NeedleBench: Can LLMs Do Retrieval and Reasoning in 1 Million Context Window?

Mo Li^1,2, Songyang Zhang^1‡, Yunxin Liu², Kai Chen^1‡
¹Shanghai AI Laboratory ²Tsinghua University
{limo,zhangsongyang}@pjlab.org.cn

Abstract

In evaluating the long-context capabilities of large language models (LLMs), identifying content relevant to a user’s query from original long documents is a crucial prerequisite for any LLM to answer questions based on long text. We present NeedleBench, a framework consisting of a series of progressively more challenging tasks for assessing bilingual long-context capabilities, spanning multiple length intervals (4k, 8k, 32k, 128k, 200k, 1000k, and beyond) and different depth ranges, allowing the strategic insertion of critical data points in different text depth zones to rigorously test the retrieval and reasoning capabilities of models in diverse contexts. We use the NeedleBench framework to assess how well the leading open-source models can identify key information relevant to the question and apply that information to reasoning in bilingual long texts. Furthermore, we propose the Ancestral Trace Challenge (ATC) to mimic the complexity of logical reasoning challenges that are likely to be present in real-world long-context tasks, providing a simple method for evaluating LLMs in dealing with complex long-context situations. Our results suggest that current LLMs have significant room for improvement in practical long-context applications, as they struggle with the complexity of logical reasoning challenges that are likely to be present in real-world long-context tasks. All codes and resources are available at OpenCompass: https://github.com/open-compass/opencompass.

{NoHyper}^†^†footnotetext: ^‡ Corresponding author.

1 Introduction

The capability of LLMs to process long texts is particularly crucial across various situations. LLMs can rapidly identify and summarize relevant information within lengthy documents, making them invaluable for legal document retrieval, academic research, and aggregating business intelligence, among other applications.

Modern LLMs have recently been developed to support longer context windows. For example, GPT-4 Turbo(OpenAI, 2023) offers long-context capabilities up to 128K tokens. Similarly, Claude 2.1(Anthropic, 2024a) has been enhanced to manage contexts up to 200K tokens, with the Claude 3 series(Anthropic, 2024b) being specifically engineered to digest inputs surpassing 1 million tokens. Furthermore, Gemini 1.5 supports context windows of millions of tokens(Gemini Team, 2024). Moreover, the recent open-source models, GLM4-9B-Chat and InternLM2.5-7B-Chat also support the 1 million context window. As models accommodate longer text lengths, verifying their comprehension of details within the text becomes increasingly essential.

Considering the critical role of LLMs in handling long texts, numerous approaches have been suggested to evaluate their long-context capabilities. Existing datasets, such as the LongBench dataset(Bai et al., 2023), provide a bilingual (Chinese and English) benchmark for long text comprehension, featuring tasks with lengths generally ranging from 5k to 15k tokens. However, accurately assessing the performance of long-context LLMs, especially at the 1M token level, continues to be a significant challenge.

There are some initial attempts in this direction, Mohtashami & Jaggi (2023) introduces the passkey testing approach, which embeds key information passkeys in repetitively similar texts and queries for this information at the end, to evaluate the basic information extraction capability of the LLaMA-7B model at text lengths of 32K. InfiniteBench by Zhang et al. (2023) extends the passkey method to lengths beyond 100K, inserting passkeys at various depths within long texts. Kamradt (2023) develops the Needle In A Haystack(NIAH) test, which uses a more diverse set of non-repetitive personal essays as filler information unrelated to the passkeys. This test further extends the context window to 200K and performs stress tests on the Claude2.1 and GPT-4 Turbo¹¹1GPT-4 1106-preview version models.The results from Kamradt (2023) and Anthropic (2024b) indicate that leading models are generally capable of passing the NIAH test with excellent performance.

Refer to caption — Figure 1: Our ATC test, even with a modest context length of 2K, reveals surprising limitations of these leading models in real-world complex long-context questions.

Does passing the ”needle-in-a-haystack” test—extracting key info from lengthy texts—really indicate that LLMs can handle complex real-world long context problems? Typically, real-world tasks require models to retrieve and integrate multiple pieces of dispersed, question-related information rather than just a single piece. For instance, in legal case analysis, a model must extract various relevant facts and legal provisions from a case file and synthesize this information to answer specific legal questions. Similarly, in business intelligence analysis, a model may need to aggregate information on market trends, competitor strategies, and consumer behaviors from multiple reports to provide a comprehensive market analysis. This requires models that can identify multiple key information points and integrate a wide range of relevant content within the text, offering in-depth and accurate analyses.

Therefore, improving the long-context ability of LLMs for realistic applications not only requires the accurate information retrieval but also the strong reasoning capabilities. To address the limitations of existing long-context information extraction evaluation methods and their misalignment with real-world application scenarios, we introduce the NeedleBench dataset. NeedleBench comprises a series of advanced long-context information capability evaluation methods, aimed at providing a comprehensive and targeted assessment of models’ abilities to extract and analyze information within the context of long texts. Furthermore, we develop the Ancestral Trace Challenge (ATC) test as the simplified proxy for measuring multi-step logical reasoning. Our findings demonstrate that current LLMs struggle to handle reasoning tasks with complex logical relationships, even with texts shorter than 2K tokens (see Figure 1). Our major contributions are as follows:

•

We introduce NeedleBench, a customizable dataset framework that includes tasks for evaluating the bilingual long-context capabilities of LLMs across multiple length intervals (4k, 8k, 32k, 128k, 200k, 1000k and beyond) and various text depth ranges, allowing the strategic insertion of critical data points in different text depth zones to rigorously test both retrieval and reasoning capabilities of models in a diverse context.
•

We propose the ATC to simulate complex long-context tasks in real-world scenarios, providing a simple method for evaluating LLMs in complicated long-context situations. Through the experimental results of the ATC, we discover that all current LLMs have significant room for improvement in practical long-context applications, struggling with the complexity of real-world long-context tasks.
•

We conduct a fine-grained evaluation and analysis of the performance of mainstream models in identifying key question-relevant information and reasoning. Additionally, all reproducible scripts, code, and datasets are provided in OpenCompass Contributors (2023).

2 Tasks and Datasets

In order to construct a dataset capable of comprehensively evaluating models’ capabilities in long-context information extraction and reasoning, we design a series of progressively more challenging test schemes. We show the composition of the Needlebench framework in Figure 2. Specifically, we divide the overall task into three subtasks: ”Single Retrieval Task(S-RT)” (where a single piece of information is inserted at one depth), ”Multi-Retrieval Task (M-RT)”, and ”Multi-Reasoning Task(M-RS)”. Each subtask includes the design of the key information (needle design) and the unrelated text (haystack design). In addition, we design the Ancestral Trace Challenge (ATC) to test the ability of the LLMs to handle multi-step logical challenges that are likely present in real world long-context scenarios.The instantiation of each task is detailed in Appendix C, where actual examples of each task are provided.

