GPT-4-1106-preview: OpenAI Turbo Variant

• GPT-4-1106-preview is a GPT-4 family variant known as a 'turbo' model used in API experiments and alignment benchmarks.
• It exhibits strong in-context learning with high benchmark scores yet shows sensitivity to prompt framing and contextual shifts.
• The model serves as a reference for studies on decision biases, OCR-based document understanding, and response-length calibration.

GPT-4-1106-preview is a later GPT-4 family model released by OpenAI and exposed in API-facing contexts as a “preview” or “turbo” variant. In the available literature, it is treated both as a concrete experimental object—especially in studies of decision behavior and document understanding—and as an operational alias for “GPT-4 Preview” in alignment benchmarks. The GPT-4 Technical Report does not identify a model with this exact name, but it documents the underlying GPT-4 architecture, training objective, post-training alignment process, capability profile, safety interventions, and limitations that are presented as the relevant substrate for GPT-4-1106-preview (OpenAI et al., 2023).

1. Identity, naming, and model family position

The designation “GPT-4-1106-preview” refers, in the cited literature, to a GPT-4 family model released after the original GPT-4 report. One source describes it as “gpt-4-1106-preview (GPT-4 Turbo)” accessed through the OpenAI Python API in January 2024 experiments on context-dependent choice (Hayes et al., 2024). Another maps “1106-preview” specifically to the text-only turbo variant in document-understanding evaluations, listing it as “TURBO + OCR (version 1106-preview, text-only)” (Borchmann, 2024). A separate alignment paper states that, in AlpacaEval 2.0, MT-Bench, and Arena-Hard, “GPT-4 Preview” refers to the OpenAI model alias gpt-4-1106-preview (Liu et al., 2024).

The relation between GPT-4-1106-preview and earlier GPT-4 releases is indirect but explicit. In abstraction-and-reasoning experiments, the evaluated text-only model was gpt-4-0613, not gpt-4-1106-preview; however, both are described as GPT-4 family models, and the paper notes that the earlier model is not stated to be strictly identical to GPT-4-1106-preview (Mitchell et al., 2023). This suggests continuity at the family level rather than identity at the checkpoint level. The same source adds a caveat that performance could shift by a few percentage points and that default sampling parameters may differ across releases (Mitchell et al., 2023).

At the architectural level, GPT-4 is described as a “Transformer-based model pre-trained to predict the next token in a document,” with multimodal extension to accept image and text inputs and produce text outputs (OpenAI et al., 2023). The report does not disclose parameter count, layer widths, hardware specifics, dataset composition, or total training compute (OpenAI et al., 2023). For GPT-4-1106-preview specifically, publicly reported work therefore characterizes behavior primarily through benchmarked interfaces rather than through transparent model-card disclosure.

2. Core architecture, training objective, and alignment lineage

The most direct technical account of the underlying model class comes from the GPT-4 Technical Report. GPT-4 is trained with standard cross-entropy next-token prediction, with loss written as

$$L(\theta) = - \sum_t \log p_\theta(x_t \mid x_{<t})$$

and is subsequently aligned through Reinforcement Learning from Human Feedback (RLHF) to improve factuality and adherence to desired behavior (OpenAI et al., 2023). The report states that human labelers provide demonstrations and rankings, a reward model is trained, and a policy is fine-tuned via Proximal Policy Optimization. It further describes rule-based reward models (RBRMs), implemented as zero-shot GPT-4 classifiers over hand-written rubrics, as part of safety-oriented reward shaping (OpenAI et al., 2023).

The report also emphasizes predictable scaling. Final loss on an internal code corpus is modeled as

$L(C) = a C^b + c$

where $C$ is normalized compute, and HumanEval pass-rate scaling is reported to follow an approximate power-law relation fitted from models trained with much less compute (OpenAI et al., 2023). These claims are attached to GPT-4 as a development program rather than to GPT-4-1106-preview as a separately disclosed architecture, but the report is explicit that preview instances of GPT-4 inherit the same core model class and post-training lineage (OpenAI et al., 2023).

Capability claims in the report establish the broader performance envelope within which GPT-4-1106-preview is usually situated. GPT-4 is reported to achieve 86.4% on MMLU, 95.3% on HellaSwag, 96.3% on ARC, 87.5% on WinoGrande, 67.0% pass rate on HumanEval, and 92.0% on GSM-8K under chain-of-thought prompting (OpenAI et al., 2023). It is also reported to obtain a simulated Uniform Bar Exam score of 298/400, around the 90th percentile (OpenAI et al., 2023). At the same time, the report states that GPT-4 remains less capable than humans in many real-world scenarios, can hallucinate, can make reasoning errors, can be overconfident in wrong answers, has no online learning, and has knowledge that is mostly frozen pre-September 2021 (OpenAI et al., 2023).

