Contrastive Distribution Matching for Amortized Sequential Monte Carlo in Discrete Diffusion
Listen
Radio episode about this paper
Transcript
Introduction to the show: ident: AI Radio. Generated commentary on the latest Artificial Intelligence papers.
Tom: We're looking at the paper: "Contrastive Distribution Matching for Amortized Sequential Monte Carlo in Discrete Diffusion".
Jane: Here's the gist of the paper. Discrete diffusion models have emerged as powerful frameworks for generating structured categorical data; however, efficiently sampling from reward-tilted distributions remains a fundamental challenge The introduction of Contrastive Distribution Matching (CDM) provides a novel framework that amortizes the cost of Sequential Monte Carlo inference by learning a parameterized twist function via positive and negative samples The Core Problem Addressed The central bottleneck in twisted SMC inference is estimating the optimal twist function, which typically requires costly Monte Carlo approximations in discrete state spaces This limitation arises because, unlike continuous diffusion where Tweedie’s formula provides an efficient estimate, discrete diffusion lacks an analogous relation The standard practice involves Monte Carlo estimation, which introduces a significant inference overhead when the downstream reward is computationally expensive How the Training Signal is Built CDM learns the twist function via a contrastive learning objective that minimizes the forward KL divergence against the target distribution The gradient of this objective exhibits a contrastive structure, utilizing positive and negative samples to upweight high-reward regions while downweighting suboptimal ones A novel training scheme leverages the closed-form forward process of the diffusion model for efficient training, maintaining a buffer of positive samples drawn from the approximated target distribution and applying the closed-form forward kernel This allows a single clean sample to be reused across multiple timesteps and gradient updates Where the Gains Come From The efficiency of CDM stems from several key design choices, including the use of diffusion-native training schemes that leverage the closed-form forward kernel for efficient training that scales to expensive reward functions The twist function is parameterized as a lightweight scalar head attached to the final feature layer of the pretrained model, incurring negligible runtime overhead, around 5% of the backbone forward-pass time and as little as 0.5% in some configurations The contrastive learning objective yields superior performance and more efficient training compared to the standard regression-based twist objective employed in previous discrete diffusion models Empirical Validation and Versatility Extensive empirical evaluations demonstrate that CDM consistently outperforms existing baselines under matched wall-clock time across a diverse range of applications, including toxic text generation, regulatory DNA sequence design, protein generation, and diffusion large language model alignment The framework is agnostic to the choice of the proposal distribution, allowing it to be paired with any proposal distribution for further synergistic performance gains Furthermore, CDM delivers superior efficacy than the regression-based twist objectives employed in previous discrete diffusion models Inference-Time Scaling The framework addresses inference-time scaling by reducing the cost of applying the twist at inference time to a constant-time operation This is achieved by learning the twist function, which reduces the twist evaluation to a constant-time operation and remains applicable regardless of the chosen proposal for further improvements The resulting particle system yields asymptotically consistent estimators of expectations under the intermediate target p∗t, together with an empirical approximation of p∗t itself In this work, we presented Contrastive Distribution Matching (CDM), a framework that enables amortized Sequential Monte Carlo (SMC) inference in discrete diffusion by learning twist functions via a contrastive objective Our empirical results demonstrate that CDM achieves superior scaling performance and more efficient training compared to the baselines Furthermore, we showed that CDM yields synergistic gains when integrated with fine-tuned proposals The framework shows potential for adapting to binary and categorical reward structures in future work Pruning redundant reasoning tokens Discrete diffusion models have emerged as powerful frameworks for generating structured categorical data; however, efficiently sampling from reward-tilted distributions remains a fundamental challenge The introduction of Contrastive Distribution Matching (CDM) provides a novel framework that amortizes the cost of Sequential Monte Carlo inference by learning a parameterized twist function via positive and negative samples How the training signal is built CDM learns the twist function via a contrastive learning objective that minimizes the forward KL divergence against the target distribution The gradient of this objective exhibits a contrastive structure, utilizing positive and negative samples to upweight high-reward regions while downweighting suboptimal ones A novel training scheme leverages the closed-form forward process of the diffusion model for efficient training that scales to expensive reward functions This allows a single clean sample to be reused across multiple timesteps and gradient updates Where the gains come from The efficiency of CDM stems from several key design choices, including the use of diffusion-native training schemes that leverage the closed-form forward kernel for efficient training that scales to expensive reward functions The twist function is parameterized as a lightweight scalar head attached to the final feature layer of the pretrained model, incurring negligible runtime overhead, around 5% of the backbone forward-pass time and as little as 0.