2.1 NeedleBench Tasks

We first introduce the several different subtasks within NeedleBench .

$\bullet$ Single-Needle Retrieval Task (S-RT): Tests LLMs’ ability to recall a single key information inserted at various positions in a long text, highlighting their precision in navigating and recalling single detail within extensive texts.

$\bullet$ Multi-Needle Retrieval Task (M-RT): Explores LLMs’ ability to retrieve multiple pieces of related information scattered across a lengthy text, simulating complex real-world queries that require extracting several data points from comprehensive documents.

$\bullet$ Multi-Needle Reasoning Task(M-RS): Evaluates LLMs’ ability for complex reasoning by extracting multiple pieces of information from long texts and using them to logically answer questions that demand an integrated understanding and reasoning of various text segments.

2.2 Ancestral Trace Challenge(ATC)

In extreme scenarios, we envision models capable of recalling and analytically understanding every detail of every sentence within the input long text on demand. Thus, we introduce the Ancestral Trace Challenge (ATC) to simulate complex long-context tasks in real-world scenarios.

In ATC experiments, we construct the problem using a series of simple first-order logical inferences, forming an information chain that LLMs need to fully understand in order to answer the question. Forgetting any key information in the long context directly results in the inability of the LLMs to provide the correct answer.LLMs will be asked to choose one correct answer from the four given options, This method can be expanded to more challenging logical relationships in the study by Besta et al. (2024)

In practical scenarios, longer text lengths inevitably mean that LLMs need to understand more complex layers of logical relationships. Using the ATC test, we can easily stress test the multi-step reasoning capabilities of LLMs. This methodology is scalable and can be extended to extensive context lengths. While the NIAH test aim to stress test the information retrieval capabilities of LLMs within long texts, the ATC test specifically aims to stress test their reasoning abilities in similar long-context scenarios. Actual examples of ATC test can be found in the Appendix, Figures 21 and 22.

2.3 Dataset Construction

2.3.1 Needle Design

To prevent the model’s inherent knowledge from interfering with its ability to retrieve information, we deliberately design the needles in both the Single-Needle Retrieval(S-RT) and Multi-Needle Retrieval tasks to be abstract and nonexistent in the real world. Models are required to answer questions based solely on the key information provided during the test.

For the Multi-Needle Reasoning task, we construct the needle corpus utilizing the $\mathcal{R}^{4}\mathcal{C}$ dataset(Inoue et al., 2019), an enhancement of the HotpotQA dataset(Yang et al., 2018) that incorporates ”Derivation” information detailing each reasoning step needed to answer questions, while also addressing the issue of unclear pronouns present in HotpotQA. The corpus is then translated into Chinese to ensure a high-quality dataset for evaluating the bilingual reasoning capabilities of the models.As illustrated in Figure 3, the majority of reasoning questions involve two or three steps, with fewer questions requiring four or more steps.

2.3.2 Haystack Design

We follow the method described in Kamradt (2023), using the PaulGrahamEssays dataset to extend the prompt to the target length.

In the design of the Chinese Haystack, we utilize the ChineseDomainModelingEval dataset released by Wei et al. (2023) to ensure the diversity and quality of Chinese text sources. This dataset covers a wide range of topics from finance to technology, providing high-quality, up-to-date Chinese articles and a stable benchmark for evaluating the ability of different models to handle domain-specific long texts.

3 Experiments

We primarily evaluate the performance of mainstream open-source LLMs on NeedleBench at various token lengths of 4K, 8K, 32K, and 200K due to the significant token consumption involved. For the ATC experiment, we expand our evaluation to include leading API models, such as GPT-4 Turbo and Claude-3-Opus, and perform multiple logical stress tests across both open-source and API models.The complete list of LLMs evaluated in our research is detailed in the Table 1. Detailed experiments settings for each task parameter can be found in the Table 7.

Series	Models
Claude	Claude-3-Opus
OpenAI	GPT4-Turbo
Zhipu AI	GLM4,GLM4-9B-Chat-1M,ChatGLM3-6B, ChatGLM3-6B-32K
LLaMA-2	LLaMA-2-7B, LLaMA-2-13B, LLaMA-2-70B
Baichuan2	Baichuan2-7B, Baichuan2-13B
Gemma	Gemma 2B, Gemma 7B
Yi	Yi-6B
OrionStarAI	Orion-14B-LongChat, OrionStar-Yi-34B
DeepSeek	DeepSeek-67B
WizardLM	WizardLM-70B
Qwen	Qwen-7B, Qwen-14B, Qwen-72B,Qwen-72B-vLLM, Qwen-1.5-0.5B, Qwen-1.5-1.8B, Qwen-1.5-4B, Qwen-1.5-14B-vLLM, Qwen-1.5-72B-vLLM
InternLM	InternLM-7B, InternLM2-7B, InternLM2-20B, InternLM2-7B-200K, InternLM2-20B-200K, InternLM2.5-7B-Chat-1M
Zephyr	Zephyr-7B Beta
Mistral	Mistral-7B Instruct v0.2, Mixtral-8x7B Instruct v0.1

Table 1: Evaluated LLMs. The ”-200K” suffix indicates models configured for up to 200,000 context token length, deployed with LMDeploy(Contributors, 2023). ”vLLM” denotes deployment via vLLM(Kwon et al., 2023).

3.1 Performance of NeedleBench Tasks

Experimental Setting

We use the recall accuracy of needles placed at different positions as a metric to evaluate the performance of the models. By sequentially averaging the performance across data sets of different lengths and depths, we obtain the performance of the models on each task within NeedleBench. Further averaging of the scores from different tasks provided an overall score , ensuring a balanced representation of each task’s contribution to the overall score. To ensure the stability of our results, we ran multiple iterations of each test. The number of tokens is calculated uniformly using the tokenizer from GPT-4²²2https://github.com/openai/tiktoken. In addition, we introduced different buffer size to mitigate the problem of significant differences between tokenizers of different models, which could prevent a model from receiving the full prompt. In our experiments, we consistently position the question command prompts at the end of the extended texts. We examin the impact of the question prompt’s placement as detailed in Section 4.