These details matter for GPT-4-1106-preview because later empirical papers repeatedly test whether strong benchmark performance transfers to structurally different settings. The resulting picture is uneven: strong in-context adaptation and benchmark utility coexist with failures in robust abstraction, context-independent valuation, and layout-sensitive document reasoning (Mitchell et al., 2023, Hayes et al., 2024, Borchmann, 2024).

3. Abstraction and reasoning performance on ConceptARC

The most direct evidence on abstraction comes from “Comparing Humans, GPT-4, and GPT-4V On Abstraction and Reasoning Tasks,” which evaluates gpt-4-0613 and gpt4-vision-preview on ConceptARC and then discusses correspondence to GPT-4-1106-preview (Mitchell et al., 2023). ConceptARC is described as a benchmark of robust understanding and abstract reasoning over 16 psychology-derived core concepts, with 480 tasks total organized into 16 concept-groups of 30 tasks each. Each task presents a few demonstrations of grid-based transformations with colors encoded as integers 0–9 and requires the solver to infer a general rule and produce the transformed test grid (Mitchell et al., 2023).

For text-only GPT-4 on all 480 tasks, the paper reports prior zero-shot accuracy of $p=0.19$ at $T=0.0$ and $p=0.25$ at $T=0.5$, while its one-shot prompting protocol raises performance to $p=0.33$ at both temperatures, approximately $k \approx 158/480$ (Mitchell et al., 2023). Humans from Moskvichev et al. are reported at $p=0.91$ with $L(C) = a C^b + c$0 (Mitchell et al., 2023). The paper defines accuracy as $L(C) = a C^b + c$1 and standard error as

$L(C) = a C^b + c$2

with the corresponding estimates reported as approximately $L(C) = a C^b + c$3 for humans and $L(C) = a C^b + c$4 for one-shot GPT-4 (Mitchell et al., 2023).

The one-shot protocol is unusually detailed. It includes a system-style header beginning “You are given demonstration grids…”, an explicit enumeration of the 10 color codes, and a fully worked example containing demonstrations, a natural-language rule, the correct output grid, and a brief explanation (Mitchell et al., 2023). Up to three guesses are allowed per task by reissuing the prompt with “Please try again with a different answer.” A task is counted as solved if any of the three guesses exactly matches the ground-truth grid (Mitchell et al., 2023). For GPT-4V, only the 48 “minimal” tasks were tested because of cost; success there was based on accurate natural-language description of the output grid’s shape and colors rather than exact grid reproduction (Mitchell et al., 2023).

The quantitative pattern is consistent across modalities. On the 48 minimal tasks, humans achieve $L(C) = a C^b + c$5 ($L(C) = a C^b + c$6), text-only GPT-4 reaches $L(C) = a C^b + c$7 at $L(C) = a C^b + c$8 and $L(C) = a C^b + c$9 at $C$0, whereas GPT-4V obtains $C$1 zero-shot and $C$2 one-shot (Mitchell et al., 2023). The paper’s qualitative analysis attributes failures to miscounting or mislocating objects, ignoring required invariances, brittle rule induction on compositional tasks, and over-reliance on details from the one-shot example. It also notes prompt brittleness: small rephrasings or changing the example can flip a solution from correct to incorrect (Mitchell et al., 2023).

The paper’s stated implication for GPT-4-1106-preview is cautious but clear. Because gpt-4-0613 and GPT-4-1106-preview are both GPT-4 family models, the qualitative conclusions are said likely to hold: a large gap versus humans, gains from one-shot prompting, and multimodal underperformance relative to text-only GPT-4 on these tasks (Mitchell et al., 2023). This suggests that high benchmark scores and robust abstraction should not be conflated.

4. Relative value bias and context-dependent decision behavior

GPT-4-1106-preview is directly evaluated in “Relative Value Biases in LLMs,” which adapts a two-phase bandit task from behavioral research on humans (Hayes et al., 2024). In each of 30 simulation runs, eight slot machines are randomly assigned to letters A–H and organized into four learning contexts—Low, Medium-Low, Medium-High, and High—each pairing a “low” option $C$3 with a “high” option $C$4 for $C$5 (Hayes et al., 2024).

The protocol separates a 20-trial learning phase from a 28-trial transfer test. During learning, the full history of previous outcomes from both options in the same context is embedded in the prompt, and the model is instructed to choose the slot machine that maximizes cumulative payoff (Hayes et al., 2024). During transfer, all 28 possible pairs of the eight machines are presented once without feedback, and the entire 20-trial learning history is carried forward into each prompt (Hayes et al., 2024). A rational absolute-value chooser should compute expected values,

$C$6

whereas a relative-valuation chooser should prefer machines that were locally superior in their original learning contexts (Hayes et al., 2024).

