Technical Summary: "Don't Say It!": Constraints, Compliance, and Communication when Language Models Play Taboo

1. Problem Statement

The game of Taboo requires a player to describe a target word without using a specific set of forbidden words, relying on semantic associations to allow a guesser to identify the target. This task presents a unique challenge for Large Language Models (LLMs) because it demands the simultaneous satisfaction of two competing objectives:

1. Strict Lexical Constraints: The model must suppress lexically and conceptually salient words (the forbidden list) and their morphological variants.
2. Communicative Effectiveness: The model must generate a description sufficiently informative to evoke the target concept for a human or machine guesser.

Current benchmarks often test instruction following in isolation. This paper investigates whether LLMs can navigate the trade-off between rule compliance and communicative utility in a setting that requires "lexical grounding under constraint." The authors specifically question whether descriptions generated under constraint are detailed enough to be useful and whether the strategies models adopt align with human cognitive processes in the game.

2. Methodology

2.1. Data and Models

• Dataset: The study utilizes 194 Italian Taboo cards manually transcribed from a physical board game. Each card contains a target word and five forbidden words.
• Models: Two open-weight models were evaluated:
  • google/gemma-3-4b-it: A lightweight instruction-tuned model selected for its general performance and compatibility with pre-trained Sparse AutoEncoders (SAEs) for interpretability experiments.
  • openai/gpt-oss-20b: A Mixture-of-Experts reasoning model post-trained with Chain-of-Thought (CoT) reinforcement learning. It was evaluated both with and without explicit reasoning to assess the role of deliberation in constraint adherence.

2.2. Experimental Conditions (Intervention Levels)

The authors investigated three distinct levels of intervention to enforce constraints:

1. Prompting (Instruction Level): The model is instructed via the user prompt to avoid the target word and the forbidden list (including morphological derivatives). The prompt explicitly requests a description under 30 words with no surrounding text.
2. Generation-Time Constraints (Decoding Level):
   • A Priori Stems Censorship: Before generation, a static set of banned tokens is computed based on the stems of forbidden words. Logits for these tokens are masked to $-\infty$ at every decoding step.
   • Inference-Time Constrained Generation: A dynamic approach where the model generates freely until a complete word matching a forbidden stem is detected. Generation then backtracks to the start of the offending word, which is added to a persistent position-level mask, and generation resumes with the token suppressed.
3. Model Manipulations (Representation Level):
   • SAE Steering: Using Sparse AutoEncoders (SAEs), the authors identified features in the model's residual stream (Layer 29) most strongly associated with each forbidden word. During generation, an activation-norm-scaled intervention was applied to steer the residual stream away from these feature directions, suppressing the underlying concepts without modifying model weights or explicitly mentioning forbidden words in the prompt.

2.3. Evaluation Metrics

• Compliance:
  • Accuracy: Percentage of outputs with zero violations (verbatim or morphological).
  • Violation Types: Share of exact violations, morphological violations, and "target leaks" (accidentally using the target word).
  • No : Rate of reasoning budget exhaustion for the GPT model.
• Communicative Effectiveness:
  • LLM-as-a-Judge: Two models (Gemma and GPT) acted as judges, reading descriptions and guessing the target. Effectiveness was measured by Pass@k (whether the target's first token is in the top-k predictions) and Raw token likelihood.
  • Human Evaluation: 8 annotators participated in two phases:
    • Guessing: Annotators guessed targets from model-generated descriptions.
    • Describing: Annotators generated descriptions under "spontaneous" (oral-style) and "deliberate" (written) conditions to serve as a baseline for model performance.

3. Key Results

3.1. Constraint Compliance

• Prompting: Significantly improved compliance over the unconstrained baseline. Gemma-3-4b-it reached 91.2% accuracy; GPT-oss-20b with reasoning reached 97.4%. Reasoning helped GPT suppress verbatim words but was less effective against morphological variants.
• Constrained Generation: Achieved near-perfect compliance. Gemma-3-4b-it reached 98.5% (a priori) and 96.4% (online). GPT-oss-20b with reasoning reached 92.3% (a priori) but dropped to 87.1% (online), with a higher rate of target word leaks.
• SAE Steering: Performed poorly on compliance (55.7%), significantly lower than prompting and constrained generation. This suggests that representation-level steering alone cannot reliably enforce surface-level lexical constraints.

3.2. Communicative Effectiveness (Informativeness)

• Trade-offs: The method of constraint imposition shaped the quality of the description.
  • Prompting tended to produce overly cautious, informationally sparse descriptions.
  • Constrained Generation forced the model toward more semantically precise paraphrases, resulting in higher Pass@10 scores for judges compared to prompting.
  • SAE Steering showed a striking dissociation: despite low compliance, it produced descriptions that were competitively informative (Pass@10 up to 21.1%), suggesting that blocking surface tokens imposes a descriptive cost that representation-level intervention avoids.
• Judge Sensitivity: The two judge models differed significantly in sensitivity. GPT-oss-20b assigned near-zero probability to targets across most conditions, while Gemma-3-4b-it was more generous, though both agreed on the relative ordering of methods.

3.3. Model vs. Human Performance

• Asymmetry in Guessing: Models performed substantially worse than humans as guessers.
  • Human annotators achieved 30.77% accuracy guessing from model descriptions.
  • Model judges (even Gemma-3-4b-it) required 5 to 10 guesses (Pass@5 to Pass@10) to approach the accuracy humans achieved with just 1 guess (Pass@1) on human-written descriptions.
  • GPT-oss-20b as a judge assigned near-zero probability to correct targets in almost all cases.
• Asymmetry in Describing:
  • For models, reasoning (CoT) significantly improved compliance (72.2% vs. 97.4% for GPT).
  • For humans, spontaneous oral descriptions yielded higher guessing accuracy (95%) than deliberate written descriptions (85%), likely due to the interactive, real-time refinement possible in oral communication, which static text generation lacks.

4. Contributions and Significance

Key Contributions

1. Multi-Level Constraint Framework: The paper systematically compares constraint enforcement strategies across three levels of the generative process: prompting, decoding-time masking, and internal representation manipulation (SAEs).
2. Bidirectional Human-Model Evaluation: Unlike previous studies that rely solely on automatic metrics, this work grounds evaluation in actual gameplay by reversing roles: humans guess model descriptions, and models guess human descriptions.
3. Characterization of Trade-offs: The study demonstrates that rule compliance and communicative effectiveness are not independently controllable. The mechanism used to impose a constraint (instruction vs. decoding vs. representation) fundamentally alters the nature of the output, creating distinct trade-offs between safety (avoiding forbidden words) and utility (evoking the target).

Significance and Claims

The authors claim that lexical grounding under constraint remains an open challenge for current language models.

• The "Guessing Gap": The inability of models to guess targets from indirect descriptions with human-level efficiency suggests that LLMs lack the same network of salient lexical associations that make Taboo intuitive for humans. The authors suggest this points to a difference between the fast, associative "System 1" processing used by humans and the deliberate "System 2" reasoning of models.
• Limitations: The authors modestly note that the benchmark is currently Italian-only, the shared evaluation set is small (10 words), and SAE steering was only tested on one architecture. They also highlight that the study evaluates models in a single-turn setting, whereas the original game is inherently interactive.
• Future Direction: The paper argues that games like Taboo constitute a natural, underexplored testbed for language model capabilities, specifically for probing pragmatic adaptation and collaborative grounding in multi-turn, multi-agent configurations.