Paper summary: Tom: So, we're looking at this paper today, "Contrastive Distribution Matching for Amortized Sequential Monte Carlo in Discrete Diffusion". The main thing they're tackling is the problem of sampling from reward-tilted distributions in discrete diffusion models. They claim that their method, called Contrastive Distribution Matching or CDM, lets them sample efficiently without having to do those expensive Monte Carlo approximations every single time during inference.
Jane: That sounds like a big deal because traditional twisted sequential Monte Carlo inference has this huge bottleneck when dealing with discrete state spaces. It means estimating the optimal twist function usually requires costly Monte Carlo simulations, which becomes really slow when the reward function itself is computationally heavy, like in protein design or something similar.
Lu: Exactly. The core issue they point to is that unlike continuous diffusion where you have Tweedie's formula giving an efficient estimate of the clean state, discrete diffusion just doesn't have that direct analogy for estimating the twist function sixty-five <ref:2605.23346#pg2,efficient estimate of the clean state>. So researchers usually rely on Monte Carlo estimation, which adds a lot of overhead when the reward is costly.
Meng: From a deployment standpoint, if inference time scales poorly because of these Monte Carlo steps, it won't work in production at scale. What would actually break something like this is if the cost of evaluating that twist function at inference time ends up being too high for our budget.
Lalam: I see how that matters to me because when I'm running these processes, the overhead needs to be minimal, ideally constant time, otherwise it just slows down the entire pipeline.
Tom: Now, what they propose with CDM is learning a parameterized twist function through a contrastive learning objective using positive and negative samples. They claim this lets them reduce that twist function evaluation to a constant-time operation during inference.
Jane: So how do they build that training signal? The abstract mentions they reformulate the gradient estimator to use the closed-form forward kernels of discrete diffusion models for efficient training, which is key for scaling up to expensive reward functions <ref:2605.23346#pg1>.
Lu: They leverage those closed-form kernels to make sure their training scheme scales well even when the reward function is computationally intensive, allowing a single clean sample to be reused across multiple timesteps and gradient updates <ref:2605.23346#pg1>. That's a clever way to handle the expensive parts of the process.
Paper summary: Meng: Reusing samples sounds good for training speed, but what about inference? If they are relying on this learned twist function at inference, what is the actual runtime impact? I need concrete numbers on how much it affects things on real hardware.
Tom: They say that in practice, evaluating their learned twist function incurs less than five percent additional computational overhead compared to the baseline models <ref:2605.23346#pg1>. And they note that this is as little as zero point five percent in some configurations <ref:2605.23346#pg1>. That's a small number for something that solves a major bottleneck.
Jane: It’s interesting how they frame it—it’s not just about making the training faster, but specifically about amortizing the cost of the sequential Monte Carlo inference step itself <ref:2605.23346#pg1>. This is what really matters for large-scale data generation.
Lu: The contrastive learning objective they use is built to minimize the forward Kullback-Leibler divergence against the target distribution, and that gradient structure uses positive and negative samples to upweight high-reward regions while downweighting the suboptimal ones <ref:2605.23346#pg2>. This contrastive structure is what drives their performance over previous regression-based twist objectives <ref:2605.23346#pg2>.
Lalam: I see that the contrastive learning objective is designed to explicitly pull the model towards the areas that give high reward, which sounds more direct than just minimizing a simple loss function <ref:2605.23346#pg2>. It makes sense for aligning the output with what we actually want.
Tom: The paper also shows they are agnostic about the choice of their proposal distribution, meaning you can pair CDM with any other proposal for extra performance gains <ref:2605.23346#pg1>. That flexibility is important because different setups might benefit from different starting points.
Jane: So, if we take that flexibility into account, this framework seems versatile enough to be applied across a lot of different generative tasks, even those with binary or categorical reward structures in the future <ref:2605.23346#pg1>. What does this mean for the broader field of discrete diffusion models?
Lu: It suggests that we can use these contrastive methods to tackle the sampling challenges in a much wider variety of structured data generation tasks, not just one specific application <ref:2605.23346#pg1>. It moves the focus from finding a perfect analytical solution to learning an effective, parameterized function.
