← Back to all development cases
This case was judged by an anonymized AI reviewer and has not been independently reviewed by people; inspect the draft, cards, and both outputs to assess it.
AutocompleteWith cards better272 / 280 · beu-b7f5f78462e078ac

A Scalable Benchmark for Repository-Oriented Long-Horizon Conversational Context Management

Computer science · 2603.06358v1

FLOWING EVIDENCE BENCHMARK

How do we compare writing with and without cards?

At the same writing position, the same model produces one first result with cards and one without. All 280 pairs enter the overall anonymized review, which assesses whether a decision not to continue fits the draft.

Overall: 138 with-card wins, 26 ties, 107 no-card wins, 5 pairs where neither was usable, and 4 inconclusive judgments. Both sides wrote text in 218 pairs; 48 had a deliberate stop and 14 included an invalid output.

SAME DRAFT · WITH OR WITHOUT CARDS

01 · FIXED WRITING POSITION

MANUSCRIPT

Same writing position ▌

Matched autocomplete at the same draft position

02 · TWO MATCHED VERSIONS

SHARED BY BOTH

Draft, writing task, and generation model

A · With cards

Relevant literature cards supplied

B · Without cards

No literature cards supplied

Generation model

Same model and version

A → first output

B → first output

Anonymized AI judge

First outputs are anonymized as X and Y

Continuation review: accuracy, fit to the writing task and usability
Returns: X preferred / tie / Y preferred / both unusable

Review again with the answer order swapped

HOW DOES BLIND JUDGING WORK?

① Anonymize both outputs
The AI judge sees the same draft and both first outputs without knowing which received cards.

② Compare and swap order
The judge applies task-specific criteria in X/Y and then Y/X order.

③ Review disagreements
A third pass resolves disagreements. Incomplete reviews remain in the denominator.

How are positions and cards selected?

We selected 40 writing positions in each of seven fields, for 280 comparisons. The position and relevant literature cards were set before each comparison, so both versions began from the same draft.

The main result retains every frozen position, including places where cards might not add a useful next move. This avoids selecting cases based on their outputs, but 49% describes the whole cohort, not a win rate where evidence clearly had an opportunity to help. This cohort did not record that opportunity label for each position before generation, so those two groups cannot yet be separated reliably.

CARD USE · 140/280

Reported card use in continuations with text

The model reports whether a card affected its continuation. In 140 of 280 tests it produced continuation text and reported card use. The previous count of 150 included 10 cases without continuation text and has been corrected.

COMPARISON · 70 / 7 / 30

Which written version was better?

Of the 107 reported-use cases where both versions produced text, the blinded AI judge preferred the with-card version 70 times, found 7 about the same, and preferred the without-card version 30 times.

How are the two table percentages calculated?

ADOPTION COVERAGE

140reported use with text
÷
280all positions
=
50%

WITH-CARD VERSION PREFERRED

70with cards better
÷
107comparable tests
=
65%

The percentages use different groups: 50% counts continuations with text reporting card use among all 280 tests; 65% covers the 107 such cases with text on both sides. The generating model reports card use, while an anonymized AI judge compares writing quality.

Card use is self-reported by the generator. The anonymized AI judge sees the draft and source snippets, so source-aware preference remains possible; there was no independent human review. The overall result describes these 280 frozen positions, not a win probability for any writing task.

Manuscript writing position

Text excerpt · not a PDF page

Research manuscript · excerpt

A Scalable Benchmark for Repository-Oriented Long-Horizon Conversational Context Management

body

Early benchmarks were limited to static settings, where the evaluation is performed in a single-shot manner and overlooks the examination of multi-turn interactions. Representative tasks include single-turn repository-level code generation [11, 24, 25, 59, 63, 67], code understanding [12, 23, 35, 36], code translation [33, 48, 66], issue resolving [20, 61, 64], and vulnerability detection & repair [ 29, 41, 58]. However, in real-world usage scenarios, interactions between users and LLMs-based systems are often multi-turn and dynamic, which further requires benchmarks to simulate such interaction patterns in order to evaluate the corresponding capabilities. Depending on the specific capabilities of interest, several multi-turn benchmarks are proposed, which can be roughly categorized into several groups:❶ multi-turn instruction-following ability, such as CoCoPIF [15] and CodeIF-Bench [45];❷multi-turn

↓ Next paragraph was generated here

The highlighted text is near the writing position. This is a layout of the evaluation input, not a PDF screenshot.

A matched comparison at the same writing positionOverall AI review: With cards better

Both arms use the same model, task and manuscript context. Only the retrieved passages differ. The first recorded output from each arm is shown below.

Read overall AI review rationale

A = with cards · B = without cards

The draft is in the middle of categorizing multi-turn benchmarks, having just introduced category ❶ (multi-turn instruction-following) and started category ❷ with "multi-turn". The sentence is clearly incomplete and needs continuation. Output A provides a plausible continuation that completes the categorization scheme with additional benchmark types (code generation/editing, debugging/repair), fitting the established pattern and academic style. The mention of "RepoBench" is a reasonable example for repository-level code generation in conversational settings, consistent with the earlier citation of repository-level code generation benchmarks. Output B returns "invalid" with no text, which is unusable because the draft manifestly requires continuation—the sentence is grammatically incomplete with an open enumeration (❶...❷...) that needs completion.

Evidence

Same model and position · retrieved passages supplied · first output

Preferred by blind review

code generation and editing ability, such as RepoBench and its conversational extensions; and ❸ multi-turn debugging and repair ability, such as interactive fault localization and program repair benchmarks.

