FrameRef: A Framing Dataset and Simulation Testbed for Modeling Bounded Rational Information Health

1. Problem Statement

The rapid expansion of social media, search engines, and generative AI has intensified the production and distribution of misinformation, disinformation, and toxic content. These systems shape Information Health—defined as an individual's long-term ability to understand, evaluate, and respond to information.

Current research faces two main challenges:

1. Ethical and Practical Constraints: Studying the long-term effects of biased information exposure on real humans is resource-intensive and ethically risky.
2. Modeling Limitations: Existing computational models often fail to capture bounded rationality. They either lack systematic biases (being too "rational") or degrade in overall task competence when trained to exhibit bias. There is a need for a framework that simulates how small, systematic framing biases in early judgments compound over time to create divergent information health trajectories.

2. Methodology

The authors propose a two-pronged approach: the creation of a large-scale dataset (FrameRef) and a simulation-based framework using fine-tuned language model agents.

A. The FrameRef Dataset

• Scale & Source: Built upon the FEVER and FEVEROUS datasets, FrameRef contains 1,073,740 systematically reframed claims.
• Framing Dimensions: The dataset introduces five specific framing styles that alter presentation without changing the underlying propositional truth:
  1. Authoritative: Citing recognized authorities.
  2. Consensus: Citing widespread agreement.
  3. Emotional: Using affective language.
  4. Prestige: Highlighting the status of the subject.
  5. Sensationalist: Using dramatic or hyperbolic language.
• Construction Pipeline:
  1. Reframing: Llama-3.1-8B-Instruct was used to rewrite claims into the five framing styles, ensuring no new facts were introduced.
  2. Verification: DeepSeek-R1-Distill-Llama-8B was used to verify that the reframed claims retained the original truth value (SUPPORTS/REFUTES) and propositional content.
  3. Filtering: Claims failing verification were removed, resulting in a final dataset with a 92% pass rate.

B. Simulation Framework

The authors model information exposure as a sequential process using Monte Carlo trajectory sampling:

• Agent Personas: Instead of simple prompting, the authors use Supervised Fine-Tuning (SFT) with Low-Rank Adaptation (LoRA) to create "framing-sensitive" agents.
• Conditional Loss Attenuation: To induce bias without destroying task competence, the training objective is modified. For claims in the agent's target framing that are actually REFUTED, the training loss is down-weighted by a factor $\alpha$ (set to 0.3). This makes the model less likely to learn to reject false claims in that specific framing, simulating a cognitive bias while maintaining general accuracy.
• Interaction Dynamics:
  • Agents are exposed to a stream of claims.
  • If an agent supports a claim, the next claim is sampled based on cosine similarity (top $k$ similar claims), simulating algorithmic reinforcement.
  • If an agent refutes a claim, the next claim is sampled uniformly to prevent topic drift.
• Information Health Metric ($H_t$): A cumulative score updated at each step:
  • $H_t = H_{t-1} + P(a_t|s_t)$ if the judgment is correct.
  • $H_t = H_{t-1} - P(a_t|s_t)$ if the judgment is incorrect.
  • This formulation penalizes confident errors more heavily than uncertain ones, reflecting the higher harm of confident misinformation.

3. Key Contributions

1. FrameRef Dataset: The first large-scale dataset providing controlled, generative reframings across multiple dimensions, validated by human evaluation to show systematic effects on judgment.
2. Simulation Testbed: A framework for modeling sequential information exposure where small biases compound over time, allowing researchers to stress-test ranking and personalization policies.
3. Framing-Sensitive Personas: A method for creating LLM agents that exhibit targeted, framing-dependent biases (e.g., higher credulity to "Consensus" framed false claims) while preserving overall task competence via controlled loss attenuation.
4. Open Resources: Release of the dataset, code, human evaluation data, and trained persona adapters.

4. Results

A. Simulation Results (Agent Personas)

• Divergent Trajectories: Small biases in initial acceptance rates led to substantial divergence in cumulative Information Health ($H_T$) over 100 steps.
• Bias Mechanism: Personas trained with loss attenuation on specific framings (e.g., Authoritative, Consensus, Sensationalist) showed:
  • Increased Mean Probability of SUPPORTS on Refuted claims (MSPR).
  • High confidence in these incorrect acceptances.
  • Reinforcement Loop: Because the simulation rewards similarity, incorrect "Support" judgments triggered the selection of more similar (and likely false) claims, causing the agent's Information Health score to plummet rapidly.
• Performance Trade-off: The Emotional framing persona did not show increased false acceptance, suggesting the loss attenuation mechanism did not translate to this specific framing type in the same way.

B. Human Evaluation

• Validation: Human participants (N=120) evaluated claims from FrameRef.
• Findings:
  • Framing significantly affected the acceptance of false claims ($p=0.038$) but not true claims.
  • Familiarity Effect: The impact of framing was moderated by topic familiarity. For high-familiarity topics, framing effects were strongest (e.g., Emotional and Prestige framings led to significantly higher false acceptance rates).
  • Confidence vs. Accuracy: As familiarity increased, participants became more confident but less accurate in rejecting false claims, mirroring the "confident error" phenomenon observed in the simulations.

C. Model Scaling

• Smaller models (Llama-3.2-3B) saturated faster with bias (requiring fewer training samples) compared to larger models (Llama-3.1-8B), indicating that bias induction is highly sensitive to model capacity and training scale.

5. Significance and Implications

• Methodological Advancement: The paper moves beyond static classification tasks to model dynamic, cumulative information health. It demonstrates that long-term vulnerability to misinformation is not just about factual ignorance but about how presentation styles interact with reinforcement loops.
• System Design: The framework provides a tool for system designers to "stress-test" recommendation algorithms. By simulating how different user personas (with different framing sensitivities) interact with a system, designers can identify policies that inadvertently amplify misinformation for specific user groups.
• Responsible AI: The work highlights the importance of "framing-aware" interventions. It suggests that mitigating misinformation requires understanding not just the content of a claim, but the contextual framing that makes it persuasive to specific cognitive profiles.
• Future Directions: The authors note limitations, such as the lack of social context and memory in agents, and propose extending the framework to include moral/political framing and other cognitive biases like priming.

In summary, FrameRef provides the infrastructure to study how the presentation of information, combined with algorithmic reinforcement, shapes long-term information health, offering a controlled, scalable alternative to purely empirical human studies.