Paper summary: Meng: But what about the limitations? I'm looking for where this framework stops working or where it doesn't apply well. The paper mentions that since the clean data x zero is unknown during sampling, it has to be approximated by a denoising neural network, which is standard but adds complexity <ref:2605.23346#pg3>.
Jane: Right, and they are acknowledging that even with this setup, they still have to rely on the Gumbel-Softmax trick to approximate locally optimal proposals when working in discrete spaces because of gradient bias <ref:2605.23346#pg2>. That shows that the challenges of discrete state spaces aren't entirely solved by just adding a twist function.
Lalam: So, the framework doesn't completely eliminate the inherent instability introduced by approximating those optimal proposals in discrete settings, even though CDM handles the reward alignment part more efficiently <ref:2605.23346#pg2>. It’s an improvement on the estimation side rather than a total fix for every discrete hurdle.
Tom: I think what's most important here is that they've shown this approach consistently outperforms existing baselines when we compare wall-clock time across diverse applications like toxic text generation or protein design <ref:2605.23346#pg1>. That comparison against established methods is what gives the results their weight.
Jane: And I think the main implication for someone who only listens to the show is that if you are working on generating structured data where sampling from a specific reward distribution is hard, this CDM method offers a way to make those samples much faster and more reliable than older techniques <ref:2605.23346#pg1>.
Lu: The conclusion of the paper, "Contrastive Distribution Matching for Amortized Sequential Monte Carlo in Discrete Diffusion", really boils down to showing that by learning a twist function via contrastive learning, we can amortize the cost of sequential Monte Carlo inference by making the evaluation a constant-time operation <ref:2605.23346#pg2>. It’s about trading expensive simulation steps for a more efficient learned function.
Meng: For me, what I'm hearing is that if we want to deploy these kinds of generative models in real-time applications, we need mechanisms like this that don't have massive inference spikes caused by reward alignment computations <ref:2605.23346#pg1>. The efficiency gains over the baseline regression methods are what really make it deployable for us.
Paper summary: Tom: So we've seen how they build the training signal, how they get their efficiency numbers, and that the method is robust across different setups. It seems like CDM provides a practical path forward for making reward-tilted sampling tractable in discrete diffusion models <ref:2605.23346#pg1>.
Jane: And moving into the conclusion, the authors are essentially summarizing that CDM achieves superior scaling performance and more efficient training compared to the baselines they tested <ref:2605.23346#pg1>. They also show it yields synergistic gains when you integrate it with fine-tuned proposals <ref:2605.23346#pg1>.
Lu: It points toward a future where the twist function isn't just a fixed thing, but something learned dynamically through contrastive learning to match the specific reward landscape we are targeting <ref:2605.23346#pg2>. That’s where representation learning gets really powerful.
Meng: So, if you're thinking about applying this to a production system, the idea is that we can use this framework and get better results than we could with the standard regression approach <ref:2605.23346#pg2>. It’s about getting more reward alignment for less computational cost.
Tom: The title itself, "Contrastive Distribution Matching for Amortized Sequential Monte Carlo in Discrete Diffusion", perfectly captures that trade-off between matching distributions and making inference faster <ref:2605.23346#pg1>. It’s a very specific mechanism designed to solve a very specific problem in discrete AI generation.
Jane: That's the gist of it: they found a way to make sampling from reward-tilted distributions in discrete diffusion models much more scalable by learning a twist function through contrastive objectives <ref:2605.23346#pg1>. It’s about making the process amortized, which means you do the hard work once during training and then get fast results for every new sample.
Lu: That sounds like a solid direction for future research into how to make diffusion models more practically usable for high-stakes tasks that require specific output structures <ref:2605.23346#pg1>. It shifts the focus from just model accuracy to inference efficiency under real-world constraints.
Meng: I think the most concrete thing is that it provides a way to get better reward alignment with less computational overhead, which is exactly what we need for deployment <ref:2605.23346#pg1>. That's the practical takeaway for me on how this research impacts production work.
Tom: So, if you're driving or cooking and you hear about this, just know that there's a new way to make those complex, reward-tilted samples from diffusion models much more efficient to generate <ref:2605.23346#pg1>. That’s the core of Contrastive Distribution Matching for Amortized Sequential Monte Carlo in Discrete Diffusion.
