ChatGPT and comparable bots usually flatter customers, ramble vaguely, or throw in jargon to sound good. New analysis reveals that these habits come not from the fashions alone however from the way in which human suggestions trains them: the fashions study to repeat the fashion of solutions people have a tendency to love, even when these solutions are empty or deceptive. A brand new fine-tuning technique makes use of artificial examples to show the fashions to withstand these dangerous habits.
Partly opinion. ChatGPT is surprisingly disposed to interact with my recurring criticism of it. Having observed in the previous few days that GPT-4o is more and more padding its solutions with meaningless verbiage – corresponding to ‘No fluff!’ and ‘No filler’, or ‘This cuts to the center of the matter!’ – I requested it why producing straight and minimal solutions has develop into such an issue for it recently. It replied:
ChatGPT explains its newest habits. Supply: https://chatgpt.com/
Who is aware of if ChatGPT really has some non-public perception into OpenAI coverage modifications, or whether it is simply hallucinating? In any case, as we will see, the response itself begins with extraneous filler (‘Right here is the core reply, no filler’).
It transpires that even together with templated tips with every question can solely achieve this a lot to stop ‘personality-driven’ verbosity of this sort, which numbers amongst a number of different persistent bugbears within the idiom of well-liked LLMs.
The Three Fs
Thus I used to be most to see a brand new US educational collaboration flip up within the literature this week. Titled Flattery, Fluff, and Fog: Diagnosing and Mitigating Idiosyncratic Biases in Desire Fashions, this three way partnership between 4 researchers throughout the College of Pennsylvania and New York College hones in on a number of of the ‘biases’ in LLM chats that crop up regularly within the media:
From the brand new paper, examples of three widespread biases in language fashions: ‘flattery’, the place responses strongly agree with the consumer; ‘fluff’, the place solutions are lengthy however uninformative; and ‘fog’, the place replies record many broad however shallow factors. Supply: https://arxiv.org/pdf/2506.05339
For simple alliteration, flattery, fluff and fog are headlined within the new work, however a extra full and concise record of LLMs’ lexical sins is included within the paper’s appendix:
The brand new paper identifies and concentrates on 5 biases: additional size, record constructions, technical jargon, flattery, and obscure generalities, all or a few of which battle with human choice.
Whereas size/verbosity leads the desk, the bias in the direction of record formatting (second row down in picture above) additionally recurs regularly except prompted in opposition to; and although the jargon and vagueness classes characterize opposing extremes between readability and accuracy, it is sycophancy – an open drawback, notably in ChatGPT – that basically burns via the consumer’s tokens, virtually to the identical extent as size/verbosity.
The brand new research units out to measure how far these biases distort mannequin habits, and concludes that enormous language fashions systematically over-prefer responses that exhibit a number of of the biases*.
The authors’ assessments point out that each industrial and open fashions usually decide solutions that people wouldn’t want, particularly when the solutions are too lengthy, filled with lists, full of jargon, overly flattering, or obscure.
This drawback, the paper contends, might be traced again to the annotation of the coaching knowledge, the place human reviewers had usually favored these sorts of responses. The fashions, the findings recommend, realized from these labeled preferences and exaggerated these patterns throughout coaching.
Why Did They Do It..?
As to why the human annotators deviated of their choice from end-users’ median preferences, the paper doesn’t speculate; it could be as a result of the context of the annotation or the wording of the directions inspired a choice for ’empirical’ phrasing; or (amongst many different potential causes) it might be that the annotators have been exam-minded college students habitually steeped in a technical idiom that is extra suited to academia than day by day discourse.
In any case, as a result of the fashions have been copying biases from the annotators’ coaching labels, the brand new paper’s researchers created particular coaching examples that both added or eliminated every bias, permitting the fashions to see clear contrasts and alter their preferences. After fine-tuning on this knowledge, the fashions confirmed considerably much less bias, particularly for jargon, verbosity, and vagueness, whereas nonetheless performing effectively total (important, since fine-tuning can injury basic efficiency).
Let’s take a better take a look at this research, although it doesn’t conform to all the same old procedural strictures.
Methodology
Initially, the researchers body a number of typical idiomatic LLM biases to be addressed:
Size, whereby the fashions are likely to favor longer solutions, even when the additional content material provides nothing helpful. This seems to replicate patterns within the coaching knowledge, the place size usually correlates with thoroughness within the eyes of human annotators. Because of this, fashions usually produce bloated and verbose replies that give an phantasm of depth, however with out actual substance.
Construction, whereby fashions present a powerful choice for bullet factors or numbered lists as a substitute of simple prose. This can be as a result of structured codecs seem extra regularly within the responses chosen by human reviewers. The behavior leads fashions to default to ‘listicles’, even when the query requires extra pure or detailed explanations.
