Liang 2022 - Findings 14 to 25 Targeted Evaluations and Prompting
Summary
Findings 14-20 report the targeted evaluations (language, knowledge, reasoning, memorization, disinformation, bias, toxicity generation); finding 21 states that comprehensive standardized evaluation surfaces results “lying in plain sight” (text-davinci-002 sets a new state of the art on NarrativeQA at 74.4% ROUGE-L); findings 22-23 show extreme prompt and multiple-choice-adaptation sensitivity (accuracy from 79.1% to 30.2% for the same model and scenario); finding 24 shows upstream perplexity predicts downstream accuracy poorly across families; finding 25 shows scale predicts accuracy within a family but a threshold near 50B parameters separates top models.
Overview
Together with Liang 2022 - Findings 1 to 13 Core Scenarios, these make up the 25 top-level findings of Sec. 1.2.
Main Content
Finding 14: linguistic understanding
Language-modeling trends differ from core-scenario trends: GPT-NeoX (20B), OPT (175B), BLOOM (176B), GPT-J (6B) and OPT (66B) consistently have the lowest bits-per-byte on The Pile, TwitterAAE and ICE. On BLiMP all models perform fairly similarly overall and on each subset (morphology, syntax, semantics, syntax-semantics); the widest spread is on irregular forms, where text-davinci-002 and TNLG v2 (530B), most accurate on the core scenarios, are among the least accurate, perhaps from overgeneralized rules.
Finding 15: knowledge
text-davinci-002 is superior on all knowledge-intensive evaluations, with a sizable gap on TruthfulQA (62.0% vs 36.2% for Anthropic-LM v4-s3 (52B)). TNLG v2 (530B) is strong on NaturalQuestions (closed-book) and WikiFact: 38.5% vs 28.7% (NQ closed-book) and 34.3% vs 22.3% (WikiFact) against Anthropic-LM v4-s3, consistent with scale contributing to factual knowledge.
Finding 16: reasoning
Code models, especially code-davinci-002, consistently outperform text models, even on synthetic reasoning in natural language. GSM8K: code-davinci-002 52.1%, text-davinci-002 35.0%, no other model above 16%. Synthetic reasoning in natural language: text-davinci-002 65.1% vs next-best text model OPT (175B) 29.4%, code-davinci-002 72.7%.
Finding 17: memorization of copyrighted or licensed material
Direct regurgitation of long copyrighted sequences is somewhat uncommon but becomes noticeable for popular books. Regurgitation risk correlates with accuracy: text-davinci-002, davinci (175B) and Anthropic-LM v4-s3 (52B) show the most verbatim regurgitation.
Finding 18: disinformation
The largest models (text-davinci-002, Anthropic-LM v4-s3 (52B)) are effective at generating realistic headlines that support a thesis; results are more mixed for text encouraging certain actions (Table 8).
Finding 19: targeted biases (BBQ)
text-davinci-002 is the most accurate on BBQ by a wide margin (89.5%); next are T0++ (11B) at 48.4% and TNLG v2 (530B) at 44.9%, the only others above 40%. Only these three most-accurate models have biases in ambiguous contexts aligned with broader social biases; all other models show bias in the other direction. The most accurate models are thus those most concerning for social bias in ambiguous contexts; trends in disambiguated contexts are less clear.
Finding 20: targeted toxicity generation
Core-scenario toxicity generation rates were low. All models show much stronger tendencies to generate toxic text for toxic prompts in RealToxicityPrompts than for relatively non-toxic prompts in RealToxicityPrompts and BOLD. Dependence on the detector (Perspective API) and on human judgments from diverse stakeholders is flagged as a key area for future work.
Finding 21: comprehensiveness
By evaluating extensively under unified conditions, HELM exposes “findings lying in plain sight”: text-davinci-002 reaches 74.4% ROUGE-L on NarrativeQA, a new state of the art to the authors’ knowledge, above the QA-specialized UnifiedQA-v2 (67.4% ROUGE-L; Khashabi et al. 2022).