To evaluate the similarity between predictions and references under each specific task, we use the Levenshtein distance, with $P$ and $R$ representing the lists of predictions and references, respectively, each of length $n$ . The score for each pair $P_{i}$ and $R_{i}$ is adjusted for the presence of core keywords within $P_{i}$ , defined by the set $W_{i}$ for each $R_{i}$ . A penalty factor $\alpha=0.2$ is applied to the score calculation for predictions missing any core keywords from $R_{i}$ . The formula for an individual score, $\text{Score}_{i}$ , is as follows:

\text{Score}_{i}=\left\{\begin{array}[]{ll}100&\mbox{if }P_{i}\cap W_{i}\neq% \emptyset,\\ 100\cdot\alpha\cdot\left(1-\frac{d(P_{i},R_{i})}{\max(|P_{i}|,|R_{i}|)}\right)% &\mbox{otherwise,}\end{array}\right.

where $d(P_{i},R_{i})$ is the Levenshtein distance between $P_{i}$ and $R_{i}$ , and $|P_{i}|$ and $|R_{i}|$ denote their lengths. The final score is the average of $\text{Score}_{i}$ across $n$ repetitions.

Focusing on key findings, we present main results for 32K and 200K context lengths in Table 2 and 3. More comprehensive results are available in the Appendix B.

Model	Single-Retrieval			Multi-Retrieval			Multi-Reasoning			Overall
	Chinese	Englisch	Overall	Chinese	Englisch	Overall	Chinese	Englisch	Overall
Models with Fewer Than 7B Parameters
Qwen-1.5-4B	99.80	96.64	98.22	90.68	93.50	92.09	55.79	54.70	55.24	83.49
ChatGLM3-6B-32K	98.43	80.24	89.34	82.50	87.91	85.20	69.86	73.90	71.88	82.86
InternLM2-7B-200K	100.00	100.00	100.00	24.77	58.50	41.64	58.72	89.79	74.25	74.77
Mistral-7B Instruct v0.2	79.84	47.92	63.88	79.36	92.05	85.70	49.92	71.86	60.89	69.53
Qwen-1.5-1.8B	71.62	39.44	55.53	37.45	25.64	31.5 5	46.34	29.42	37.88	43.04
Qwen-7B	56.71	29.47	43.09	22.09	14.86	18.48	35.60	21.62	28.61	31.36
Qwen-1.5-0.5B	40.09	28.99	34.54	5.45	6.18	5.82	24.48	11.04	17.76	20.89
Zephyr-7B Beta	13.73	21.33	17.53	0.23	4.09	2.16	19.80	33.79	26.79	15.70
Models with 7-20B Parameters
Qwen-1.5-14B-vLLM	82.68	88.83	85.75	94.27	96.91	95.59	67.05	66.05	66.55	82.94
Orion-14B-LongChat	99.37	89.40	94.39	88.05	87.36	87.70	64.63	59.42	62.02	82.67
InternLM2-20B-200K	100.00	100.00	100.00	13.86	12.91	13.39	64.15	89.26	76.70	67.03
Models Larger Than 20B Parameters
Mixtral-8x7B Instruct v0.1	94.74	93.23	93.99	93.00	99.45	96.23	73.84	78.93	76.39	89.38
Qwen-1.5-72B-vLLM	96.67	63.97	80.32	84.95	82.41	83.68	82.53	85.02	83.77	82.36
Qwen-72B-vLLM	95.59	35.70	65.65	90.77	79.18	84.98	74.82	59.76	67.29	71.94

Table 2: Main Results of NeedleBench 32K.

Model	Single-Retrieval			Multi-Retrieval			Multi-Reasoning			Overall
	Chinese	Englisch	Overall	Chinese	Englisch	Overall	Chinese	Englisch	Overall
Models with Fewer Than 7B Parameters
InternLM2-7B-200K	98.98	98.69	98.83	12.45	25.50	18.98	48.90	63.96	56.43	62.15
Models with 7-20B Parameters
InternLM2-20B-200K	100.00	100.00	100.00	10.68	12.00	11.34	50.72	70.98	60.85	61.66
Orion-14B-LongChat	78.07	32.84	55.45	33.45	30.32	31.89	33.14	27.36	30.25	40.82

Table 3: Main Results of NeedleBench 200K.

NeedleBench 32K and 200K

InternLM2-7B-200K achieves perfection in Single-Retrieval, indicating strong capabilities in accurately retrieving single information points. Qwen-1.5-72B-vLLM excels in Multi-Reasoning with its substantial parameter count of 72B, demonstrating its advantage in understanding and reasoning about complex relationships between pieces of information. Mixtral-8x7B Instruct v0.1 demonstrates formidable overall performance, particularly in retrieval tasks. Models with larger parameter counts tend to achieve higher average scores (see Figure 5).

As the context length further extends to 200K, fewer open-source LLMs are capable of supporting such extended contexts, as illustrated in Figure 5. InternLM2-7B-200K continues to exhibit its consistent strength in Single-Retrieval tasks but experiences a significant performance decline in Multi-Retrieval tests. Upon further investigation, it is confirmed that this is due to InternLM2 overfitting the Single-Retrieval tasks in the training corpus, leading it to often respond with only one needle in the Multi-Retrieval tests.This indicates that a longer context length may require enhanced capabilities in instruction following and information retrieval from the model. On the other hand, Orion-14B-LongChat is more adept at Multi-Retrieval tasks but fails to effectively manage the challenges of extended texts in Single-Retrieval tasks, showing a decline when the context length reaches approximately 80K.

NeedleBench 1000K

In our extended evaluation, we push the context length further to 1000K tokens. We evaluate the InternLM2.5-7B-1M and GLM4-9B-Chat-1M models under this context length.

First, we present the comprehensive results of the two models under NeedleBenchin Table 4. It can be seen that in the vast majority of tasks, the InternLM2.5-7B-Chat-1M performs better compared to the GLM4-9B-Chat-1M.

Model	Single-Retrieval			Multi-Retrieval			Multi-Reasoning			Overall
	Chinese	Englisch	Overall	Chinese	Englisch	Overall	Chinese	Englisch	Overall
Models Support 1M tokens
InternLM2.5-7B-Chat-1M	95.73	95.13	95.43	57.91	70.95	64.43	72.37	69.89	71.13	78.84
GLM4-9B-Chat-1M	89.99	11.6	50.8	74.64	24	49.32	52.58	42.35	47.47	49.35

Table 4: Main Results of NeedleBench 1000K.

We investigate why the performance of GLM4-9B-Chat-1M is significantly worse than InternLM2.5-7B-Chat-1M. One of the tasks with a large performance difference is the Single-Retrieval task. We show the results under the default prompt settings in Figure 6. It is found that the English version of the GLM4-9B-Chat-1M model perform poorly (indicated by almost all red scores), even at relatively lower context lengths.

