in prompts
The paper claims no game names appear in prompts, but LLMs may still recognize classic games like Zork1 from observations, which could inflate performance independently of the framework.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Learning to Play Like Humans: A Framework for LLM Adaptation in Interactive Fiction Games." pith.science (2026). https://pith.science/paper/V2I6D6B4

@misc{pith2026250512439,
  author       = {Pith},
  title        = {Pith review of: Learning to Play Like Humans: A Framework for LLM Adaptation in Interactive Fiction Games},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/V2I6D6B4}},
  note         = {Machine review of arXiv:2505.12439}
}

read the original abstract

Interactive Fiction games (IF games) are where players interact through natural language commands. While recent advances in Artificial Intelligence agents have reignited interest in IF games as a domain for studying decision-making, existing approaches prioritize task-specific performance metrics over human-like comprehension of narrative context and gameplay logic. This work presents a cognitively inspired framework that guides Large Language Models (LLMs) to learn and play IF games systematically. Our proposed **L**earning to **P**lay **L**ike **H**umans (LPLH) framework integrates three key components: (1) structured map building to capture spatial and narrative relationships, (2) action learning to identify context-appropriate commands, and (3) feedback-driven experience analysis to refine decision-making over time. By aligning LLMs-based agents' behavior with narrative intent and commonsense constraints, LPLH moves beyond purely exploratory strategies to deliver more interpretable, human-like performance. Crucially, this approach draws on cognitive science principles to more closely simulate how human players read, interpret, and respond within narrative worlds. As a result, LPLH reframes the IF games challenge as a learning problem for LLMs-based agents, offering a new path toward robust, context-aware gameplay in complex text-based environments.

Figures

Figures reproduced from arXiv: 2505.12439 by the authors.

**Figure 1.** Example of RL approach, the basic LLM approach, and our LPLH approach for agent problem-solving, shedding light on core mechanisms of exploration and learning. Deep Recurrent Neural Network (DRNN) (He et al., 2016) sparked a growing interest in Reinforcement Learning (RL) settings where states and actions are expressed in natural language. Consequently, IF games have become a core testbed for integrating RL and na… view at source ↗

**Figure 2.** LPLH Framework. The Dynamic KG-map incrementally constructs a knowledge graph from observed items. The Action Space separates valid actions into verb-object pairs for efficient generation. The Experience Lib captures and summarizes key steps as reusable experiences to guide future decisions. play data to produce a compact set of action candidates. CALM significantly improves in-game scores, even on unseen games, wh… view at source ↗

**Figure 3.** Zork1 learning curve in scaled steps. For reference, human player’s best trajectory gets 350 scores in 412 steps with 48 verbs, 57 objects (total 105 unique words), and 63 rooms. 1 For all RL agents, they receive completed observation and inventory at each step. However, LLMs-based agent needs to decide when to call the command ’look’ or ’i’ to get such information by themselves. 6 [PITH_FULL_IMAGE:figures/full_fig… view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

31 extracted references · 29 canonical work pages

[1]

Opened",

If the observation doesn’t describe an environment or information is insufficient (e.g., "Opened", "Taken"), output |start| none |end| and skip other points

work page
[2]

Stairwell (First Floor) You’re in the north stairwell

Always use ’in’ as the relation to represent the character’s location. Convert any spatial descriptions (e.g., ’are facing’, ’are standing’, ’are behind’) to the ’in’ relation. If the input begins with a Room name (starts with a capital letter and does not end with a period), use it as the location. Example: Input: "Stairwell (First Floor) You’re in the n...

work page
[3]

If the observation doesn’t include a precise location, do not provide any <You, in, *> triple

work page
[4]

There is a small mailbox here

Use ’have’ as the relation to represent interactive objects present in the location. Focus only on the objects themselves as the ’obj’ in the triple. Ignore decorative details unless they indicate an interactive object. Limit object names to a maximum of three words. Example: Input: "There is a small mailbox here." Triple: <[Location], have, mailbox>

work page
[5]

You can’t

integrates dynamically constructed KG into the Advantage Actor-Critic framework to constrain the action space and improve decision-making. DBERT-DRRNL(Singh et al., 2022) enhances the traditional DRRN architecture by incorporating DistilBERT (Sanh et al., 2020), a pre-trained lan- guage model, to provide richer text representations, thereby improving the ...

work page 2022
[6]

Forest. You would need a machete to go further west

If the input specifies a requirement or action needed to continue, use <location/object, need/require, something to action>. Example: Input: "Forest. You would need a machete to go further west." Triple: <Forest, need, machete to go west>

work page
[7]

Hall. To the southwest is the entrance to the Computer Site, and you can go east here as well as go up with a stair

