When building Large Language Model (LLM) systems for administrative tasks—like deciding whether to grant a refund or approve an insurance claim—developers often use Retrieval-Augmented Generation (RAG). The idea is simple: show the model the current request alongside a few "precedents" (similar past cases) to ensure consistency.
However, a new paper titled "Precedent Labels and Facts Shift LLM Approval Decisions in Opposite Directions" reveals a fascinating and counter-intuitive tug-of-war happening inside the model’s "mind." It turns out that the labels of past decisions and the facts of those cases pull the LLM in completely opposite directions.
TL;DR
When an LLM evaluates a borderline request, seeing past cases with "APPROVE" labels makes it much more likely to approve. Conversely, seeing the facts of those same approvable cases (without the labels) actually makes the model less likely to approve, as the current request looks weaker by comparison—a phenomenon known as the "contrast effect."
What Problem Does This Paper Address?
As we move toward "AI-as-a-judge" systems, consistency is the gold standard. If a customer gets a refund for a broken screen today, a similar customer should get one tomorrow. RAG is the primary tool used to achieve this "case law" style of grounding.
The authors wanted to know: what exactly is the model learning from these precedents? Is it following the logic of the past facts, or is it just being swayed by the labels (the final "APPROVE" or "DENY" verdict)? Understanding this is critical for building fair systems that don't just blindly anchor to past mistakes.
Methods
The researchers tested three major models: DeepSeek V4.1 Flash, GPT-5.6 Luna, and Claude Haiku 5.5. They focused on "borderline" cases—requests where the model, when left to its own devices, was roughly 50/50 on whether to approve or deny.
The experimental design was clever in its simplicity. For a fixed "target" request, they provided three precedents and systematically varied two things:
- The Label Effect: They kept the facts of the precedents the same but flipped the recorded decisions between APPROVE and DENY.
- The Facts Contrast: They removed the decision labels entirely and changed the facts of the precedents to be either clearly "approve-side" (meeting all criteria) or "deny-side" (failing criteria).
Datasets
The study utilized two distinct sets of materials:
- Public Policies: 9 real-world policies (e.g., Steam refunds, IRS penalty relief, Airbnb rebooking) paraphrased into 5–6 clear rules. They generated 65 borderline target requests for these.
- Fictional Policies: 6 entirely new organizations (retailers and insurers) to ensure the models hadn't seen the data during training.
- Human Benchmarks: Three human annotators provided a "ground truth" consensus on 40 of the targets to see if the model's shifts were corrections toward the "right" answer or merely biases.
Key Insights and Ideas
- The Tug-of-War: Labels and facts are in a constant battle. Labels act as an anchor: if the model sees "APPROVE" on similar cases, it feels social pressure to do the same. Facts, however, trigger a contrast effect: if the precedents are "perfect" examples of a valid claim, the borderline target request looks "shabby" or "weak" by comparison.
- Similarity Amplifies Labels: The "pull" of the label is strongest when the precedents are near-copies of the target request. When the precedents are from a different type of request (even under the same policy), the label effect shrinks significantly.
- The Contrast Effect is a Bias, Not a Correction: One might hope the "facts push" helps the model find the right answer. However, human annotation showed the push happened even on requests that humans judged as clearly approvable. The model wasn't getting "smarter"; it was just being biased by the comparison.
Key Results
The quantitative shifts were massive, especially for DeepSeek:
- Label Power: Switching precedent labels from DENY to APPROVE raised approval rates by 30 to 55 percentage points across the three models.
- The Facts Push: In the absence of labels, showing "approve-side" facts (which should theoretically encourage approval) actually lowered DeepSeek's approval rate by 26.2 points.
As shown in the figure below, the label effect ($L$) is consistently positive (pulling toward the label), while the facts contrast ($F$) is consistently negative (pushing away from the "strong" facts).
Limitations and Future Work
The authors acknowledge several constraints:
- Single-Step Logic: The study used relatively short fact sheets and single-call decisions. Real-world claims are often much more complex.
- Borderline Focus: The effects are most pronounced on "close calls." Clear-cut cases (e.g., a refund request 100 days past the limit) are much harder to sway.
- Retrieval Dynamics: The study used "near-copy" variants. In a real production system, the retrieval tool might not always find such close matches, which could change the balance between the label and fact effects.
The takeaway for engineers? If you're building a RAG system for decision-making, test your labels and facts separately. Don't assume that showing "good" precedents will lead to "good" decisions; you might accidentally be setting a bar so high that your model starts denying perfectly valid requests.
Summary updated