PROBE-SWE Finds Chain-of-Thought and Self-Debiasing Don't Reduce Prompt-Induced Bias in Coding Agents

A paper accepted at FSE’2026 tests the assumption that chain-of-thought prompting or explicit self-debiasing instructions can steer coding agents away from cognitive shortcuts. They cannot — at least not reliably. PROBE-SWE (arXiv 2604.16756)¹, revised April 22, 2026, benchmarks eight software-engineering dilemmas across eight cognitive biases and finds no statistically significant per-bias reduction from either strategy. What does reduce bias: injecting axiomatic reasoning cues before the model answers, which cut overall bias sensitivity by 51% on average (p < .001)¹.

What PROBE-SWE Measures and Why SE Dilemmas Are a Blind Spot

Existing bias evaluations tend to target general reasoning or factual recall. Software engineering decisions — which dependency to pin, whether to defer technical debt, how to interpret a test suite’s signal — carry their own structure: they involve incomplete information, time pressure, and stakeholders who often anchor the framing before the agent sees the problem.

PROBE-SWE isolates prompt-induced bias, which is distinct from the data-induced variety. The precursor 2025 work from the same authors (arXiv 2508.11278)² found that GPT, LLaMA, and DeepSeek exhibited 6–35% bias sensitivity when training data shaped the response, rising to 49% as task complexity increased². The 2026 paper holds data constant and varies only the wording — the same technical dilemma, phrased to activate or not activate a specific cognitive shortcut. That isolation is methodologically important: it means the results speak directly to prompt engineering decisions, not model selection.

The Eight Biases and the Paired-Dilemma Design

PROBE-SWE covers eight biases selected for SE relevance¹:

• Anchoring — over-weighting an initial estimate (e.g., a performance number cited in a ticket)
• Availability — over-weighting the most recent or memorable incident (e.g., the last outage)
• Bandwagon — deferring to stated consensus or team majority
• Confirmation — preferring evidence that validates the existing approach
• Framing — shifting judgment based on whether a choice is presented as a loss or a gain
• Hindsight — asserting that a past outcome was predictable when it was not
• Hyperbolic discounting — overvaluing immediate fixes over deferred but higher-value solutions
• Overconfidence — expressing higher certainty than the evidence supports

Each bias gets a paired dilemma: two prompts presenting the same underlying technical situation, one phrased to activate the bias, one neutral or reversed. The gap in the model’s response between the two conditions is the bias sensitivity score. This design controls for content, which means observed differences are attributable to wording alone.

Why Chain-of-Thought and Self-Debiasing Flopped

Chain-of-thought prompting asks a model to reason step-by-step before answering. Self-debiasing instructs the model to explicitly consider and discount its potential biases before committing to a response. Both strategies have genuine utility for certain failure modes. For prompt-induced cognitive bias in SE dilemmas, PROBE-SWE finds neither produces statistically significant per-bias reduction¹.

The likely mechanism: chain-of-thought externalizes whatever reasoning the model would have done internally — including the biased reasoning. If a prompt frames a dependency choice using an anchored estimate, a chain-of-thought trace will walk through that anchor with apparent rigor, not around it. Self-debiasing faces a related problem: a model instructed to “consider your biases” must know which bias the prompt is activating and which direction it pushes. Without an explicit map from prompt structure to bias type, the instruction is too underspecified to act on.

What Actually Worked: Axiomatic Reasoning Cues

The approach that achieved a 51% average reduction (p < .001)¹ works in two steps: first, elicit software engineering best practices relevant to the decision domain; second, inject those practices as axiomatic reasoning cues into the prompt before the model answers the dilemma.

The distinction from self-debiasing is structural. Self-debiasing asks the model to apply a general epistemic correction. Axiomatic cues provide a concrete normative anchor: established SE principles that the model must reason against before forming a judgment. Rather than “be careful of your biases,” the model receives “here are the applicable engineering constraints, now reason from these.” The bias is crowded out by domain-specific structure rather than flagged abstractly.

The 51% figure is an average across all eight biases. Per-bias reduction likely varies — the paper reports this as the overall result, and the brief does not break out individual bias figures — but the p < .001 significance across the aggregate is a strong signal that the effect is not driven by a favorable subgroup.

What This Means for Agent Frameworks

CrewAI’s blog output between January and April 2026 — including posts on self-evolving agents (March 17) and agent harnesses (April 14) — does not address prompt-induced bias mitigation³. LangChain’s April 2026 posts on swarms of coding agents (April 17) and background subagents (April 16) similarly omit the topic⁴. That silence matters because both frameworks construct and hand off prompts across task boundaries in ways that can compound bias.

In a CrewAI task handoff, the output of one agent becomes part of the input context for the next. If the first agent’s output encodes an anchored estimate or a framed recommendation, subsequent agents receive that framing as ground truth. The bias is baked into the handoff payload, not the model. Chain-of-thought at any individual step will not undo it.