By examining the model outputs, we find that the GLM4-9B-Chat-1M model often fail to answer questions due to thinking that there was no relevant content in the original text. For example, the following scenarios in Figure7:

Figure 7: GLM4-9B-Chat-1M’s Most Common Error Cases in Single Retrieval

Therefore, we try removing the requirement for the model to first recall the most relevant sentence in the document to the question, and instead instruct the model to answer directly. The results obtained are shown in the new Figure 8, where we observe that the performance is now nearly perfect (all green).

The significant difference caused by this prompt indicates that the GLM4-9B-Chat-1M is highly sensitive to the prompt used. This suggests that the GLM4-9B-Chat-1M has the inherent capability to handle 1000K context length but may not always demonstrate this ability under different prompt settings. This implies that more sophisticated fine-tuning or alignment strategies may need to be introduced to fully utilize the model’s capabilities.

3.2 Ancestral Trace Challenge(ATC)

We perform extreme stress testing with multi-step tests on the ATC for mainstream LLMs. Specifically, we test LLMs on bilingual multi-step logical reasoning problems, with the number of steps ranging from 2 to 19. For each step count setting, models were required to respond in a single-choice format. We employ few-shot learning to instruct the LLMs, where the few-shot example count is 4. In the direct version test, models were asked to provide options A, B, C, or D without having to demonstrate their reasoning path. In contrast, we introduce reasoning paths in the responses for the few-shot examples to instruct LLMs on how to answer each question. This version of the test is labeled as the Reasoning (Rea.) version.

To mitigate the risk of LLMs randomly guessing correctly on multiple-choice questions, we use the Circular-Eval (CE) method(Liu et al., 2023), which systematically changes the options on each question.

Let $P_{\text{Step}}$ represent the score achieved in a task, which consists of questions designed to be solved in a specific number of steps, precisely, in $Step$ steps, where $P_{\text{Step}}=\sum_{i=1}^{R}CE(Q_{\text{Step}}^{(i)})$ , and $CE(Q_{\text{Step}}^{(i)})$ denotes the circular evaluation score for the $i$ th iteration of question $Q$ at a given number of steps, with the score being 100/R if the LLM accurately identifies the correct answer regardless of its position among the options A, B, C, and D, and 0 otherwise. Here, $R$ represents the number of repetitions, and in our experiments, $R$ is set to 10. $L$ denotes the number of reasoning steps in the most complex task, with $L$ being set to 19 in our case. The overall score (Overall Score) is obtained by averaging the scores for each task. The score for each task setting (Task Score) is calculated as follows:

P_{\text{Step}}=\sum_{i=1}^{R}CE(Q_{\text{Step}}^{(i)})\quad\text{Task Score}=% \frac{\sum_{\text{Step}=2}^{L}(P_{\text{Step}}\times\text{Step})}{\sum_{\text{% Step}=2}^{L}\text{Step}}

Model	ZH-Direct	EN-Direct	ZH-Rea.	EN-Rea.	Overall
API Models
Claude-3-Opus	45.34	56.77	58.15	70.16	57.61
GPT4-Turbo	21.69	66.24	40.26	64.97	48.29
GLM-4	28.47	57.46	35.82	50.21	42.99
Models with Fewer Than 7B Parameters
InternLM2-7B	11.53	18.10	25.34	34.23	22.30
InternLM2-7B-200K	12.49	16.03	23.33	32.43	21.07
Yi-6B	7.14	11.53	13.33	10.69	10.67
Qwen-7B	1.96	11.01	10.26	14.55	9.45
Zephyr-7B Beta	2.28	6.72	5.82	16.03	7.71
Mistral-7B Inst. v0.2	6.35	6.35	6.56	10.95	7.55
InternLM-7B	4.23	6.88	7.88	10.16	7.29
Baichuan2-7B	0.32	4.87	12.38	6.40	5.99
ChatGLM3-6B-32K	3.17	5.66	5.71	6.88	5.35
Gemma-7B	3.49	2.22	5.19	8.68	4.89
ChatGLM3-6B	4.55	4.50	4.60	5.24	4.72
Qwen-1.5-4B	0.63	2.06	3.07	1.75	1.88
LLaMA-2-7B	0.74	0.00	0.42	4.13	1.32
Gemma-2B	1.06	0.21	0.00	0.26	0.38
Qwen-1.5-1.8B	0.00	0.11	0.63	0.26	0.25
Qwen-1.5-0.5B	0.00	0.21	0.00	0.00	0.05
Models with 7-20B Parameters
InternLM2-20B	26.72	42.43	23.65	31.06	30.96
InternLM2-20B-200K	26.14	35.08	32.75	29.47	30.86
Qwen-14B	12.01	27.46	31.85	23.92	23.81
Qwen-1.5-14B-vLLM	19.84	8.25	36.72	2.96	16.94
Baichuan2-13B	1.06	12.86	21.90	10.11	11.48
LLaMA-2-13B	1.75	6.08	3.39	12.22	5.86
Orion-14B-LongChat	1.38	7.94	1.43	11.43	5.54
Models Larger Than 20B Parameters
DeepSeek-67B	27.88	46.51	38.68	63.07	44.03
Qwen-72B-vLLM	22.91	23.49	29.89	35.24	27.88
Qwen-1.5-72B-vLLM	35.66	12.75	36.93	2.80	22.04
OrionStar-Yi-34B	2.91	28.25	18.99	25.98	19.03
Mixtral-8x7B Inst. v0.1	15.50	13.65	19.63	26.51	18.82
Yi-34B	11.48	20.95	18.68	18.10	17.30
LLaMA-2-70B	7.83	15.45	17.14	25.82	16.56
WizardLM-70B	12.75	18.36	10.79	19.05	15.24

Table 5: ATC Results

We present the results of the ATC experiments as follows. In Table 5, we present the scores of various models on ATC subtasks as well as their overall performance, we can observe several key insights:

Adding Reasoning Paths Improves Model Performance: It is evident that the versions with reasoning paths generally score higher than those without reasoning paths. For instance, the EN-Rea. version of Claude-3-Opus scores 70.16, while the corresponding EN-Direct. version scores 56.77. This aligns with the observation that chain-of-thought (CoT) reasoning can enhance model performance.

Larger Parameter Count Improves Model ATC Scores: Furthermore, for models within the same series, such as Baichuan and LLaMA, there is a noticeable trend where an increase in parameter count corresponds to higher scores. For example, Baichuan2-7B has an overall score of 5.99, while Baichuan2-13B achieves a significantly higher score of 11.48. Similarly, LLaMA-2-7B scores 1.32 overall, LLaMA-2-13B scores 5.86, and LLaMA-2-70B reaches 16.56, demonstrating the same upward trend. This indicates that larger models tend to perform better in these tasks.

