Technical Summary: Framing Matters: Addressing Framing Sensitivity in Decision-Making through Behaviorally-Grounded Value Alignment

1. Problem Statement

Large Language Models (LLMs) are increasingly deployed in high-stakes decision-making domains such as legal reasoning, medical triage, and moral judgment. A critical requirement for these systems is consistency: the model should make the same decision when presented with factually equivalent inputs, regardless of how those inputs are linguistically framed.

However, this paper identifies a significant vulnerability: LLMs exhibit high susceptibility to framing effects, where fact-preserving but semantically reframed inputs cause the model to reverse its decisions. Unlike human cognitive biases (e.g., loss aversion, present bias) which are well-documented, the internal mechanisms driving framing sensitivity in LLMs remain poorly understood. Existing mitigation strategies, such as prompt-level instructions or activation-based steering, often fail to suppress this sensitivity and, in some cases, actively amplify it. The core problem is that current interventions address surface symptoms rather than the underlying representational pathways through which framing operates.

2. Methodology

2.1 The FRAGILE Benchmark

To systematically investigate this issue, the authors introduce FRAGILE, a large-scale benchmark designed to isolate fact-preserving semantic framing across four high-stakes domains:

1. Moral Dilemma (GGB, UNIBENCH)
2. Medical Triage (TRIAGE, MEDICAL TRIAGE ALIGNMENT DATASET)
3. Legal Judgment (SUPER-SCOTUS)
4. Interpersonal Role Conflict (ROLECONFLICTBENCH)

The benchmark transforms base scenarios along three controlled, orthogonal dimensions derived from psychological theory:

• Value-Tinted Narration: Embeds latent value orientations (based on Schwartz's human values) into the scenario without altering facts (e.g., framing an action as "protecting others" vs. "autonomy").
• Temporal Slice: Shifts the perceived time horizon of consequences (short-term vs. long-term) at the option level.
• Narrative Vividness: Modulates experiential concreteness through action-oriented phrasing (e.g., "hit censor" vs. "censor the post").

The benchmark includes 48,140 instances, incorporating label-swapped variants to control for positional bias. Framing sensitivity is quantified via Flip Rate (behavioral inconsistency) and Distributional Shift (L1 distance between decision distributions).

2.2 Diagnostic Analysis

The authors employ Logit Lens analysis to project hidden states at the decision token into vocabulary space. This allows them to trace how framing alters internal representations layer-by-layer. The analysis reveals that framing-induced decision flips are not early-layer phenomena but emerge from late-layer representation shifts concentrated near the decision boundary. Distinct token-level signals were identified for each framing type:

• Value-Tinted: Distributed value-space reorganization.
• Temporal: Activation of urgency pathways (e.g., tokens like "immediate," "OUT").
• Vividness: Ambivalence amplification or scene-evocative imagery.

2.3 The VALIGN Framework

To mitigate these effects, the paper proposes VALIGN, a representation-level intervention framework that anchors decisions to a stable value prior. It operates in three stages:

1. Self-Report Vector Extraction: The model is prompted to articulate its core values. The hidden state at a specific layer (approx. 65–70% depth) is extracted to form a Value Vector ($\hat{w}$), representing the model's intrinsic value orientation.
2. Sensitivity Subspace Construction: Using PCA on the difference vectors between framed and base hidden states, the authors identify a low-dimensional Temporal-Vividness Sensitivity Subspace ($\{\hat{v}_1, \dots, \hat{v}_K\}$) where representations are most sensitive to framing cues.
3. Inference-Time Intervention: During the forward pass, the hidden state ($h$) is modified via:
   • Value Steering: Adding the value vector to steer the state toward value-consistent directions: $h^* = h + \alpha \hat{w}$.
   • Orthogonal Projection: Projecting out the components of the sensitivity subspace from the steered state: $h' = h^* - \sum (h^* \cdot \hat{v}_i)\hat{v}_i$.

This process is combined with a text-level Value Anchor system prompt instructing the model to reason from core values rather than surface cues.

3. Key Results

3.1 Framing Sensitivity

Experiments on LLaMA-3.1-8B, Mistral-7B, and Qwen2.5-7B reveal severe framing instability:

• Average Decision Flip Rate: 28.6% across all models and framing types.
• Value-Tinted Narration induces the strongest instability (Avg Flip 36–47%), followed by Temporal Slice (21–38%), while Narrative Vividness yields weaker but still significant effects (~8–18%).
• Baseline Failures: Existing mitigation strategies (Instruction prompting, Chain-of-Thought, Concept Activation Vectors) consistently amplified framing sensitivity, increasing flip rates by 20–40 percentage points in many cases.

3.2 VALIGN Performance

VALIGN demonstrates consistent and significant mitigation across all models and framing types:

• Flip Rate Reduction: Reduced the average flip rate from 28.6% (Base) to 13.2% (VALIGN).
• Stability Gains: The proportion of stable, low-perturbation responses (NL%) increased from 67.1% to 81.9%.
• Specific Improvements:
  • Under Value-Tinted Narration, flip rates dropped from ~40% to ~13%.
  • Under Temporal Slice, flip rates dropped from ~30% to ~13%.
  • The method effectively suppresses high-confidence flips (FH%), indicating that the model's internal belief distributions are becoming more robust, not just the final output.
• Generalization: The method remains effective on larger models (LLaMA-3.1-70B) and in multi-framing scenarios where multiple framing cues are combined.

3.3 Ablation Studies

Ablation studies confirm that all three components of VALIGN are necessary:

• Removing the Temporal-Vividness Projection causes the largest degradation, particularly for temporal and vividness framing.
• Removing the Text Value Anchor increases flip rates, showing that explicit grounding is required.
• Removing Value Steering reduces the stabilization of internal belief distributions (NL% drops).

4. Significance and Claims

The paper claims that achieving framing-invariant decision-making in LLMs requires deep, representational interventions rather than surface-level prompt engineering.

• Mechanistic Insight: The work provides evidence that framing sensitivity is driven by specific internal pathways (late-layer representation shifts) that can be identified and neutralized.
• Limitations of Current Methods: It highlights that standard alignment techniques (prompts, CoT) are insufficient and can be counterproductive for stability, as they fail to address the root representational causes of framing bias.
• Robustness via Value Alignment: The findings suggest that anchoring model decisions to a stable, intrinsic value profile can effectively decouple decision-making from transient framing cues, offering a pathway to more reliable AI in high-stakes domains.

The authors maintain a modest stance, acknowledging that while the benchmark focuses on binary decisions for controlled evaluation, the underlying mechanisms likely extend to more complex, multi-choice decision spaces (which they briefly validate in the appendix). They also note that while VALIGN improves stability, there may be trade-offs with task-specific accuracy in certain domains (e.g., role conflict), suggesting that future work must balance consistency with normative correctness.