💻 computer science

CASHEW: Stabilizing Multimodal Reasoning via Iterative Trajectory Aggregation

The paper introduces CASHEW, an inference-time framework, and CASHEW-RL, a learned variant trained with Group Sequence Policy Optimization, to stabilize multimodal reasoning by iteratively aggregating candidate trajectories and filtering hallucinations through visual verification, achieving significant performance gains across 13 image and video understanding benchmarks.

Original authors: Chaoyu Li, Deeparghya Dutta Barua, Fei Tao, Pooyan Fazli

Published 2026-06-16

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Chaoyu Li, Deeparghya Dutta Barua, Fei Tao, Pooyan Fazli

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are trying to solve a tricky puzzle, like figuring out exactly what a cup of water is doing in a painting video. If you ask a standard AI (a Vision-Language Model) just once, it might guess, "Maybe the artist is storing the brush there?" It's confident, but it's wrong. If you ask it again, it might guess, "Maybe it's for resting the brush?" It's still guessing, and its answers keep changing. This is the problem the paper calls unstable reasoning.

The authors of this paper, from Arizona State University and NewsBreak, created a solution called CASHEW (and a smarter version called CASHEW-RL) to fix this. Here is how it works, using simple analogies:

1. The Problem: The "Lonely Thinker"

Current AI models are like a single person trying to solve a puzzle in a dark room. They might trip over their own feet (make a mistake) and keep walking in the wrong direction because they can't see the whole picture clearly. They often "hallucinate," meaning they invent facts that aren't there, like saying a cup is for storage when it's actually for cleaning.

2. The Solution: The "Study Group" (CASHEW)

Instead of letting the AI think alone, CASHEW turns the AI into a study group.

The Process: When the AI gets a question, it doesn't just give one answer. It generates a whole crowd of different "thought paths" (trajectories). Imagine 8 different students in a room, each trying to solve the puzzle on their own.
The Reality Check (Visual Verification): This is the secret sauce. Before the group agrees on an answer, a "referee" (a visual verification tool) checks their work against the actual image or video. If a student says, "The cup is for storage," the referee looks at the video and says, "Nope, I don't see a storage shelf. I see the artist dipping the brush to clean it."
The Synthesis: The AI then takes the best parts of all 8 students' ideas, filters out the lies (hallucinations) using the referee's notes, and combines them into one super-accurate, evidence-based answer.

It's like taking a messy brainstorming session and turning it into a polished, fact-checked report.

3. The Upgrade: The "Internalized Expert" (CASHEW-RL)

The first version (CASHEW) is great, but it requires the computer to run the "study group" process every single time it answers a question, which can be slow.

The authors also created CASHEW-RL. Think of this as taking that study group and teaching the AI to become the group leader itself.

They trained the AI using a special reward system (like a video game score) that gave points for:
1. Getting the right answer.
2. Citing the correct visual evidence (like pointing to the cup).
3. Not wasting time on easy questions (being efficient).
After this training, the AI learned to do the "group thinking" and "fact-checking" all by itself, inside its own brain, much faster and more efficiently.

What Did They Find?

The paper tested this on 13 different benchmarks involving images and videos. The results were like finding a magic wand for AI reasoning:

Big Jumps in Accuracy: On a science quiz called ScienceQA, the AI's score jumped by 26.2 percentage points. On a video reasoning test called EgoSchema, it jumped by 9.1 percentage points.
Less Hallucination: The AI stopped making up facts. It started sticking to what it could actually see in the picture.
Better Than the Competition: When compared to other methods that try to fix AI reasoning (like just asking the AI the same question 8 times and picking the most common answer), CASHEW won every time.

In a Nutshell

The paper claims that by forcing AI to think in groups, check its work against the visual evidence, and learn from its mistakes, we can make it much more reliable. It stops the AI from confidently guessing and starts it on a path of careful, evidence-based reasoning.

Technical Summary: CASHEW – Stabilizing Multimodal Reasoning via Iterative Trajectory Aggregation

1. Problem Statement

Vision-language models (VLMs) have achieved significant success in multimodal understanding and reasoning tasks. However, their multi-step reasoning processes remain fragile and unstable. Key issues include:

Instability: Repeated sampling on the same input often yields divergent reasoning trajectories and inconsistent final predictions.
Hallucinations: Models are prone to visual hallucinations, where reasoning steps are not grounded in actual visual evidence.
Limitations of Current Test-Time Scaling: Existing approaches often follow a "sample-and-select" paradigm, generating multiple independent chains and discarding partial insights from rejected trajectories. Furthermore, simply increasing reasoning length or sampling count without explicit grounding can amplify early perceptual errors rather than correcting them, leading models to "think longer" but not "think better."