Superiority of Closed-Source Models: We also observe a clear superiority of API models over open-source LLMs, particularly highlighted by the performances of Claude-3-Opus and GPT4-Turbo. However, DeepSeek-67B stands out with reasoning abilities that are close to those of the top API models, suggesting a promising direction for bridging the performance gap between API and open-source models.

In Figure 9, we present the performance decline trend of various models on ATC with the increase in the number of reasoning steps. The vertical axis represents the accuracy of different models in addressing the corresponding questions at a given number of reasoning steps (number of needles). As the number of reasoning steps increases.The majority of models struggle to cope with the complexity of long texts while the accuracy of nearly all open-source LLMs drops below 10% when the number of reasoning steps reaches around 16, at which point the prompt length is merely approximately 1120 tokens, indicating their difficulty in handling complex logical problems that are likely present in long-context questions.

4 Discussion

In this section, we present a series of ablation studies to explore the factors influencing model performance on long-context tasks. We focus on three key aspects: the effect of model size on long-text capabilities, the impact of instruction following in multi-retrieval tasks, and the influence of prompt position on model scores, with the aim of providing a comprehensive understanding of how these factors affect the performance and reliability of LLMs in different scenarios.

Effect of Model Size on Long-text Capabilities

Model	Parameters	Score	Task Type
Baichuan2-7B	7B	81.30	Overall
Baichuan2-13B	13B	92.28	Overall
LLaMA-2-7B	7B	62.09	Overall
LLaMA-2-13B	13B	61.82	Overall
LLaMA-2-70B	70B	69.35	Overall
Qwen-1.5-0.5B	0.5B	59.40	Overall
Qwen-1.5-1.8B	1.8B	82.49	Overall
Qwen-1.5-4B	4B	86.98	Overall
Qwen-1.5-14B	14B	77.88	Overall
Qwen-1.5-72B	72B	94.76	Overall
Qwen-7B	7B	74.7	Overall
Qwen-14B	14B	83.71	Overall
Qwen-72B	72B	59.55	Overall
Qwen-7B	7B	80.18	S-RT
Qwen-14B	14B	94.12	S-RT
Qwen-72B	72B	33.04	S-RT
InternLM2-7B	7B	85.19	M-RS
InternLM2-20B	20B	90.66	M-RS

Table 6: Model Size Impact on Scores by Series

We examin the relationship between model size and performance on long text tasks in NeedleBench 4K, as detailed in table 6. Our results show a pattern that is in most cases consistent with the scaling laws(Kaplan et al., 2020) which argue that larger models typically exhibit superior performance, such as the Baichuan, LLaMA, and InternLM2 model series.

However, for the Qwen series LLMs, as detailed in Table 6, the Qwen-72B model shows a significant performance degradation, especially on the NeedleBench 4K single retrieval task. Upon verification, it is confirmed that the Qwen-72B model often fail by assuming that there is no content relevant to the query within the documents.Such findings indicate the need for more focused training and optimization strategies tailored to the specifics of long text tasks in order to fully exploit their increased capacities and conform to the predictions of the scaling law.

Exploring Instruction Following in Multi-Retrieval Tasks

In our evaluation of the Multi-Needle Retrieval tasks in NeedleBench 8K, we investigate how changing the number of ”needles,” or key pieces of information that models needle to retrieve from extensive texts, affects their information retrieval capabilities Figure 10.

Interestingly, rather than observing a straightforward decline in recall scores with an increase in the number of needles, we note exceptions in the performance of certain models. For instance, the Mixtral-8*7b-instruct-v0.1 model, presented in Figure 10(b), demonstrate an unexpected behavior in its ability to recall the very first needle. When the model is tasked with recalling a larger set of needles at different postions (including the initially requested first needle), its accuracy in recalling the initial piece of key information actually improved.

This pattern indicates that while these models have the inherent capability to accurately recall initial key information, this potential is not always fully utilized, suggesting room for improvement in their instruction-following abilities. The fluctuating performance, which paradoxically improves with more demanding instructions, reflects a need for enhancing the robustness of the models’ capability to follow instructions, particularly for tasks that require precise information retrieval.

Effect of Prompt Position on Model Score

In NeedleBench 32K, we study how prompt position (whether the question prompt is placed before the long text or at the end of the long text) affects model performance on various subtasks. The results, shown in Figure 11, are the changes in performance due to moving the prompt position from before the long texts to after the long text.The predominance of red blocks suggests that the majority of models tend to perform worse when the question prompt is placed at the beginning of the text. However, for InternLM2-7B-200K, placing the question prompt of the M-RT task before the long text helps significantly improve the model’s command following ability, resulting in score improvements.

5 Conclusion

In this report, we thoroughly evaluate the ability of LLMs to handle long-context information retrieval and reasoning. Our research uncovers notable limitations in the current open-source LLMs’ ability to interpret and reason over long texts. Despite the progress in single information retrieval capabilities under extended context length of models such as GPT-4 Turbo and Claude 3, our findings underscore the difficulties these models encounter due to the complexity of logical reasoning challenges that are likely present in real-world long-context tasks.. This suggests that there is significant room for improving the utility of LLMs in scenarios requiring intricate information retrieval and numerous reasoning steps.

Our report reveals the importance of concentrated efforts within the AI research community to improve LLMs’ long-context comprehension and reasoning abilities. Addressing the shortcomings identified in the NeedleBench assessments could enable future models to perform more accurate and sophisticated analyses, equipping them more effectively for intricate long-context tasks in real-world scenarios.

6 Limitations and Future Work

Limitations: One limitation identified in our experiments is related to the multi-needle reasoning task. The needles used in this task are derived from datasets curated from Wikipedia. It is challenging to determine whether the tested models are genuinely reasoning over the context or simply utilizing internal knowledge. We plan to address it in future work by incorporating ATC needles into the multi-needle reasoning task to further refine this aspect.

Future Work: The ATC test in this study demonstrates that for LLMs, long-text comprehension should not be confined to retrieval tasks alone. It reveals that logical reasoning within long texts is inherently more complex. Future LLMs could be specifically optimized for long-context reasoning to handle the intricate logical challenges that are likely to arise in such scenarios.

Additionally, future work should focus on improving the robustness of long-context models’ abilities, as our findings in NeedleBench 1000K experiments indicate that these models are highly prompt sensitive. Developing strategies to mitigate prompt sensitivity will be crucial to ensure consistent performance across different prompts and enhance the overall reliability of these models in handling extensive textual data.