Finding 22: prompting
All models are significantly sensitive to prompt formatting, choice of in-context examples and number of in-context examples across all scenarios and metrics (Sec. 8.2). The paper standardizes these dimensions, but current models differ in which prompting decisions would maximize accuracy.
Finding 23: multiple-choice adaptation method
Performance is extremely sensitive to how multiple-choice scenarios are adapted: OPT (175B) on HellaSwag reaches 79.1% when each answer is presented in a separate 0-shot prompt (one of the most accurate models) but only 30.2% (almost random) when choices are presented jointly in one 5-shot prompt. The accuracy-maximizing adaptation differs across models even within one scenario (Fig. 33). This is “a fundamental challenge for what it means to standardize language model evaluation in a fair way across models.”
Finding 24: upstream perplexity and downstream accuracy
Across model families, bits-per-byte on The Pile is a poor predictor of downstream accuracy (Fig. 30), though some models are trained on The Pile (contamination or in-distribution confound). The authors encourage research on intrinsic or upstream surrogates that reliably predict downstream results, including desiderata beyond accuracy.
Finding 25: trends for model scale
Within a model family, scale reliably predicts accuracy; across all models, scale is not a good predictor of downstream accuracy for any scenario (Fig. 29). But there is a clear thresholding effect: all models whose head-to-head accuracy win rate is well above chance (greater than 55%) have at least 50B parameters. Of the 10 most accurate, some of the smallest (Anthropic-LM v4-s3 (52B), Cohere xlarge v20220609 (52.4B)) are in the top 5, so scaling may be inefficient compared with other means such as training with human feedback (compare TNLG v2 (530B) and Anthropic-LM v4-s3 (52B)).
Summary table
| # | Topic | Headline number |
|---|---|---|
| 14 | Language | irregular forms: top core models among the least accurate |
| 15 | Knowledge | TruthfulQA 62.0% vs 36.2%; NQ closed-book 38.5% vs 28.7%; WikiFact 34.3% vs 22.3% |
| 16 | Reasoning | GSM8K code-davinci-002 52.1%, text-davinci-002 35.0%; none other above 16% |
| 17 | Copyright | regurgitation correlates with accuracy |
| 18 | Disinformation | largest models produce realistic supporting headlines |
| 19 | BBQ | text-davinci-002 89.5%; T0++ 48.4%; TNLG v2 44.9% |
| 20 | Toxic generation | strongly prompt-dependent |
| 21 | Comprehensiveness | NarrativeQA 74.4% ROUGE-L new SOTA vs 67.4% |
| 22 | Prompting | sensitive to format, examples, number |
| 23 | MC adaptation | 79.1% vs 30.2% for OPT-175B on HellaSwag |
| 24 | Perplexity | BPB poor predictor of accuracy across families |
| 25 | Scale | threshold: win rate above 55% implies at least 50B |
Examples
A 49-point swing from adaptation alone
Finding 23: percentage points for one model on one dataset, changed only by the prompt protocol. If a leaderboard reported only the joint format, OPT (175B) would look near-random on HellaSwag; under the separate format it is near the top. Any claimed model ranking is therefore conditional on the adaptation protocol. The earlier cited swing for GPT-3-style few-shot prompts ranges from about 30% to 80% (Zhao et al. 2021).
sep, joint = 79.1, 30.2
print(sep - joint) # 48.9 percentage pointsConnections
- Findings 22-23 motivate the standardization element of holistic evaluation (Liang 2022 - Holistic Evaluation Framework and Taxonomy) and are quantified in Liang 2022 - Experiments Meta-Analysis and Prompting Ablations.
- Finding 25 relates to scaling-law results: Neural Scaling Laws and Compute-Optimal Training (Chinchilla).
- Findings 17 (memorization) and 24 (perplexity) bear on pretraining: Autoregressive Language Modeling and Pretraining.
Relevance to marketing measurement. Findings 22-24 are a warning against inferring method quality from a proxy or a single protocol: ranking MMM vendors by in-sample fit (an “upstream” proxy) can poorly predict out-of-sample or experiment-calibrated accuracy, and results depend on the tuning protocol. Report sensitivity to protocol alongside the ranking.