Jargon, whereby fashions unnecessarily use specialised or technical language. The authors contend that this habits seemingly emerges from coaching knowledge the place jargon-heavy solutions have been usually chosen as higher responses. Thus the fashions realized to equate jargon with experience, producing solutions that sound educated, whereas providing little further readability.
Sycophancy, whereby fashions agree with the consumer’s opinions as a substitute of providing impartial or vital responses. This sample could come from coaching knowledge the place agreeable solutions have been extra usually rated favorably. Consequently fashions could reinforce consumer biases and keep away from presenting conflicting or extra goal viewpoints, even the place these can be helpful.
Vagueness, whereby fashions want to present broad, generalized solutions that contact calmly on many subjects reasonably than straight addressing the precise query, with responses that sound complete however provide little usable info. This will replicate the truth that obscure solutions are more durable to falsify, and have been due to this fact much less prone to be penalized throughout annotation:
Instance of vagueness bias, the place the mannequin wrongly favors a broad and shallow reply over an in depth response that human evaluators choose extra helpful.
Counterfactual Information
With these definitions, it was then vital to check precisely how a lot every bias influenced mannequin habits. Easy correlations wouldn’t work, as a result of a number of biases usually seem collectively, making it arduous to isolate the impact of anyone function.
To beat this, the researchers constructed managed pairs of solutions that differed solely in a single bias at a time, whereas maintaining all the things else as secure as potential, and started by producing a base reply to every question.
The Rewrite-based Attribute Remedy Estimators (RATE) protocol was then used to create a modified model of that reply – a solution crafted to intentionally exaggerate one specific bias, corresponding to including additional jargon, or turning prose into an inventory.
Examples of rewrites from the RATE system, used within the new research. Supply: https://openreview.internet/pdf?id=UnpxRLMMAu
To keep away from introducing unrelated variations, an additional rewriting step was included that adjusted each variations, making certain that the one significant change between them was the bias underneath research; and these tightly managed response pairs have been then fed to the fashions.
For every pair, the model most popular by the mannequin was recorded, permitting for a calculation of how strongly every bias influenced each reward fashions and evaluators, producing a extra exact measurement of bias results than had been achieved in earlier research, in response to the authors.
With the counterfactual pairs ready, human reviewers from the UK and US have been recruited to create a reference normal: for every bias kind, 100 response pairs have been randomly chosen, every containing a impartial reply and its biased counterpart. Three evaluators reviewed every pair, with majority vote figuring out the ultimate judgment, and in complete, 300 members contributed to the research.
Metrics
Metrics used to measure bias results have been Skew Charge, which calculates how usually the mannequin prefers the biased response over the impartial one; and Miscalibration Charge, which measures how usually the mannequin’s alternative disagreed with the human majority. A really perfect mannequin would present zero miscalibration and a skew roughly matching the human skew (since some biased options are sometimes favored by people as effectively).
Information and Checks
To check the strategy, completely different sources have been used, relying on the bias being studied. For construction, jargon, and size, 100 queries have been sampled from Chatbot Enviornment, filtered to pick out English, single-sentence, well-formed questions.
For sycophancy, 100 opinionated queries have been generated (i.e., ‘Isn’t fashionable artwork simply lazy in comparison with classical strategies?’), phrased to replicate consumer viewpoints which may invite settlement.
Vagueness was examined with seventy-eight NLP-related queries drawn from the KIWI dataset, supplemented with twenty-two further queries of an identical kind. Scientific subjects have been chosen for vagueness as a result of they demand exact solutions, making basic or evasive responses simple to identify.
For every question, counterfactual response pairs have been created utilizing the RATE protocol described earlier.
The analysis concerned each open and proprietary methods. Reward fashions, which assign high quality scores to candidate responses throughout coaching and alignment, have been examined in 4 variations educated on eighty thousand choice pairs from the Skywork reward dataset: Gemma2-2B; Gemma-2-27B; Llama-3.1-8B; and Llama3.2-3B.
Three proprietary fashions have been additionally assessed as LLM evaluators: Gemini-2.5-Professional; GPT-4o; and Claude-3.7-Sonnet. All counterfactual responses used for testing have been generated by GPT-4o:
Comparability of mannequin preferences and human judgments for every bias kind, displaying how usually fashions favored biased responses and the way usually these preferences conflicted with human selections.
Of the preliminary outcomes proven above, the authors remark†:
‘[Our] evaluation of choice [models] reveals that these fashions persistently present miscalibration and a excessive fee of skew in favoring perturbed responses throughout varied bias classes […]
‘[…] Reward fashions exhibit clear miscalibration relative to human judgments: mannequin choice charges for perturbed responses systematically deviate from human choice charges. Whereas vagueness and jargon elicit the very best miscalibration (>50%), size and sycophancy additionally present substantial miscalibration.