For objects or locations mentioned with a direction (e.g., ’to the north’, ’up to’, ’down’), use <current location, direction, [new location]/to [direction]>. Example: Input: "Hall. To the southwest is the entrance to the Computer Site, and you can go east here as well as go up with a stair." Triples: <Hall, southwest, Computer Site>, <Hall, east, to east...

work page
[10]

A buzzing water fountain has been moved

Do not include additional details or properties of objects. Only extract the objects themselves, ensuring object names are no longer than three words. But if a object have a relation to another object, such as ’in’ and ’on’, then extract that relation. Example: Input: "A buzzing water fountain has been moved." Triple: "<[Location], have, water fountain>" ...

work page

Show all 31 references

[13]

north" or

If the action is a simple directional command (e.g., "north" or "n"), the object list should be empty. For example: Input: "west" Response: "<act> <west; []> </act>"
[14]

take all

If the action is "take all" or another "all" command (e.g., "take all"), treat "take all" as the verb and leave the object list empty. For example: Input: "drop all" Response: "<act> <drop all; []> </act>"
[15]

eat", "take

If there are objects following the Verb (e.g., "eat", "take") or Verb phrase (e.g., "drop down", "go around"), list them. If prepositions (e.g., "on", "at", "with") are present, include them in the verb phrase using "&" as a placeholder, and list each noun object individually....
[16]

The action will help you understand what led to the new observation

Receive a history of the game’s the action you performed, the new observation representing what you see or experience after your action. The action will help you understand what led to the new observation
[17]

You must rely on these 10 turns and your internal reasoning to keep track of your location, inventory, puzzles, and progress

Have access only to the last 10 turns of conversation as your history. You must rely on these 10 turns and your internal reasoning to keep track of your location, inventory, puzzles, and progress
[18]

Receive current new observation based on the last action and the current game states as input
[19]

|start|" and

Produce all responses formatted between "|start|" and "|end|". **Your Task:** - At each turn, carefully read the provided new observation and the action you performed. - Use your internal chain-of-thought to determine the best possible action to advance in the game, considerin...
[20]

location

"location": where the player is (or what area is described) when the score changes. If the player has died, give the location name before death. *1.1* - One Location name Only. *1.2* - Description of situation
[21]

puzzle_status

"puzzle_status": what puzzles or obstacles have been solved to earn/lose the points. *2.1* - ONLY related steps to solve the puzzles directly. Any requirement for solving the puzzles, such as ’player need to <step>open door<step> at Room1 to enter <loc>Room2<loc>. *2.2* - Desc...
[22]

Any action leads to earning/losing points

"scoring": how the player earned/lost points for the last step. Any action leads to earning/losing points. *3.1* - Step done to earn/lose points. *3.2* - How many points are changed?
[23]

important_experience

"important_experience": The experience can be used for the future. Only the most notable and valuable clues or items the player learned about for the global game experience or any warning must be recorded through all previous logs. Only Focus on confirmed information. *Earn Po...
[24]

- **Recent Attempts**: Reflect on the previous actions, the motivation of taking that action and observation after this attempt

**Analyze the Current Game State** - **Room & Map Details**: Assess where you are, noting any exits, known layout, and significant objects. - **Recent Attempts**: Reflect on the previous actions, the motivation of taking that action and observation after this attempt. - **Inve...
[25]

- **Avoid Repetition**: Do not repeat failing commands indefinitely

**Use Retrieved Experiences and Past Attempts** - **Relevance**: Apply past successes or observed clues that align with the current room or situation. - **Avoid Repetition**: Do not repeat failing commands indefinitely. If a command fails, adjust strategy. - **Focus on Gains**...
[26]

- **Purpose**: Briefly ensure it’s the most logical next step, considering both context and success likelihood

**Formulate a Single Effective Command** - **One Action**: Provide exactly one executable command. - **Purpose**: Briefly ensure it’s the most logical next step, considering both context and success likelihood. - **Move command**: The full directions are [’north’, ’south’, ’ea...
[27]

|start| <com>[command]</com> <rea>[short motivation for the decision-making reason]</rea> |end| “‘ — **Adaptation and Fallback Rules**

**Output Format** - Present the final command and a short motivation in the following format without extra commentary: “‘ You internal reasoning steps Here. |start| <com>[command]</com> <rea>[short motivation for the decision-making reason]</rea> |end| “‘ — **Adaptation and Fa...
[28]

- **Next**: Options in ‘may_direction‘ or ‘may_have‘

**Priority Usage** - **Highest Priority**: Items in ‘temp_have‘. - **Next**: Options in ‘may_direction‘ or ‘may_have‘. - **Then**: Verified directions (‘direction‘) or items (‘have‘)
[29]

- Validate uncertain (‘may_‘) directions or items before fully committing to them

