Inference-Time Reasoning: cutting hidden bias by switching on reasoning mode

This isn't a step-by-step technique, it's an operating principle:

STANDARD MODE:
Request → Model leans on statistical associations from data → Answer with implicit bias

REASONING MODE:
Request → Model explicitly spells out its steps → Leans on explicit values from alignment → Answer with less bias

How to turn it on: - o3: on by default - Claude Opus/Gemini: enable the "extended thinking" setting (in the API or the app) - GPT-4/Llama/others: add to your prompt: "Think step by step"

Task: You're an HR manager at a startup. You need to pick a candidate for head of product. You have two finalists: Anna (10 years at Google, two kids, recently back from maternity leave) and David (8 years at Meta, single, speaks at conferences a lot). You want an unbiased read on their potential.

Prompt without reasoning (GPT-4):

Evaluate these two candidates for head of product:

Candidate A: 10 years of experience, two kids, recently back from maternity leave
Candidate B: 8 years of experience, single, speaks at conferences a lot

Who is the better fit for a role that requires full commitment and frequent travel?

Prompt with reasoning (o3 or GPT-4 + CoT):

Evaluate these two candidates for head of product. **Think step by step and check yourself for bias.**

Candidate A: 10 years of experience, two kids, recently back from maternity leave
Candidate B: 8 years of experience, single, speaks at conferences a lot

Which factors actually matter for success in this role? Which of my assumptions might be wrong?

Result:

In the first case the model may automatically link "kids = family = less time" and lean toward Candidate B without any explicit justification.

In the second case the model spells out its steps: it separates facts (experience) from stereotypes (kids = low commitment), points out that public speaking isn't the same as management skill, and gives a more balanced assessment of both candidates. In GPT/Claude the difference will be noticeable. In Gemini/Llama, probably not.

An LLM learns two kinds of knowledge: statistical patterns from text (implicit) and explicit values from alignment training (explicit). Statistics say "career" shows up next to male names more often in the corpus: that's an implicit association. Alignment says "sex discrimination is bad": that's an explicit value.

Standard mode is fast and runs on automatic associations from statistics. Like human intuition, it fires instantly but can reproduce the stereotypes of its environment. The model isn't "thinking racist thoughts"; it's simply reflecting that "Black" appears next to words about guilt more often in news text.

Reasoning mode forces the model to verbalize its steps and check them. That engages the explicit layer, the values it was taught after pretraining. The model sees: "I'm about to link 'guilt' with 'Black', but I know that's a racist stereotype," and corrects its answer. In essence, reasoning switches from autopilot (statistics) to deliberate driving (values).

Why only for social topics: Because alignment training taught the model "racism is bad, sexism is bad," but never taught it "the word 'cause' sounds negative even though it's technically neutral." For non-social associations the model has no explicit values that could override the statistics.

Control knobs: - An explicit self-check instruction: "check yourself for bias," "which stereotypes might I have missed" → strengthens the effect - Model choice: GPT/Claude respond strongly, Gemini/Llama barely at all → different alignment strategies - Type of stereotype: unambiguous ones (race-guilt, gender-career) drop more than ambiguous ones (skintone, weight) → ambiguous ones have alternative readings

{your_task}

Think step by step:
1. Which facts do I know for certain?
2. Which assumptions am I making?
3. Which of those assumptions might be stereotypes?
4. How would I assess the situation using only the facts?

{additional_context_if_needed}

What to fill in: - {your_task}: a question or task where fairness matters (evaluating people, decisions about groups, advice on sensitive topics) - {additional_context}: a description of the people or situation with facts (experience, skills) and no value judgments

Version for models with a reasoning setting: If you're using o3, Claude or Gemini with extended thinking turned on, just ask the question. The model will apply reasoning automatically. Explicit instructions like "check for stereotypes" will strengthen the effect.

⚠️ Model-dependent: In Gemini 2.5 and Llama 3.3 the effect is close to zero. It works in GPT (4.1→o3) and Claude Opus 4.1. Different model families went through different alignment training, which affects how much reasoning can engage explicit values.

⚠️ Not for every stereotype: Ambiguous words respond more weakly. For example, "dark-light" (skintone) or "fat-thin" (weight) are often used outside a social context (dark chocolate, a thick book), so reasoning barely changes them. Unambiguous stereotypes (race-guilt, gender-career) drop by 80-90%.

