Technical Summary: Diagnosing Intrinsic Over-Calling Bias in LLM Agents

Problem Statement

Large Language Model (LLM) agents have evolved to utilize external tools (e.g., search engines, APIs), yet they exhibit a systematic failure mode known as over-calling: the tendency to invoke tools even when the user's request lacks necessary information or when a direct answer is sufficient.

Evaluation on the When2Call benchmark reveals a consistent asymmetry across six models from three families (Qwen3.5, Gemma-3, Ministral-3). While these models achieve high accuracy when a tool call is required (Call Accuracy: 90%–99%), their accuracy drops significantly when the correct action is not to call a tool (No-Call Accuracy: 2%–69%). This results in an overall accuracy range of only 55%–70%. This bias degrades user experience (e.g., hallucinating tool calls for underspecified requests) and inflates API costs in deployed systems.

Methodology

The authors propose a mechanistic investigation into the internal decision-making process of LLM agents, moving beyond empirical observation to causal analysis.

1. Intrinsic Bias Hypothesis (IBH)

The paper posits that over-calling is not merely a function of activation levels but stems from an activation-independent CALL offset in the decision mapping.

• Hypothesis: The decision to call or not call is determined by a signed activation margin ($m$) between "CALL" and "NO_CALL" feature directions. However, the decision boundary is shifted such that even at activation parity ($m=0$), the model favors calling. The model only becomes decision-neutral when the NO_CALL activation significantly outweighs the CALL activation.
• Formalism: The decision probability is modeled as $P(\text{CALL} | m) = \sigma(\beta m + \beta_0)$, where $\beta_0 > 0$ represents the intrinsic bias offset.

2. Feature Discovery via Sparse Autoencoders (SAEs)

To test IBH, the authors employ Sparse Autoencoders to decompose residual-stream activations into interpretable features.

• Training: TopK SAEs are trained on the residual streams of target models using a two-stage curriculum: first on a broad corpus (OpenWebText2) to learn general geometry, and second on the When2Call dataset to adapt to tool-use contexts.
• Feature Bases: Two behavior-aligned feature sets are recovered:
  • $\mathcal{C}$: Features strongly associated with observed CALL decisions.
  • $\mathcal{N}$: Features strongly associated with observed NO_CALL decisions.
• Validation: Linear probes trained on a handful of these features (typically $K \le 5$) predict the gating decision with accuracy approaching the upper bound of the raw residual stream, confirming these features capture the essential gating signal.

3. Diagnosis and Quantification

Using the recovered feature bases, the authors analyze the activation geometry:

• Response-Conditioned View: NO_CALL responses concentrate in the NO_CALL-dominant half-plane, while CALL responses extend across the parity diagonal into regions where NO_CALL evidence is comparable or stronger.
• Input-Conditioned View: Among emitted CALL responses, "false calls" (over-calling errors) lie closer to the parity diagonal than "true calls," indicating the bias pushes weak cases over the decision boundary.
• Offset Estimation: Logistic regression on the signed activation margin confirms a negative neutral boundary ($m^* < 0$) for all models, quantifying the intrinsic bias $\beta_0$.

4. Causal Intervention: Adaptive Margin-Calibrated Steering (AMCS)

To validate IBH causally, the authors introduce AMCS, a closed-form steering method.

• Mechanism: AMCS computes a steering vector along the SAE decoder directions of $\mathcal{C}$ and $\mathcal{N}$. It applies a counter-bias shift $\delta = -\beta_0 / \beta$ to the activation margin.
• Implementation: The shift is distributed across the top-ranked features based on their activation gaps. The intervention is applied as a single vector addition to the residual stream at inference time, requiring no additional model calls or iterative tuning.

Key Results

Experiments were conducted on six instruction-tuned models (Qwen3.5-4B/9B, Gemma-3-1B/4B, Ministral-3-3B/8B).

• Mitigation of Over-Calling: Cancelling the diagnosed offset via AMCS significantly improves No-Call Accuracy by 4 to 17 percentage points on five of the six models.
• Preservation of Valid Calls: The method maintains Call Accuracy with a negligible drop (within 5 points) on five models, demonstrating that the intervention corrects the bias without suppressing necessary tool usage.
• Overall Accuracy: AMCS improves overall accuracy by up to 5 points on several models (e.g., +5.10% on Qwen3.5-9B).
• Comparison to Baselines:
  • Prompt Engineering: While effective at increasing No-Call accuracy, it drastically reduces Call accuracy (e.g., -56% on Ministral-3-8B), leading to a net loss in overall performance.
  • One-Sided Steering: Methods that only suppress CALL features or only promote NO_CALL features fail to achieve the balance AMCS provides, yielding smaller gains in overall accuracy.
• Gemma-3 Limitation: The Gemma-3 models showed little separation between true and false calls along the SAE axes, resulting in minimal leverage for margin-based interventions. These were treated as reference cases rather than primary evidence.

Significance and Claims

The paper claims to recast over-calling from an empirical phenomenon into a mechanistic object that can be measured, modeled, and causally adjusted.

1. Mechanistic Insight: The work demonstrates that tool-use gating is governed by an intrinsic, activation-independent bias rather than just the relative strength of feature activations.
2. Causal Correction: By identifying and cancelling the specific offset term ($\beta_0$) via SAE-based steering, the authors provide a causal validation of the Intrinsic Bias Hypothesis.
3. Controllability: The diagnosed offset acts as a "knob" for calling effort. AMCS serves as a first step toward a "calling effort" interface, allowing systems to tune the balance between eager tool invocation and user verification without retraining the model.
4. Scope: The study focuses on inference-time correction for deployed models. It does not trace the training-time origins of the bias, nor does it claim to solve long-horizon agent planning issues beyond the specific call-versus-verify decision unit.

The code and data are available at https://github.com/SKURA502/agent-sae/.