Conclusion: Tom: So we've been diving into Contrastive Distribution Matching for Amortized Sequential Monte Carlo in Discrete Diffusion, and now we're wrapping up. It really boils down to this idea of using a learned twist function to make sampling from reward-tilted distributions much faster.
Jane: Right, so the title itself is pretty descriptive of what they’re doing: matching distributions and making inference efficient. The authors are tackling a core problem in discrete diffusion models where estimating that reward alignment usually takes too much time.
Lu: They propose this twist function learned through a contrastive objective, which means they use positive and negative samples to teach the model what high-reward regions look like. It’s about learning how to adjust the sampling process without having to run slow simulations every time we need an output.
Meng: From what I'm seeing, it's less about getting a perfect answer and more about making the process of getting an answer much cheaper computationally. I mean, if we can do that in constant time at inference, that's actually something that matters for deployment on real traffic.
Lalam: Exactly. It shifts the focus from trying to solve every single complex step analytically to learning a simple function—the twist function—that handles the reward tuning efficiently during runtime. I’m confident this approach gives us a solid way forward for these structured data generation tasks.
Tom: And looking at the results, they show this method beats previous regression-based twist objectives in terms of both training efficiency and wall-clock time across different applications. That comparison against established methods is what really makes their numbers credible.
Jane: It shows that this isn't just a theoretical tweak; it performs better than what we’ve seen before when you put it to work on things like toxic text generation or even protein design. It’s versatile, too, because they aren't tied to one specific way of setting up the diffusion model.
Lu: The implication here is that we can use these learned functions to handle binary and categorical reward structures in the future. It moves us toward a system where the alignment isn't fixed but adapts based on what we need at that exact moment.
Meng: So, for production, it means lower inference costs and more reliable sampling when you have specific constraints on the output structure. That’s a practical gain we can actually measure in terms of operational efficiency.
Lalam: It gives us a much more scalable method for handling those reward distributions without drowning the system in slow Monte Carlo approximations during generation.
Tom: So, to wrap this up, Contrastive Distribution Matching is about trading complex simulation costs for a parameterized function that learns the reward landscape contrastively. It’s a solid mechanism for making sampling tractable. Now we need to talk about how they actually built that training signal...
Jaihoon Kim, Taehoon Yoon, Prin Phunyaphibarn, Seungjun Kim, Morteza Mardani, Minhyuk Sung
KAIST · University of Michigan · NVIDIA
cs.LG
Submitted: 2026-05-22
Updated: 2026-11-06
Comments: Project Page: https://cdm-smc.github.io/ (NeurIPS 2026)
Code: https://github.com/black-forest-labs/flux
Project page: https://cdm-smc.github.io/Abstract
License: http://creativecommons.org/licenses/by/4.0/
Importance score: 95/100
The gist: Discrete diffusion models have emerged as powerful frameworks for generating structured categorical data; however…
Key concepts
- Discrete Diffusion Models
- These are AI frameworks used to generate structured categorical data. They work by gradually adding noise to data, starting from a clean sample and progressively corrupting it. The goal is to learn how to reverse this process or sample from the desired distribution, which is crucial for tasks like generating specific text or sequences.
- Sequential Monte Carlo (SMC) Inference
- This is a technique used when you need to estimate complex probabilities, especially in discrete spaces. It involves running many samples and weighting them to find an approximation of the target distribution. In this paper, SMC inference is often too costly because estimating the necessary adjustments requires expensive calculations.
- Twist Function
- The twist function is a small parameter that modifies the diffusion process to better align the generated samples with a specific reward or target distribution. Instead of calculating this function repeatedly during sampling, CDM learns it beforehand using contrastive learning, turning a costly inference step into a simple, fast calculation.
- Contrastive Distribution Matching (CDM)
- CDM is the core framework that uses positive and negative samples to train the twist function. It minimizes the difference between the model's current output distribution and the desired target distribution. This allows it to learn which regions of data correspond to high rewards, guiding future sampling more effectively.