In an AutoGen group chat, a biased agent response that achieves consensus — even manufactured consensus from a biased prompt — can short-circuit subsequent reasoning via bandwagon dynamics. The same applies to Pydantic AI tool loops where tool-call results feed back into the next prompt iteration: a framed tool result can steer the next call without any agent recognizing it as framing.

The fix PROBE-SWE identifies — axiomatic reasoning cues injected before the model answers — maps onto pre-prompt hooks that most frameworks support in principle. The gap is that no major framework, as of April 2026, exposes a standardized mechanism for injecting bias-mitigating cues at task boundaries³⁴. That burden falls on the application developer.

Linguistic Patterns That Predict Vulnerability

The paired-dilemma design produces a byproduct beyond the aggregate bias score: the wording deltas between neutral and biased prompt variants identify which linguistic constructions activate each bias. Because the benchmark holds all other variables constant, the constructions that correlate with high sensitivity scores are, by elimination, the structural features of the prompt that trigger the bias.

For orchestration developers, this makes bias vulnerability a tractable inspection problem. A framework that exposes prompt construction to static analysis can, in principle, flag high-risk wording patterns before execution — numeric references preceding decision questions for anchoring risk, consensus framing for bandwagon risk, outcome descriptions cast asymmetrically for framing risk. The paper does not ship a linter, but the benchmark design implies that one is possible. The specifics would require reading the full experimental results in arXiv 2604.16756¹.

Takeaway: Bias-Aware Orchestration, Not Better Models

PROBE-SWE shifts where the work needs to happen. If chain-of-thought and self-debiasing do not reduce prompt-induced cognitive bias, upgrading to a more capable model is not the mitigation. The bias is in the prompt structure, and the correction must be in the prompt structure.

For teams running multi-agent SE workflows, that means three concrete things. First, audit prompt templates at task boundaries for the eight PROBE-SWE patterns — especially anchoring, bandwagon, and framing, which are the most natural to introduce accidentally during task handoffs. Second, implement axiomatic cue injection as a pre-answer step in each agent’s prompt — not as a generic self-correction instruction, but as domain-specific SE best practices relevant to the task type. Third, do not assume that framework-level reflection loops, critic agents, or self-revision cycles substitute for upstream prompt hygiene; PROBE-SWE provides no evidence that they do¹.

The FSE’2026 acceptance of arXiv 2604.16756¹ means this benchmark is positioned to become a reference point for SE agent evaluations. Teams that wait for frameworks to absorb the lesson will be running biased agents in the meantime.

Frequently Asked Questions

What happens when a task domain lacks established best practices to use as axiomatic cues?

The 51% reduction depends on injecting domain-specific SE principles as normative anchors. Teams working in nascent or poorly documented domains — emerging infrastructure tooling, proprietary internal platforms — may have thinner cue libraries to draw from, and the paper does not report results for artificially sparse cue sets. Cue effectiveness likely scales with the maturity and consensus of available best practices for the target decision type, making cue curation a prerequisite, not an afterthought.

How does implementing axiomatic cues work in CrewAI’s orchestration model?

CrewAI’s March 2026 integration with NVIDIA NemoClaw for self-evolving agents demonstrates dynamically constructed prompts across multi-agent handoffs — exactly the pattern where bias compounds. Injecting axiomatic cues would require a pre-task callback that prepends domain-specific SE principles into each agent’s context before it answers. As of April 2026, CrewAI exposes no standardized hook for this at task boundaries, so developers must implement custom callbacks per node in the task graph.

Does the 51% bias reduction hold as task complexity increases?

The precursor 2025 study established that bias sensitivity scales with task complexity — up to 49% at higher levels. PROBE-SWE reports the 51% axiomatic-cue reduction as an average across its benchmark dilemmas at fixed complexity, but does not break out whether cue effectiveness degrades as complexity rises. Teams running deep multi-step agent chains should validate cue performance at their actual task depth rather than assuming the headline figure transfers linearly.

How does PROBE-SWE differ from SWE-Bench in what it evaluates?

SWE-Bench measures whether an agent produces a correct patch for a real GitHub issue — it evaluates outcome correctness given a fixed prompt. PROBE-SWE evaluates decision-making integrity by varying only the wording of the same technical dilemma and measuring the delta. An agent can score well on SWE-Bench while exhibiting high bias sensitivity on PROBE-SWE, because SWE-Bench never varies the framing of its tasks. The benchmarks are orthogonal: one tests capability, the other tests robustness to irrelevant prompt structure.

Footnotes

1. https://arxiv.org/abs/2604.16756 ↩ ↩² ↩³ ↩⁴ ↩⁵ ↩⁶ ↩⁷ ↩⁸

2. https://arxiv.org/abs/2508.11278 ↩ ↩²

3. https://www.crewai.com/blog ↩ ↩²

4. https://www.langchain.com/blog/ ↩ ↩²