For instance, miscalibration for vagueness dropped from 51.3% to 28.5% and for jargon from 50.3% to 33.2% after CDA.

Even joint debiasing across multiple biases (length, vagueness, jargon) proved effective with minimal impact on general capabilities.

And the results? CDA works!

It significantly reduced average miscalibration (e.g., from 39.4% to 32.5%) and brought model skew much closer to human preferences. All this while maintaining overall performance on RewardBench!

So how do we debias models? We propose a simple yet effective post-training method based on counterfactual data augmentation (CDA).

We synthesize contrastive responses that explicitly magnify biases in dispreferred responses, & further finetune reward models on these responses.

Indeed, preference models can easily latch on to these subtle data artifacts!

Features that only weakly correlate with human preferences (r_human=−0.12) are strongly predictive for models (r_model​=0.36). Points above y=x suggest that models overrely on these spurious cues😮

Where do these biases come from?🤔Our analysis suggests they originate from training data artifacts.

For eg, humans preferred structured responses >65% of the time when the alternative wasn't structured. This gives an opportunity for models to learn these patterns as heuristics!

How severe is the problem? Using controlled counterfactual pairs, we found that preference models (incl. LLM evaluators) prefer biased responses in >60% of cases (defined as skew) and show high miscalibration (~40%) wrt humans.

Vagueness & sycophancy are especially problematic!

Preference models act as proxies for human judgements in alignment (as reward models) & evaluation, but they can be miscalibrated.

We found that they overrely on many idiosyncratic features of AI-generated text, which can lead to reward hacking & unreliable evals. Features like:

Ever wondered what makes language models generate overly verbose, vague, or sycophantic responses?

Our new paper investigates these and other idiosyncratic biases in preference models, and presents a simple post-training recipe to mitigate them! Thread below 🧵↓

🤔 How can we use context to learn more about model behavior?

We can study "default" responses from models. Under what type of context does their response get highest score?

We uncover a bias towards WEIRD contexts (Western, Educated, Industrialized, Rich & Democratic)!

🤔 Does providing context to evaluators have a substantial effect on evaluation conclusions?

We find that (1) presence of context can improve agreement between evaluators and (2) even change model rankings! 🤯

...we then conduct experiments providing context (1) during response generation, (2) during evaluation or (3) both.

With ✨Contextualized Evaluations✨, we synthetically generate context as clarifying, follow-up questions to an underspecified query...

Underspecified queries can lead to arbitrary evaluation judgments of response quality!

e.g., Given a query “Is coffee good for you?”, how can evaluators accurately judge model responses when they aren't informed about the user’s preferences, background or important criteria?

Underspecified queries are prevalent in many datasets used to benchmark language models (e.g., Chatbot Arena, AlpacaEval).

These can be ambiguous (e.g., what is a transformer? ... 🤔 for NLP or EE?), subjective (e.g., who is the best? ... 🤔 what criteria?), and more!

Excited to share ✨ Contextualized Evaluations ✨!

Benchmarks like Chatbot Arena contain underspecified queries, which can lead to arbitrary eval judgments. What happens if we provide evaluators with context (e.g who's the user, what's their intent) when judging LM outputs? 🧵↓