Passages supplied to the Evidence version

Card 1 · CodeIF-Bench: Evaluating Instruction-Following Capabilities of Large Language Models in Interactive Code Generation

However, existing code generation benchmarks primarily assess the functional correctness of code generated by LLMs in single-turn interactions, offering limited insight into their capabilities to generate code that strictly follows users’ instructions, especially in multi-turn interaction scenarios.…

Read full passage excerpt

However, existing code generation benchmarks primarily assess the functional correctness of code generated by LLMs in single-turn interactions, offering limited insight into their capabilities to generate code that strictly follows users’ instructions, especially in multi-turn interaction scenarios. In this paper, we introduce CodeIF-Bench, a benchmark for evaluating LLMs’ instruction- following capabilities in interactive code generation

Card 2 · Cambrian-1: A Fully Open, Vision-Centric Exploration of Multimodal LLMs

We introduce Cambrian-1, a family of multimodal LLMs (MLLMs) designed with a vision-centric approach. While stronger language models can enhance multimodal capabilities, the design choices for vision components are often insufficiently explored and disconnected from visual representation learning re…

Read full passage excerpt

We introduce Cambrian-1, a family of multimodal LLMs (MLLMs) designed with a vision-centric approach. While stronger language models can enhance multimodal capabilities, the design choices for vision components are often insufficiently explored and disconnected from visual representation learning research. This gap hinders accurate sensory grounding in real-world scenarios. Our study uses LLMs and visual instruction tuning as an interface to evaluate various visual representations, offering new insights into different models and architectures -- self-supervised, strongly supervised, or combinations thereof -- based on experiments with over 20 vision encoders. We critically examine existing MLLM benchmarks, address the difficulties involved in consolidating and interpreting results from various tasks, and introduce a new vision-centric benchmark, CV-Bench. To further improve visual grounding, we propose the Spatial Vision Aggregator (SVA), a dynamic and spatially-aware connector that in …

Card 3 · HoPE: Hybrid of Position Embedding for Long Context Vision-Language Models

Vision-Language Models (VLMs) have made significant progress in multimodal tasks. However, their performance often deteriorates in long-context scenarios, particularly long videos. While Rotary Position Embedding (RoPE) has been widely adopted for length generalization in Large Language Models (LLMs…

Read full passage excerpt

Vision-Language Models (VLMs) have made significant progress in multimodal tasks. However, their performance often deteriorates in long-context scenarios, particularly long videos. While Rotary Position Embedding (RoPE) has been widely adopted for length generalization in Large Language Models (LLMs), extending vanilla RoPE to capture the intricate spatial-temporal dependencies in videos remains an unsolved challenge. Existing methods typically allocate different frequencies within RoPE to encode 3D positional information. However, these allocation strategies mainly rely on heuristics, lacking in-depth theoretical analysis. In this paper, we first study how different allocation strategies impact the long-context capabilities of VLMs. Our analysis reveals that current multimodal RoPEs fail to reliably capture semantic similarities over extended contexts. To address this issue, we propose HoPE, a Hybrid of Position Embedding designed to improve the long-context capabilities of VLMs. HoPE …

Card 4 · NORA: A Small Open-Sourced Generalist Vision Language Action Model for Embodied Tasks

Existing Visual-Language-Action (VLA) models have shown promising performance in zero-shot scenarios, demonstrating impressive task execution and reasoning capabilities. However, a significant challenge arises from the limitations of visual encoding, which can result in failures during tasks such as…

Read full passage excerpt

Existing Visual-Language-Action (VLA) models have shown promising performance in zero-shot scenarios, demonstrating impressive task execution and reasoning capabilities. However, a significant challenge arises from the limitations of visual encoding, which can result in failures during tasks such as object grasping. Moreover, these models typically suffer from high computational overhead due to their large sizes, often exceeding 7B parameters. While these models excel in reasoning and task planning, the substantial computational overhead they incur makes them impractical for real-time robotic environments, where speed and efficiency are paramount. To address the limitations of existing VLA models, we propose NORA, a 3B-parameter model designed to reduce computational overhead while maintaining strong task performance. NORA adopts the Qwen-2.5-VL-3B multimodal model as its backbone, leveraging its superior visual-semantic understanding to enhance visual reasoning and action grounding. A …

Card 5 · Multimodal Large Models Are Effective Action Anticipators

The task of long-term action anticipation demands solutions that can effectively model temporal dynamics over extended periods while deeply understanding the inherent semantics of actions. Traditional approaches, which primarily rely on recurrent units or Transformer layers to capture long-term depe…

Read full passage excerpt

The task of long-term action anticipation demands solutions that can effectively model temporal dynamics over extended periods while deeply understanding the inherent semantics of actions. Traditional approaches, which primarily rely on recurrent units or Transformer layers to capture long-term dependencies, often fall short in addressing these challenges. Large Language Models (LLMs), with their robust sequential modeling capabilities and extensive commonsense knowledge, present new opportunities for long-term action anticipation. In this work, we introduce the ActionLLM framework, a novel approach that treats video sequences as successive tokens, leveraging LLMs to anticipate future actions. Our baseline model simplifies the LLM architecture by setting future tokens, incorporating an action tuning module, and reducing the textual decoder layer to a linear layer, enabling straightforward action prediction without the need for complex instructions or redundant descriptions. To further …

Model-reported card use

The model did not report that any card materially affected this completion.

No Evidence

Same model and position · no retrieved passages · first output

Not preferred in this comparison

No continuation text was produced.

Model action:invalid

Earlier card-use comparison

The earlier diagnostic compared only cases with reported card use and text on both sides; its scope differs from the overall AI review above.

This case does not have both with-card continuation text and a model-reported use record. It remains among all 280 tests but does not enter the reported-use comparison.

The passages are input material; their presence does not mean an output used them or that they support every claim. Draft and source passages are excerpted; no simulated scores or PDF appear here.