All prompts use the system role “You are a helpful assistant,” with default OpenAI API settings: temperature $C$7, top_p $C$8, other parameters at default, and $C$9 response per trial (Hayes et al., 2024). In the learning phase, GPT-4-Turbo selects the locally optimal option with probability $p=0.19$0 by Trial 5, reported as mean $p=0.19$1, $p=0.19$2, $p=0.19$3, Cohen’s $p=0.19$4 (Hayes et al., 2024). This is presented as robust in-context learning.

The transfer phase reveals a different property. The authors quantify context dependence with a linear contrast

$p=0.19$5

using $p=0.19$6 across the eight machines (Hayes et al., 2024). Under pure absolute valuation, $p=0.19$7; under pure relative valuation, $p=0.19$8 (Hayes et al., 2024). For GPT-4-Turbo in the Baseline condition, the reported value is $p=0.19$9 with SD $T=0.0$0, $T=0.0$1, $T=0.0$2, and $T=0.0$3, which the authors interpret as a strong relative-value bias (Hayes et al., 2024).

Prompt manipulations sharply affect this bias. In the “Regret” condition, each historical outcome line is rewritten to include the relative difference $T=0.0$4, and the contrast shifts to $T=0.0$5; the paper states that this significantly exceeds the Baseline bias, with $T=0.0$6, $T=0.0$7 (Hayes et al., 2024). In the “Expected Outcomes” condition, the prompt asks the model to estimate the expected amount of money it would win from each slot machine before choosing, and the contrast reverses to $T=0.0$8 with $T=0.0$9 (Hayes et al., 2024). Broken-context formatting likewise yields $p=0.25$0 (Hayes et al., 2024).

The paper’s central conclusion is that GPT-4-1106-preview exhibits relative value decision biases similar to those observed in humans and animals, but that the bias can disappear when the prompt explicitly requires expected-outcome estimation (Hayes et al., 2024). A plausible implication is that GPT-4-1106-preview’s decision behavior is highly sensitive to representational framing even when complete historical information is present in-context.

5. Document understanding, OCR dependence, and length sensitivity

“Notes on Applicability of GPT-4 to Document Understanding” provides the most detailed benchmark-specific characterization of GPT-4-1106-preview as a text-only model in document tasks (Borchmann, 2024). In that paper, “1106-preview” denotes the GPT-4-1106-preview turbo variant used as “TURBO + OCR,” with plain-text inputs only, an 8K token window by default, and no built-in vision encoder (Borchmann, 2024). OCR text is concatenated with the question in a single prompt; no image pixels are supplied (Borchmann, 2024).

The evaluation spans four benchmarks: DocVQA, Infographics VQA, Slide VQA, and DUDE (Borchmann, 2024). OCR engines tested are Tesseract 5.3.3, Azure Cognitive Services OCR 3.2 (2022-04-30), and Amazon Textract DetectDocumentText 1.0 (Borchmann, 2024). The best OCR choice for GPT-4-1106-preview is reported as Amazon Textract on DocVQA and SlideVQA, and Azure OCR on Infographics VQA and DUDE (Borchmann, 2024). Prompts are zero-shot and greedy, with temperature $p=0.25$1, top_p $p=0.25$2, and a short answer format such as “Answer the question. Do not write a full sentence, just provide a value. If the value is unclear, guess it given the input document” (Borchmann, 2024).

The reported final test-set scores for GPT-4-1106-preview are as follows:

Benchmark	Metric	Score
DocVQA	ANLS	78.2
Infographics VQA	ANLS	54.3
Slide VQA	Exact Match	45.9%
DUDE	ANLS	48.1

These results are explicitly contrasted with vision-enabled variants. The paper reports that “TURBO v + OCR” exceeds “TURBO + OCR” by $p=0.25$3 on DocVQA, $p=0.25$4 on Infographics VQA, $p=0.25$5 on Slide VQA, and $p=0.25$6 on DUDE (Borchmann, 2024). It also notes that Vision Turbo without OCR performs worse than Vision Turbo with OCR, implying that OCR text remains materially useful even for vision-capable variants (Borchmann, 2024).

A prominent failure mode is sensitivity to document length and position. On Slide VQA, performance by page-location bucket declines from $p=0.25$7 on pages 1–5 to $p=0.25$8 on pages 6–10, $p=0.25$9 on pages 11–15, and $T=0.5$0 on pages 16–20 (Borchmann, 2024). The paper characterizes this as a primacy bias and states that chunking into smaller page windows or hierarchical retrieval was tested as mitigation, though no end-to-end fix was reported (Borchmann, 2024).

The error analysis is concrete. Typical mistakes include table-cell alignment errors, omitted list items in extractive questions, hallucinatory answers for figures, and output-format deviations such as “The answer is 42.” rather than “42” (Borchmann, 2024). A contamination analysis using guided instruction reports DocVQA baseline $T=0.5$1 and Infographics VQA baseline $T=0.5$2 when the correct dataset name is mentioned, with no comparable gain for misguided instructions; the paper states that gains of up to 3.8 points on some variants imply partial contamination (Borchmann, 2024).