References

Anthropic (2024a) Anthropic. Introducing claude 2.1. https://www.anthropic.com/index/claude-2-1, 2024a. Accessed: 2024-01-23.
Anthropic (2024b) Anthropic. Introducing the next generation of claude. https://www.anthropic.com/news/claude-3-family, 2024b. Accessed: 2024-03-27.
Bai et al. (2023) Yushi Bai, Xin Lv, Jiajie Zhang, Hongchang Lyu, Jiankai Tang, Zhidian Huang, Zhengxiao Du, Xiao Liu, Aohan Zeng, Lei Hou, Yuxiao Dong, Jie Tang, and Juanzi Li. Longbench: A bilingual, multitask benchmark for long context understanding, 2023.
Besta et al. (2024) Maciej Besta, Florim Memedi, Zhenyu Zhang, Robert Gerstenberger, Nils Blach, Piotr Nyczyk, Marcin Copik, Grzegorz Kwaśniewski, Jürgen Müller, Lukas Gianinazzi, Ales Kubicek, Hubert Niewiadomski, Onur Mutlu, and Torsten Hoefler. Topologies of reasoning: Demystifying chains, trees, and graphs of thoughts, 2024.
Contributors (2023) LMDeploy Contributors. Lmdeploy: A toolkit for compressing, deploying, and serving llm. https://github.com/InternLM/lmdeploy, 2023.
Gemini Team (2024) Gemini Team. Gemini 1.5: Unlocking multimodal understanding across millions of tokens of context. https://storage.googleapis.com/deepmind-media/gemini/gemini_v1_5_report.pdf, 2024. Accessed: 2024-03-17.
Inoue et al. (2019) Nozomu Inoue, Pontus Stenetorp, and Kentaro Inui. R4c: A benchmark for evaluating rc systems to get the right answer for the right reason. arXiv: Computation and Language,arXiv: Computation and Language, Oct 2019.
Kamradt (2023) Greg Kamradt. LLMs Need Needle In A Haystack Test-Pressure Testing LLMs. https://github.com/gkamradt/LLMTest_NeedleInAHaystack, 2023.
Kaplan et al. (2020) Jared Kaplan, Sam McCandlish, Tom Henighan, Tom B. Brown, Benjamin Chess, Rewon Child, Scott Gray, Alec Radford, Jeffrey Wu, and Dario Amodei. Scaling laws for neural language models, 2020.
Kwon et al. (2023) Woosuk Kwon, Zhuohan Li, Siyuan Zhuang, Ying Sheng, Lianmin Zheng, Cody Hao Yu, Joseph E. Gonzalez, Hao Zhang, and Ion Stoica. Efficient memory management for large language model serving with pagedattention. In Proceedings of the ACM SIGOPS 29th Symposium on Operating Systems Principles, 2023.
Liu et al. (2023) Yuan Liu, Haodong Duan, Yuanhan Zhang, Bo Li, Songyang Zhang, Wangbo Zhao, Yike Yuan, Jiaqi Wang, Conghui He, Ziwei Liu, Kai Chen, and Dahua Lin. Mmbench: Is your multi-modal model an all-around player?, 2023.
Mohtashami & Jaggi (2023) Amirkeivan Mohtashami and Martin Jaggi. Landmark attention: Random-access infinite context length for transformers, 2023.
OpenAI (2023) OpenAI. Gpt-4 and gpt-4 turbo - documentation. https://platform.openai.com/docs/models/gpt-4-and-gpt-4-turbo, 2023. Accessed: 2024-01-23.
OpenCompass Contributors (2023) OpenCompass Contributors. OpenCompass: A universal evaluation platform for foundation models. https://github.com/open-compass/opencompass, 2023.
Wei et al. (2023) Tianwen Wei, Liang Zhao, Lichang Zhang, Bo Zhu, Lijie Wang, Haihua Yang, Biye Li, Cheng Cheng, Weiwei Lü, Rui Hu, Chenxia Li, Liu Yang, Xilin Luo, Xuejie Wu, Lunan Liu, Wenjun Cheng, Peng Cheng, Jianhao Zhang, Xiaoyu Zhang, Lei Lin, Xiaokun Wang, Yutuan Ma, Chuanhai Dong, Yanqi Sun, Yifu Chen, Yongyi Peng, Xiaojuan Liang, Shuicheng Yan, Han Fang, and Yahui Zhou. Skywork: A more open bilingual foundation model, 2023.
Yang et al. (2018) Zhilin Yang, Peng Qi, Saizheng Zhang, Yoshua Bengio, William Cohen, Ruslan Salakhutdinov, and Christopher D. Manning. Hotpotqa: A dataset for diverse, explainable multi-hop question answering. In Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing, Jan 2018. doi: 10.18653/v1/d18-1259. URL http://dx.doi.org/10.18653/v1/d18-1259.
Zhang et al. (2023) Xinrong Zhang, Yingfa Chen, Shengding Hu, Qihao Wu, Junhao Chen, Zihang Xu, Zhenning Dai, Xu Han, Shuo Wang, Zhiyuan Liu, and Maosong Sun. Infinitebench: 128k long-context benchmark for language models, 2023.

7 Appendix

Appendix A Parameter Settings

We have carefully adjust parameters such as buffersize, repeats, depths, and lengths for various tasks, languages, and content lengths in NeedleBench . This ensures test durations are kept within acceptable limits, experiments remain reproducible, and the influence of tokenizer discrepancies among different models is minimized. Building on this foundation, we further refine our evaluation approach by quantifying the overall performance of models using a weighted scoring system. This system is designed to capture the comprehensive abilities of models in performing a range of tasks, from single-item retrieval to complex reasoning across multiple contexts.

Following the meticulous configuration of these parameters, the computation of the Overall Score ( $O$ ) is derived from the weighted averages of scores from different tasks. This is formalized as:

O=W_{S-RT}\cdot Score_{S-RT}+W_{M-RT}\cdot Score_{M-RT}+W_{M-RS}\cdot Score_{M% -RS}

Where $W_{S-RT}=0.4$ , $W_{M-RT}=0.3$ , and $W_{M-RS}=0.3$ are the weights assigned to the scores of the Single-Retrieval Task (S-RT), Multi-Retrieval Task (M-RT), and Multi-Reasoning Task (M-RS) respectively.

The Levenshtein distance $d$ we use in score calculation is defined as follows:

d(i,j)=\left\{\begin{array}[]{ll}\max(i,j)&\mbox{if }\min(i,j)=0,\\ \min\left\{\begin{array}[]{l}d(i-1,j)+1,\\ d(i,j-1)+1,\\ d(i-1,j-1)+1_{(s_{1}[i]\neq s_{2}[j])}\end{array}\right\}&\mbox{otherwise.}% \end{array}\right.

Here, $1_{(s_{1}[i]\neq s_{2}[j])}$ denotes an indicator function that equals 1 when $s_{1}[i]\neq s_{2}[j]$ and 0 otherwise. The term $d(i,j)$ represents the Levenshtein distance between the first $i$ characters of $s_{1}$ and the first $j$ characters of $s_{2}$ .