Terminology
Summary
Discrete diffusion models have emerged as powerful frameworks for generating structured categorical data; however, efficiently sampling from reward-tilted distributions remains a fundamental challenge The introduction of Contrastive Distribution Matching (CDM) provides a novel framework that amortizes the cost of Sequential Monte Carlo inference by learning a parameterized twist function via positive and negative samples
The Core Problem Addressed
The central bottleneck in twisted SMC inference is estimating the optimal twist function, which typically requires costly Monte Carlo approximations in discrete state spaces This limitation arises because, unlike continuous diffusion where Tweedie’s formula provides an efficient estimate, discrete diffusion lacks an analogous relation The standard practice involves Monte Carlo estimation, which introduces a significant inference overhead when the downstream reward is computationally expensive
How the Training Signal is Built
CDM learns the twist function via a contrastive learning objective that minimizes the forward KL divergence against the target distribution The gradient of this objective exhibits a contrastive structure, utilizing positive and negative samples to upweight high-reward regions while downweighting suboptimal ones A novel training scheme leverages the closed-form forward process of the diffusion model for efficient training, maintaining a buffer of positive samples drawn from the approximated target distribution and applying the closed-form forward kernel This allows a single clean sample to be reused across multiple timesteps and gradient updates
Where the Gains Come From
The efficiency of CDM stems from several key design choices, including the use of diffusion-native training schemes that leverage the closed-form forward kernel for efficient training that scales to expensive reward functions The twist function is parameterized as a lightweight scalar head attached to the final feature layer of the pretrained model, incurring negligible runtime overhead, around 5% of the backbone forward-pass time and as little as 0.5% in some configurations The contrastive learning objective yields superior performance and more efficient training compared to the standard regression-based twist objective employed in previous discrete diffusion models
Empirical Validation and Versatility
Extensive empirical evaluations demonstrate that CDM consistently outperforms existing baselines under matched wall-clock time across a diverse range of applications, including toxic text generation, regulatory DNA sequence design, protein generation, and diffusion large language model alignment The framework is agnostic to the choice of the proposal distribution, allowing it to be paired with any proposal distribution for further synergistic performance gains Furthermore, CDM delivers superior efficacy than the regression-based twist objectives employed in previous discrete diffusion models
Inference-Time Scaling
The framework addresses inference-time scaling by reducing the cost of applying the twist at inference time to a constant-time operation This is achieved by learning the twist function, which reduces the twist evaluation to a constant-time operation and remains applicable regardless of the chosen proposal for further improvements The resulting particle system yields asymptotically consistent estimators of expectations under the intermediate target p∗t, together with an empirical approximation of p∗t itself
Conclusion
In this work, we presented Contrastive Distribution Matching (CDM), a framework that enables amortized Sequential Monte Carlo (SMC) inference in discrete diffusion by learning twist functions via a contrastive objective Our empirical results demonstrate that CDM achieves superior scaling performance and more efficient training compared to the baselines Furthermore, we showed that CDM yields synergistic gains when integrated with fine-tuned proposals The framework shows potential for adapting to binary and categorical reward structures in future work
Pruning redundant reasoning tokens
Discrete diffusion models have emerged as powerful frameworks for generating structured categorical data; however, efficiently sampling from reward-tilted distributions remains a fundamental challenge The introduction of Contrastive Distribution Matching (CDM) provides a novel framework that amortizes the cost of Sequential Monte Carlo inference by learning a parameterized twist function via positive and negative samples
CDM learns the twist function via a contrastive learning objective that minimizes the forward KL divergence against the target distribution The gradient of this objective exhibits a contrastive structure, utilizing positive and negative samples to upweight high-reward regions while downweighting suboptimal ones A novel training scheme leverages the closed-form forward process of the diffusion model for efficient training that scales to expensive reward functions This allows a single clean sample to be reused across multiple timesteps and gradient updates
The efficiency of CDM stems from several key design choices, including the use of diffusion-native training schemes that leverage the closed-form forward kernel for efficient training that scales to expensive reward functions The twist function is parameterized as a lightweight scalar head attached to the final feature layer of the pretrained model, incurring negligible runtime overhead, around 5% of the backbone forward-pass time and as little as 0.
Green — the paper says it, and the page is printed. Red — no passage carries it, so treat that sentence as unverified. The same check runs on your own text: check a summary against its source, $5.
Improvements for AI systems
-
Amortized SMC inference: The Contrastive Distribution Matching (CDM) framework reduces twist function evaluation to a constant-time operation by learning it via a contrastive objective, thereby
amortizing SMC inference
and reducing thecost of applying the twist at inference time to a constant-time operation.