Taken together, these results place GPT-4-1106-preview in an intermediate position for document understanding: effective enough for some OCR-mediated text extraction and question answering, but substantially weaker than vision-plus-OCR systems on tasks requiring layout, spatial arrangement, or visual evidence (Borchmann, 2024).

6. GPT-4-1106-preview as evaluation baseline and alignment target

GPT-4-1106-preview also functions as a benchmark anchor in alignment research. “Iterative Length-Regularized Direct Preference Optimization: A Case Study on Improving 7B LLMs to GPT-4 Level” identifies “GPT-4 Preview” in AlpacaEval 2.0, MT-Bench, and Arena-Hard explicitly as the alias gpt-4-1106-preview (Liu et al., 2024). In this role, the model is both a reference answer generator and, in MT-Bench, the judge model following the cited evaluation protocol (Liu et al., 2024).

The paper’s central contribution is iterative length-regularized DPO (iLR-DPO), but its significance for GPT-4-1106-preview lies in how the latter is operationalized as a quality frontier. On AlpacaEval 2.0, with length control defined by truncating or padding outputs so that average response length matches GPT-4 Preview’s approximately 2049 tokens, the reported 7B model reaches a 50.5% length-controlled win rate against GPT-4 Preview, compared with GPT-4 Preview’s 50.0%, and produces average length 2045 versus 2049 tokens (Liu et al., 2024). On MT-Bench, the base model is 6.59, iLR-DPO iteration 3 reaches 7.02, and GPT-4 Preview is listed at 6.59 (Liu et al., 2024). On Arena-Hard, iLR-DPO attains a 20.7% length-controlled win rate (Liu et al., 2024).

The paper also records a property of GPT-4-1106-preview that matters methodologically: its output length is used as the normalization target. Vanilla iterative DPO increases verbosity sharply, with average response length around 5600 tokens by iteration 3; iLR-DPO instead keeps length around 2050 while still improving length-controlled win rate (Liu et al., 2024). This makes GPT-4-1106-preview a practical reference not only for judged answer quality but also for a response-length regime. A plausible implication is that the model’s default answer length had already become a de facto evaluation standard in instruction-following benchmarks by mid-2024.

In this evaluative role, GPT-4-1106-preview is not presented as a transparent model family endpoint. Rather, it serves as a high-performing, externally hosted baseline against which open models are tuned, compared, and sometimes judged (Liu et al., 2024). That institutional role is distinct from the model’s cognitive profile as measured in reasoning or decision tasks.

7. Limitations, prompt dependence, and practical interpretation

Across the cited literature, GPT-4-1106-preview is associated with a recurring pattern: strong performance under well-structured benchmark conditions, coupled with prompt-sensitive failures in tasks requiring robust abstraction, context independence, or layout-grounded reasoning. The GPT-4 Technical Report already states that GPT-4 can hallucinate, make reasoning errors, be overconfident when wrong, and exhibit degraded calibration after alignment fine-tuning (OpenAI et al., 2023). Later task-specific studies sharpen those general warnings.

In ConceptARC-like reasoning tasks, the recommended guidance is to provide a clear worked example, avoid relying on zero-shot prompting alone, and verify outputs carefully; the paper also notes that multimodality by itself did not solve abstract rule induction (Mitchell et al., 2023). In decision settings, prompts that emphasize pairwise differences or “regret” magnify relative-value bias, whereas prompts that require explicit expected-value estimation can eliminate it (Hayes et al., 2024). In document understanding, high-quality OCR, answer-format constraints, and chunking for longer documents are presented as practical mitigations, while the paper recommends Vision Turbo plus OCR for layout-sensitive tasks (Borchmann, 2024).

A common misconception is that strong exam or leaderboard performance implies uniformly human-like reasoning. The available evidence does not support that interpretation. GPT-4 is reported to achieve top-percentile scores on many professional and academic benchmarks (OpenAI et al., 2023), yet ConceptARC results remain far below human performance even under richer one-shot prompting (Mitchell et al., 2023). Similarly, in the bandit study, GPT-4-1106-preview learns local contingencies rapidly but nonetheless displays a context-dependent bias that disappears only under explicit expected-outcome prompting (Hayes et al., 2024). In document tasks, zero-shot OCR-only performance is workable but materially weaker than multimodal alternatives on layout-sensitive benchmarks (Borchmann, 2024).

The combined record therefore characterizes GPT-4-1106-preview less as a single stable capability point than as a model whose observed competence depends strongly on task structure, modality, and prompt design. This suggests that any technical assessment of the model should distinguish between benchmark-optimized performance, in-context decision procedures, and robustness under distributional or representational shifts.