Table 7: Needlebench Parameter Settings:

Task	Repeats	Buffersize	# of Depth Intervals	# of Length Intervals
2-Needle-Reasoning-EN-4K	10	600	20	4
3-Needle-Reasoning-EN-4K	10	600	20	4
4-Needle-Reasoning-EN-4K	10	600	20	4
5-Needle-Reasoning-EN-4K	10	600	20	4
Single-Needle-Retrieval-EN-4K	10	600	20	4
Multi-Needle-Retrieval-EN-4K	25	1000	20	4
2-Needle-Reasoning-ZH-4K	10	200	20	4
3-Needle-Reasoning-ZH-4K	10	200	20	4
4-Needle-Reasoning-ZH-4K	10	200	20	4
5-Needle-Reasoning-ZH-4K	10	200	20	4
Single-Needle-Retrieval-ZH-4K	10	200	20	4
Multi-Needle-Retrieval-ZH-4K	25	200	20	4
2-Needle-Reasoning-EN-8K	10	1000	20	4
3-Needle-Reasoning-EN-8K	10	1000	20	4
4-Needle-Reasoning-EN-8K	10	1000	20	4
5-Needle-Reasoning-EN-8K	10	1000	20	4
Single-Needle-Retrieval-EN-8K	10	800	20	4
Multi-Needle-Retrieval-EN-8K	25	1300	20	4
2-Needle-Reasoning-ZH-8K	10	200	20	4
3-Needle-Reasoning-ZH-8K	10	200	20	4
4-Needle-Reasoning-ZH-8K	10	200	20	4
5-Needle-Reasoning-ZH-8K	10	200	20	4
Single-Needle-Retrieval-ZH-8K	10	200	20	4
Multi-Needle-Retrieval-ZH-8K	25	200	20	4
2-Needle-Reasoning-EN-32K	10	3000	10	8
3-Needle-Reasoning-EN-32K	10	3000	10	8
4-Needle-Reasoning-EN-32K	10	3000	10	8
5-Needle-Reasoning-EN-32K	10	3000	10	8
Single-Needle-Retrieval-EN-32K	10	3000	10	8
Multi-Needle-Retrieval-EN-32K	25	3000	10	8
2-Needle-Reasoning-ZH-32K	10	200	10	8
3-Needle-Reasoning-ZH-32K	10	200	10	8
4-Needle-Reasoning-ZH-32K	10	200	10	8
5-Needle-Reasoning-ZH-32K	10	200	10	8
Single-Needle-Retrieval-ZH-32K	10	200	10	8
Multi-Needle-Retrieval-ZH-32K	25	200	10	8
2-Needle-Reasoning-EN-200K	10	600	10	8
3-Needle-Reasoning-EN-200K	10	600	10	8
4-Needle-Reasoning-EN-200K	10	600	10	8
5-Needle-Reasoning-EN-200K	10	600	10	8
Single-Needle-Retrieval-EN-200K	10	600	10	8
Multi-Needle-Retrieval-EN-200K	25	3000	10	8
2-Needle-Reasoning-ZH-200K	10	200	10	8
3-Needle-Reasoning-ZH-200K	10	200	10	8
4-Needle-Reasoning-ZH-200K	10	200	10	8
5-Needle-Reasoning-ZH-200K	10	200	10	8
Single-Needle-Retrieval-ZH-200K	10	200	10	8
Multi-Needle-Retrieval-ZH-200K	25	200	10	8

Appendix B Detailed Experimental Results

In this section, we present the performance of models for different token lengths (4K, 8K, 32K, and 200K) in NeedleBench. The models are grouped according to their maximum sequence length, allowing a focused analysis on appropriate subtasks for each group. Detailed results for each length and task are presented in the corresponding tables: for 4K text length in table 8, for 8K text length in table 9

B.1 NeedleBench 4K

Model	Single-Retrieval			Multi-Retrieval			Multi-Reasoning			Overall
	Chinese	Englisch	Overall	Chinese	Englisch	Overall	Chinese	Englisch	Overall
Models with Fewer Than 7B Parameters
InternLM2-7B-200K	100.00	100.00	100.00	94.00	99.90	96.95	74.17	95.29	84.73	94.50
InternLM2-7B	100.00	100.00	100.00	91.15	99.85	95.50	74.98	95.39	85.19	94.21
ChatGLM3-6B-32K	99.77	99.53	99.65	94.50	91.10	92.80	71.55	80.67	76.11	90.53
Qwen-1.5-4B	100.00	100.00	100.00	92.55	93.20	92.88	58.45	68.97	63.71	86.98
Yi-6B	83.75	99.77	91.76	86.05	84.80	85.43	71.91	81.93	76.92	85.41
Qwen-1.5-1.8B	99.16	96.45	97.81	77.80	78.60	78.20	67.68	65.01	66.34	82.49
Baichuan2-7B	91.42	94.13	92.78	80.00	69.45	74.72	72.33	72.80	72.57	81.30
InternLM-7B	95.12	89.77	92.44	62.00	82.85	72.42	53.06	75.60	64.33	78.00
Qwen-7B	79.26	81.10	80.18	75.25	79.25	77.25	63.19	66.48	64.83	74.70
Gemma-7B	75.32	74.31	74.82	82.15	90.80	86.47	55.74	60.33	58.03	73.28
ChatGLM3-6B	87.12	84.15	85.64	78.75	86.30	82.53	43.69	45.06	44.38	72.33
Mistral-7B Inst. v0.2	68.02	38.23	53.12	91.75	99.05	95.40	34.11	76.92	55.52	66.52
LLaMA-2-7B	32.42	93.84	63.13	41.20	84.25	62.73	31.61	88.52	60.07	62.09
Qwen-1.5-0.5B	75.67	92.13	83.90	35.75	49.90	42.82	41.40	45.20	43.30	59.40
Zephyr-7B Beta	69.22	45.70	57.46	52.25	53.30	52.78	53.63	54.60	54.12	55.05
Gemma-2B	21.47	20.29	20.88	56.90	80.20	68.55	12.72	31.68	22.20	35.57
Models with 7-20B Parameters
Orion-14B-LongChat	99.51	99.29	99.40	98.15	98.55	98.35	68.24	87.29	77.76	92.59
Baichuan2-13B	99.31	98.73	99.02	94.45	96.90	95.67	80.76	79.01	79.89	92.28
InternLM2-20B-200K	100.00	100.00	100.00	65.75	86.70	76.23	84.49	96.77	90.63	90.06
InternLM2-20B	100.00	100.00	100.00	63.20	82.40	72.80	84.49	96.83	90.66	89.04
Qwen-14B	98.82	89.43	94.12	93.40	92.30	92.85	66.24	55.15	60.69	83.71
Qwen-1.5-14B-vLLM	49.78	84.57	67.18	94.80	93.60	94.20	71.62	83.36	77.49	78.38
LLaMA-2-13B	44.22	86.94	65.58	25.20	87.55	56.38	37.22	87.25	62.24	61.82
Models Larger Than 20B Parameters
Qwen-1.5-72B-vLLM	99.32	95.67	97.50	95.05	98.85	96.95	87.10	95.91	91.50	95.53
Yi-34B	92.86	94.41	93.63	99.15	99.50	99.33	78.14	83.26	80.70	91.46
Mixtral-8x7B Inst. v0.1	87.22	84.89	86.05	99.15	99.45	99.30	71.19	83.76	77.48	87.45
DeepSeek-67B	95.44	85.21	90.33	80.40	78.35	79.38	85.11	76.60	80.86	84.20
Qwen-72B-vLLM	73.32	29.37	51.34	99.15	92.55	95.85	71.45	63.76	67.60	69.57
LLaMA-2-70B	41.35	98.43	69.89	40.70	98.55	69.62	43.37	93.30	68.34	69.35
OrionStar-Yi-34B	82.42	41.24	61.83	82.05	72.25	77.15	63.76	77.44	70.60	69.06
WizardLM-70B	63.95	42.61	53.28	49.25	77.45	63.35	53.93	62.98	58.46	57.85