2. Methodology

The paper proposes two complementary approaches to address these limitations: CASHEW (an inference-time framework) and CASHEW-RL (a learned variant).

CASHEW: Iterative Aggregation Framework

CASHEW treats reasoning as an evolutionary process where a population of candidate trajectories is iteratively synthesized and refined. It operates in four stages:

Population Initialization: The base VLM generates an initial population of $N$ candidate reasoning trajectories.
Subset Sampling: At each iteration, a random subset of peer trajectories is sampled as reference candidates to introduce diversity and prevent collapse into a single reasoning mode.
Visual Verification: Before aggregation, a lightweight visual-text grounding module (implemented using Grounding DINO) verifies object- and attribute-level claims against the visual input. Only grounded evidence is retained to guide subsequent steps.
Grounded Aggregation: The model synthesizes the sampled trajectories and their verified visual evidence to generate an updated, higher-quality reasoning trajectory. This process repeats for $T$ iterations, culminating in a final consolidated trajectory.

CASHEW-RL: Learned Aggregation via GSPO

CASHEW-RL internalizes the aggregation behavior into a single model through a two-stage post-training process:

Supervised Fine-Tuning (SFT): The model is trained on a structured aggregation format containing three components: a reasoning trace (<thought>), a set of verified visual keys (<visual_keys>), and the final answer (<answer>). This establishes a reference policy ( $\pi_{ref}$ ) and teaches the model to articulate evidence and format outputs.
Reinforcement Learning (RL) via Group Sequence Policy Optimization (GSPO): The model is optimized using a composite reward function that balances three objectives:
- Answer Correctness ( $R_{acc}$ ): Exact match with ground truth.
- Evidence Selection Quality ( $R_{key}$ ): A balanced precision-recall formulation encouraging the selection of relevant visual keys while avoiding indiscriminate proliferation.
- Difficulty-Aware Length Penalty ( $R_{len}$ ): An adaptive penalty that discourages excessive verbosity on easy tasks (where high solve rates are observed) while allowing flexibility for difficult tasks.

The training employs a curriculum-based mixture of offline teacher-generated trajectories and on-policy trajectories, gradually increasing the proportion of self-generated data to improve robustness and reduce distribution shift.

3. Key Contributions

CASHEW: A plug-and-play inference-time framework that stabilizes multimodal reasoning by iteratively aggregating candidate trajectories with explicit visual verification, effectively filtering hallucinations and grounding reasoning in evidence.
CASHEW-RL: A learned variant that internalizes trajectory aggregation via GSPO and a composite reward, enabling the model to perform robust, visually grounded aggregation with adaptive control over reasoning effort.
Empirical Validation: Extensive experiments demonstrating consistent performance improvements across diverse model backbones and modalities.

4. Experimental Results

The methods were evaluated on 13 benchmarks covering image understanding, video understanding, and video reasoning.

Performance Gains:
- Image Benchmarks: CASHEW-RL achieved significant improvements, including +26.2 percentage points on ScienceQA (Qwen3-VL-4B) and +9.1 percentage points on EgoSchema (InternVL3.5-8B).
- Video Benchmarks: Consistent gains were observed across Video-MME, LongVideoBench, and MVBench. For instance, CASHEW-RL improved EgoSchema scores by +9.1 points for InternVL3.5-8B.
Comparison with SOTA: CASHEW outperformed existing test-time scaling baselines (Self-Consistency, Self-Selector, and Self-Synthesizer) across all metrics. Notably, on the MME benchmark, it surpassed the strongest baseline (Self-Synthesizer) by +48.2 and +82.6 points on perception and cognition scores, respectively.
Ablation Studies: Removing the visual grounding module (Grounding DINO) still yielded improvements over the base VLM, indicating that iterative aggregation is the primary driver of performance, with visual grounding further enhancing reliability.
Efficiency: CASHEW-RL achieved performance comparable to deeper CASHEW configurations with fewer aggregation iterations, suggesting improved aggregation efficiency.

5. Significance and Claims

The paper claims that CASHEW and CASHEW-RL offer a robust solution to the instability of multimodal reasoning. By shifting from a "sample-and-select" paradigm to an "iterative aggregation" paradigm, the methods ensure that reasoning consensus is anchored in visual evidence. The authors emphasize that their approach allows models to "think together" rather than just "think longer," effectively mitigating hallucinations and improving consistency. The integration of visual verification and adaptive reasoning effort via GSPO enables the model to produce answers that are not only correct but also grounded in minimal yet sufficient visual evidence.

6. Limitations

The authors acknowledge that the current work uses a fixed aggregation budget to enable controlled comparisons, which may not be optimal for resource-constrained applications where easy examples might not require extensive aggregation. While CASHEW-RL partially addresses this by internalizing aggregation behavior, a systematic study of adaptive budget allocation is identified as future work.