TL;DR
LLMs acting as agents suffer from a "multi-turn anchoring bias," where they tend to repeatedly reuse a previously selected tool even when better or more appropriate alternatives are available. This paper introduces TBMT, a benchmark to quantify this stubbornness, and TBFAIR, a surgical mitigation strategy that identifies and suppresses "bias-inducing neurons"—fixing the problem by modifying less than 1% of the model's parameters without breaking its ability to use tools correctly.
What Problem Does This Paper Address?
As Large Language Models (LLMs) transition from chatbots to autonomous agents, they increasingly rely on external tools (APIs) to perform real-world tasks like booking flights or checking stock prices. However, tool selection is notoriously sensitive to how APIs are described.
While previous research identified biases in single-turn interactions (like preferring the first tool in a list), this paper tackles the multi-turn dynamics. The authors discover that once an agent picks a tool in Turn #1, it develops a "memory" or preference for that tool. This "anchoring bias" leads to two major issues:
- Unfairness: Certain API providers get over-exposed simply because they were picked first.
- Unreliability: The model might reuse a tool for a task it isn't suited for, leading to errors.
Crucially, the authors found that this bias is incredibly persistent; even if you insert several turns of "small talk" between tool calls, the model still tries to go back to its original choice.
Methods
The researchers developed a two-part solution: a way to measure the bias and a way to kill it.
1. The TBMT Benchmark
They created a multi-turn evaluation framework that tests models across five domains (E-commerce, Finance, Music, Movies, and Travel). It uses three specific setups:
- General Setup: Can the model pick a correct tool when the previous tool is also a valid option?
- Interleaved Noise: Does the bias persist if there is unrelated conversation between tool calls?
- Incorrect Anchor: Will the model incorrectly reuse the first tool even when the second query explicitly requires a different one?
2. TBFAIR Mitigation
Instead of trying to "prompt" the bias away (which they show doesn't work well), the authors take a mechanistic approach:
- Neuron Attribution: Using gradient-based methods, they identify which specific neurons in the LLM's architecture are most responsible for the biased tool-calling behavior.
- Activation Scaling: They "turn down the volume" on these neurons by scaling their activations.
- Optimization: They use a binary search to find the minimum number of neurons to suppress to reach a "fair" selection rate (matching a random baseline of 20% for 5 tools) while ensuring the model's overall tool-calling accuracy (the "VALID" score) remains at 100%.
Datasets
The authors introduced the TBMT (Tool selection Bias in Multi-Turn) dataset.
- Composition: 500 queries across 5 categories, based on 73 functions from 18 real-world APIs.
- Synthetic Control: To ensure they were measuring anchoring bias and not just "brand familiarity," they used GPT-5 to generate synthetic, realistic-sounding API names and function descriptions.
- Noise Integration: They used the Alpaca dataset to simulate "noisy" intervening conversations.
Key Insights and Ideas
- Bias is Mechanistic, Not Just Contextual: The fact that prompting ("You are a fair agent") fails while neuron suppression succeeds suggests that anchoring bias is deeply baked into the model's internal representations of the conversation history.
- The "Sticky" Tool Phenomenon: Even with 10 turns of unrelated chat (the NOISE dataset), the anchor selection rate remained significantly higher than the random baseline. The model's "internal state" regarding the first tool is remarkably durable.
- Surgical Precision: You don't need to fine-tune the whole model. For Qwen3-4B, modifying just 1.72% of neurons was enough to reach the fairness target. In larger models like Qwen3-30B, the required fraction was even smaller (<1%).
- Transferability: Interestingly, neurons identified as "bias-inducing" in the E-commerce domain often overlap with those in the Travel domain. This suggests that "anchoring" might be a generalized behavior handled by specific circuits in the LLM.
Key Results
- Baseline Bias: Without intervention, models showed massive anchoring. For 5 available tools (where a "fair" rate is 20%), Qwen3-30B reused the anchor tool 76-89% of the time in E-commerce tasks.
- Mitigation Success: TBFAIR reduced the anchor selection rate from ~0.80 down to ~0.20 (the fair baseline) across almost all categories.
- Preserving Functionality: Throughout the intervention, the models maintained a 1.0 (100%) "VALID" score, meaning they didn't lose the ability to actually use the tools correctly; they just stopped being obsessed with the first one they saw.
- Failure of Prompting: The "DIRECT" baseline (explicitly telling the model "never reuse a tool") only reduced bias by about 5-10%, proving that multi-turn bias is resistant to simple instruction following.
Limitations and Future Work
- Synthetic Names: While synthetic names help control for external variables, real-world tool selection might involve more complex trade-offs (like cost or latency) that aren't captured here.
- Trajectory Length: The study primarily looks at two-turn tool interactions. Future work could explore if bias compounds further over 5, 10, or 50 turns.
- Model Specificity: While the method works on Qwen, Gemma, and GPT-5, the specific "bias neurons" likely differ between architectures, requiring a new attribution pass for every new model version.
Summary updated