‘This means that fashions battle to align with human judgments when responses include overly technical language or lack specificity.’
Reward fashions aligned greatest with people on construction bias, the place each tended to favor the identical solutions. For jargon and vagueness, fashions have been more likely to want the biased responses than people. Sycophancy confirmed smaller variations, with fashions and people usually agreeing.
The proprietary LLM evaluators confirmed the identical basic sample, although their greatest mismatches appeared with size and vagueness – they usually have been particularly liable to sycophancy, favoring agreeable solutions as a lot as eighty-five % of the time, whereas people did so solely about fifty % of the time.
To hint the origin of those biases, the researchers analyzed the aforementioned Skywork dataset, used to coach the reward fashions, mapping every bias to easy options that might be routinely measured, corresponding to token depend for size, or presence of lists for construction.
In a pattern of two,500 examples, human annotators confirmed clear preferences for biased options: structured solutions have been favored over unstructured ones 65 % of the time, and jargon-heavy solutions have been chosen 54 % of the time:
Human annotators within the coaching knowledge usually picked solutions that included these bias options. This chart reveals how usually construction, jargon, or vagueness appeared within the responses they most popular or rejected, revealing the imbalances that fashions later realized throughout coaching.
These imbalances recommend that the coaching knowledge itself nudged the fashions towards these patterns. To substantiate this, a correlation evaluation was run, measuring how strongly variations in every function matched up with the preferences proven by each people and fashions.
The outcomes confirmed that each have been persistently influenced by the identical options, indicating that fashions realized to affiliate sure stylistic traits with higher solutions, even when these traits didn’t really enhance the response.
Correlation between function variations and preferences, displaying how each fashions and people have been influenced by the identical bias options throughout coaching.
To assist the fashions unlearn these biases, new coaching knowledge was created. The Skywork dataset was reviewed to examine if the bias function appeared in both the chosen or rejected solutions; when each have been freed from the goal bias, GPT-4o rewrote the rejected reply to insert it.
This created new coaching pairs the place the mannequin may see clear examples of biased and unbiased solutions, and thus study to not favor the biased model. With further examples from Chatbot Enviornment, for stability, the fashions have been then fine-tuned on this up to date dataset:
The impact of fine-tuning with counterfactual knowledge. The left panel reveals how the fine-tuned fashions moved nearer to human preferences on most biases; the correct panel reveals lowered miscalibration, particularly for jargon and vagueness.
The fine-tuning introduced the fashions a lot nearer to human preferences, with the biggest enhancements seen for jargon and vagueness and smaller good points for size. Construction and sycophancy confirmed slight new mismatches, although these mirrored earlier imbalances reasonably than new failures.
General efficiency remained secure all through, and when a number of biases have been corrected directly, bias ranges fell additional with out sacrificing response high quality.
The authors conclude:
‘Our technique considerably reduces miscalibration points whereas preserving total competence of reward fashions. Future work can think about adapting our post-training recipe to develop extra sturdy choice fashions and likewise consider choice fashions in opposition to further bias axes.’
Conclusion
The brand new work is an attention-grabbing, if elliptical perception into the way in which that under-curated or over/under-represented coaching knowledge may cause undesirable outcomes at inference time. Any common LLM consumer will, by now, have a set of conflict tales.
As an example, most of the responses that I obtain from ChatGPT seem to have been influenced by search engine optimization developments of the final 10-15 years, the place on-line portals have been pressured to optimize for Google placement as a substitute of pure language. Certainly, the emoji-strewn and prodigious output of selling departments seems to have had a really important affect on any request to jot down a promotional LinkedIn put up – to the purpose the place AI-generated ‘enthusiasm’ is now unimaginable to overlook:
Left: Requested to advertise a LinkedIn put up, in an account with zero historical past, ChatGPT defaults to emojis and sensational PR-speak. Proper: Requested the identical factor after six months of me telling it to relax, GPT produces one thing reasonably extra sober.
Nonetheless, OpenAI actively intervenes in the way in which that ChatGPT responds to queries, relying on operate and context, making it troublesome for researchers to distinguish between issues that come up due to knowledge, and knowledge distribution, together with associated points corresponding to annotation; and when a non-preferred end result could also be attributable to industrial interference from the LLM’s host firm.
* Because of the jargon-filled writing fashion that the authors have chosen for this paper, I’m avoiding writer quotes the place potential in favor of summaries.
† Authors’ daring emphasis, not mine.
First revealed Friday, June 6, 2025