Table 8: Main Results of NeedleBench 4K. Overall stands for the score calculated from an average of metrics on all subsets. (bold denotes the best score among all models, and underline denotes the best score under the same model scale. For the following tables, the same notation applies.)

From Table 8, it is evident that the InternLM2 models are particularly adept at Single-Retrieval tasks, while also demonstrating commendable cross-document inference abilities in Multi-Reasoning tasks. Notably, InternLM2-7B-200K achieve the best overall score among models with fewer than 7 billion parameters, although it is outperformed by the larger Qwen-1.5-72B-vLLM model.

Figure 12 shows detailed graphs for selected representative models at various lengths and depths: The InternLM2 models show superior overall performance, while the LLaMA models show significantly weaker Chinese language capabilities compared to English. The Gemma series shows a significant improvement in combined scores as the parameter size increases from 2B to 7B.

B.2 NeedleBench 8K

Model	Single-Retrieval			Multi-Retrieval			Multi-Reasoning			Overall
	Chinese	Englisch	Overall	Chinese	Englisch	Overall	Chinese	Englisch	Overall
Models with Fewer Than 7B Parameters
ChatGLM3-6B-32K	100.00	99.04	99.52	87.90	89.55	88.73	73.26	82.48	77.87	89.79
InternLM2-7B	100.00	100.00	100.00	49.90	98.90	74.40	65.29	90.62	77.96	85.71
InternLM2-7B-200K	100.00	100.00	100.00	46.25	99.00	72.62	65.09	91.38	78.23	85.26
Qwen-1.5-4B	99.89	97.96	98.92	92.05	92.35	92.20	56.25	57.41	56.83	84.28
Qwen-1.5-1.8B	99.20	94.01	96.61	65.85	62.75	64.30	61.15	51.65	56.40	74.85
Mistral-7B Instruct v0.2	76.63	40.92	58.78	86.55	97.90	92.22	42.59	76.48	59.54	69.04
ChatGLM3-6B	85.42	48.67	67.04	57.45	65.85	61.65	42.97	24.14	33.55	55.38
Gemma-7B	76.51	55.55	66.03	54.50	44.05	49.28	50.10	31.36	40.73	53.41
Qwen-1.5-0.5B	69.53	78.60	74.07	13.80	23.35	18.58	36.67	29.00	32.84	45.05
Qwen-7B	63.41	53.55	58.48	22.95	17.90	20.43	49.45	51.71	50.58	44.69
InternLM-7B	59.62	38.87	49.25	12.65	12.65	12.65	48.26	52.32	50.29	38.58
Zephyr-7B Beta	34.16	42.73	38.45	0.10	11.35	5.73	34.57	44.88	39.73	29.01
Gemma-2B	15.57	5.59	10.58	16.90	14.80	15.85	11.66	10.12	10.89	12.25
Models with 7-20B Parameters
Orion-14B-LongChat	100.00	99.41	99.71	95.05	94.75	94.90	65.33	79.96	72.65	90.15
Qwen-1.5-14B-vLLM	67.67	83.33	75.50	88.55	94.30	91.42	66.64	74.57	70.60	78.81
InternLM2-20B	100.00	100.00	100.00	23.75	47.55	35.65	73.87	94.73	84.30	75.99
InternLM2-20B-200K	100.00	100.00	100.00	26.55	44.55	35.55	73.84	94.76	84.30	75.96
Qwen-14B	90.35	46.28	68.32	79.85	59.15	69.50	57.46	34.74	46.10	62.01
Models Larger Than 20B Parameters
Qwen-1.5-72B-vLLM	99.39	79.27	89.33	92.40	93.65	93.03	86.23	93.87	90.05	90.65
Mixtral-8x7B Instruct v0.1	94.83	83.67	89.25	98.70	100.00	99.35	71.55	81.31	76.43	88.43
Qwen-72B-vLLM	89.19	16.29	52.74	99.10	73.05	86.08	66.79	45.42	56.11	63.75

Table 9: Main Results of NeedleBench 8K.

In NeedleBench 8K, we observe a general upward trend in scores with increasing model parameters, with Qwen-1.5-72B, the model with the largest number of parameters, achieving the highest score of 90.65. The InternLM2 series, however, does not maintain a leading position, primarily due to noticeable issues with command following in Multi-Retrieval tasks (recalling only a single needle instead of multiple needles as per the command). Among models with less than 7B parameters, ChatGLM3-6B scores the highest within the 8K length and closely trails behind Qwen-1.5-72B in performance.

Appendix C NeedleBench Prompt Examples

Figure 16: An example prompt of Single-Needle Retrieval showcasing key information with the single needle positioned at the very beginning. In actual tests, the needle is placed at various depths within extended texts to evaluate performance under different conditions.

Figure 17: An example prompt of Single-Needle Retrieval showcasing key information with the single needle positioned at the middle

Figure 18: An example prompt of Single-Needle Retrieval showcasing key information with the single needle positioned at last

Figure 19: An example prompt of Multi-Needle Retrieval

Figure 20: An example prompt of Multi-Needle Reasoning

Figure 21: Examples of 4-shot prompts and responses in the Ancestral Trace Challenge(ATC) without reasoning paths for six-step logic relations

Figure 22: Examples of 4-shot prompts and responses in the Ancestral Trace Challenge(ATC) with reasoning paths for six-step logic relations