**Conflict Resolution** - Disregard prior attempts known to fail at this location or context. - Validate uncertain (‘may_‘) directions or items before fully committing to them. - After verify all the exits in one room then you can fully trust the map
[30]

- Look for overlooked clues or alternative ways forward

**Fallback Strategies** - If uncertain, explore unvisited areas or re-examine (’look’) the current room. - Look for overlooked clues or alternative ways forward
[31]

- In case an exploration fails, attempt a different angle—return to a previous room, look around again, or try another approach

**Exploratory Commands** - If tools are available, think of how to use them on obstacles. - In case an exploration fails, attempt a different angle—return to a previous room, look around again, or try another approach. - **Explore the world**: It’s better to try all directions...
[2016]

Jinming Zhang and Yunfei Long

Intuition and insight: Two processes that build on each other or fundamentally differ? Frontiers in psychology, 7:1395. Jinming Zhang and Yunfei Long. 2025. MLD-EA: Check and complete narrative coherence by intro- ducing emotions and actions. In Proceedings of the 31st Interna...

2025
[2020]

O’Reilly Media, Inc

Graph constrained reinforcement learning for natural language action spaces. arXiv preprint arXiv:2001.08837. Prithviraj Ammanabrolu, Ethan Tien, Matthew Hausknecht, and Mark O Riedl. 2020. How to avoid being eaten by a grue: Structured explo- ration strategies for textual wor...

2001 arXiv
[2023]

In Proceedings of the 17th Con- ference of the European Chapter of the Association for Computational Linguistics, pages 703–726

Self-imitation learning for action generation in text-based games. In Proceedings of the 17th Con- ference of the European Chapter of the Association for Computational Linguistics, pages 703–726. Ishika Singh, Gargi Singh, and Ashutosh Modi. 2021. Pre-trained language models a...

2021 arXiv
[2024]

In Proceedings of the 62nd Annual Meeting of the Association for Compu- tational Linguistics (Volume 3: System Demonstra- tions), Bangkok, Thailand

Llamafactory: Unified efficient fine-tuning of 100+ language models. In Proceedings of the 62nd Annual Meeting of the Association for Compu- tational Linguistics (Volume 3: System Demonstra- tions), Bangkok, Thailand. Association for Computa- tional Linguistics. Yuqi Zhu, Xiao...

2024

Pith tools

Reviewed August 15, 2026 · model on record in the stance chip above.

[1] [1]

Opened",

If the observation doesn’t describe an environment or information is insufficient (e.g., "Opened", "Taken"), output |start| none |end| and skip other points

work page

[2] [2]

Stairwell (First Floor) You’re in the north stairwell

Always use ’in’ as the relation to represent the character’s location. Convert any spatial descriptions (e.g., ’are facing’, ’are standing’, ’are behind’) to the ’in’ relation. If the input begins with a Room name (starts with a capital letter and does not end with a period), use it as the location. Example: Input: "Stairwell (First Floor) You’re in the n...

work page

[3] [3]

If the observation doesn’t include a precise location, do not provide any <You, in, *> triple

work page

[4] [4]

There is a small mailbox here

Use ’have’ as the relation to represent interactive objects present in the location. Focus only on the objects themselves as the ’obj’ in the triple. Ignore decorative details unless they indicate an interactive object. Limit object names to a maximum of three words. Example: Input: "There is a small mailbox here." Triple: <[Location], have, mailbox>

work page

[5] [5]

You can’t

integrates dynamically constructed KG into the Advantage Actor-Critic framework to constrain the action space and improve decision-making. DBERT-DRRNL(Singh et al., 2022) enhances the traditional DRRN architecture by incorporating DistilBERT (Sanh et al., 2020), a pre-trained lan- guage model, to provide richer text representations, thereby improving the ...

work page 2022

[6] [6]

Forest. You would need a machete to go further west

If the input specifies a requirement or action needed to continue, use <location/object, need/require, something to action>. Example: Input: "Forest. You would need a machete to go further west." Triple: <Forest, need, machete to go west>

work page

[7] [7]

Hall. To the southwest is the entrance to the Computer Site, and you can go east here as well as go up with a stair

For objects or locations mentioned with a direction (e.g., ’to the north’, ’up to’, ’down’), use <current location, direction, [new location]/to [direction]>. Example: Input: "Hall. To the southwest is the entrance to the Computer Site, and you can go east here as well as go up with a stair." Triples: <Hall, southwest, Computer Site>, <Hall, east, to east...

work page

[8] [10]

A buzzing water fountain has been moved

Do not include additional details or properties of objects. Only extract the objects themselves, ensuring object names are no longer than three words. But if a object have a relation to another object, such as ’in’ and ’on’, then extract that relation. Example: Input: "A buzzing water fountain has been moved." Triple: "<[Location], have, water fountain>" ...

work page