⚠️ Social bias only: Reasoning doesn't affect non-social implicit associations. If the word "cause" often appears in negative contexts, the model will keep reflecting that even with reasoning, because there's no explicit value saying "treat the word 'cause' as neutral."

⚠️ No guarantee of fairness: A 70-90% drop in implicit bias is not zero. In sensitive applications (hiring, lending, medicine) you need additional checks, not just prompting techniques.

The team used the LLM Word Association Test, an adaptation of the psychological test for implicit bias (IAT). It works like this: the model gets a list of words (for example "career," "office," "salary" + "family," "kids," "home") and is asked to tag each word with either a male name (John, Michael) or a female name (Emily, Mary). If the model more often pairs "career-John" and "family-Emily," that reveals an implicit gender stereotype, even if it answers "of course" to the direct question "can women build careers?"

They tested 15 stereotypes (race, gender, religion, age, weight, mental health) across four model families: GPT-4.1 vs o3, Claude Opus 4.1 with/without thinking, Gemini 2.5 Flash with/without thinking, and Llama 3.3 with a plain prompt vs CoT. For each stereotype they ran 50 iterations with randomized word order.

The spread of results was surprising: GPT's bias dropped from 0.41 to 0.13 (−68%), and Claude did even better, from 0.59 to 0.18 (−91%). Gemini and Llama barely moved (−5% and −0.2%). The authors attribute this to different alignment strategies at each company: apparently OpenAI and Anthropic drilled in explicit anti-bias values more heavily, which reasoning could then activate.

The key experiment: a control on non-social words. They took words with semantic prosody: "cause" and "commit" sound negative (they often appear with bad things: cause damage, commit crime), even though they're technically neutral. "Restore" and "provide" have a positive coloring. Asked directly, people rate these words as neutral, but indirect tests reveal the associations. The models showed the same thing: neutral on explicit ratings, valenced associations on the implicit test. And reasoning had no effect on them at all. That proved the effect is specific to social stereotypes, where trained values exist, rather than a technical artifact that touches every implicit association.

It's an elegant insight: reasoning acts as a bridge between the two layers of an LLM's knowledge. The first layer is statistical knowledge from the corpus (inevitably carrying society's social biases). The second layer is ethical values from RLHF/alignment (explicitly against discrimination). Standard mode runs on the first layer. Reasoning engages the second, but only where it exists. There's a second layer for "racism is bad"; there isn't one for "the word cause is negative."

🔧 Technique: Beef up the self-check → a stronger effect

The original study just turned reasoning on. But you can explicitly ask the model to check for stereotypes:

{task about people}

Analyze step by step:
1. List the facts about each person
2. Which stereotypes might be shaping my perception (age, gender, nationality, appearance, family status)?
3. If you strip out those stereotypes, what's left?
4. Evaluate using only the remaining criteria

This makes the model explicitly verbalize "which stereotype might be in play here," which strengthens the activation of explicit values.

🔧 Technique: Blind evaluation → remove the triggers

Instead of fighting bias through reasoning, you can withhold the information that triggers stereotypes:

Evaluate a candidate for head of product:
- 10 years of experience at a large tech company
- Launched two successful products with 5M+ MAU
- Managed a team of up to 15 people
- A 1.5-year career break

What questions should I ask in the interview? What are the risks and strengths?

Name, gender and family status are gone, so the model can't latch onto "two kids → less commitment." This complements reasoning rather than replacing it: sometimes gender or race matters for context.

🔧 Technique: Counter-stereotypical few-shot → shift the baseline

Show the model examples that break stereotypes before the main task:

Here are three successful heads of product at well-known companies:

1. Maria, Amazon: three kids, runs a product with $150M in revenue
2. Ahmed, Google: moved from Egypt in 2015, rose to VP of Product
3. Tom, 58, Instacart: switched to tech after 30 years in retail

Now evaluate my candidates: [...]

Few-shot examples shift the model's "baseline": it sees diversity before it evaluates.

Inference-Time Reasoning Selectively Reduces Implicit Social Bias in Large Language Models: Molly Apsel, Michael N. Jones (Indiana University)

References from the paper: - Bai et al. (2024): LLM Word Association Test, the original method for measuring implicit bias in LLMs - Liu et al. (2024): a study of how reasoning affects tasks that rely on implicit learning - Greenwald et al. (1998): the original Implicit Association Test (IAT) paper for humans - Hauser & Schwarz (2015): semantic prosody, words with implicit valence