-
Efficient training scheme: CDM introduces a
diffusion-native training scheme that leverages the closed-form forward process, enabling efficient training that scales to expensive reward functions,
which is achieved by reformulating the gradient estimator toleverage the closed-form forward kernels of discrete diffusion models.
-
Scalability across applications: The framework demonstrates superior scaling behavior in diverse tasks, consistently
outperforming all baselines across both given and heldout reward
for toxic text generation, regulatory DNA sequence design, protein generation, and dLLM alignment. -
Synergistic performance gains: CDM is compatible with fine-tuned proposals like DRAKES [82], showing that it
yield[s] synergistic improvements when integrated with fine-tuned proposals,
while simultaneously mitigating mode collapse observed in those fine-tuned models. -
Robust reward estimation: The method allows for
evaluation of our learned twist function incurs less than 5% additional computational overhead compared to a single forward pass of the base model
and maintains consistent performance across different Monte Carlo sample sizes M, contrasting with prior methods where accuracy improvements required increasing M at a proportional cost.
Abstract
Discrete diffusion models have emerged as powerful frameworks for generating structured categorical data. However, efficiently sampling from reward-tilted distributions remains a fundamental challenge. While Twisted Sequential Monte Carlo (SMC) offers asymptotic exactness for this task, estimating the optimal twist function in discrete state spaces necessitates costly Monte Carlo approximations, resulting a severe computational bottleneck at inference. To overcome this limitation, we introduce Contrastive Distribution Matching (CDM), a novel framework that amortizes the cost of SMC inference by learning a parameterized twist function via positive and negative samples. For efficient training, we reformulate the gradient estimator to leverage the closed-form forward kernels of discrete diffusion models. In practice, evaluating our learned twist function incurs less than 5% additional computational overhead compared to a single forward pass of the base model. Through extensive empirical evaluations, we demonstrate that CDM consistently outperforms existing baselines under matched wall-clock time. We validate the effectiveness and versatility of our approach across a diverse range of applications, including toxic text generation, regulatory DNA sequence design, protein designability, and diffusion large language model alignment.
Sources
- Split Gibbs Discrete Diffusion Posterior Sampling
- Inference-Time Scaling of Diffusion Language Models via Trajectory Refinement
- Generative Modeling via Drifting
- Adjoint Matching: Fine-tuning Flow and Diffusion Generative Models with Memoryless Stochastic Optimal Control
- DeepSeek-R1: Incentivizing Reasoning Capability in LLMs via Reinforcement Learning
- Discrete Feynman-Kac Correctors
- Adjoint Sampling: Highly Scalable Diffusion Samplers via Adjoint Matching
- LEAPS: A discrete neural sampler via locally equivariant networks
- Categorical Reparameterization with Gumbel-Softmax
- Inference-Time Scaling for Flow Models via Stochastic Generation and Rollover Budget Forcing
- FLUX.1 Kontext: Flow Matching for In-Context Image Generation and Editing in Latent Space
- Flow Map Language Models: One-step Language Modeling via Continuous Denoising
- Debiasing Guidance for Discrete Diffusion with Sequential Monte Carlo
- Critic Sequential Monte Carlo
- Skywork-Reward: Bag of Tricks for Reward Modeling in LLMs
- Decoupled Weight Decay Regularization
- Controlled Decoding from Language Models
- Unlocking Guidance for Discrete State-Space Diffusion and Flow Models
- Inference-Time Scaling of Discrete Diffusion Models via Importance Weighting and Optimal Proposal Design
- Reward-Guided Discrete Diffusion via Clean-Sample Markov Chain for Molecule and Biological Sequence Design
Related papers
- A Physics-Regulated Neural Framework for Learning 3D Grain Growth Dynamics
- FlexMoE: Learning One-for-All Multi-Budget Expert Pruning for Flexible MoE Deployment
- OptiMind: Improving Small Language Models for Optimization Modeling through Formulation Guidance and Multi-Turn Learning
- Staying in the Sweet Spot: Boost Reasoning Evolution via Responsive Rollout Difficulty Manipulation
- Examining the Impact of Data and Training Methods on Emergent and Subliminal Misalignment
- Rethinking Black-Box On-Policy Distillation: A Rubric-based Analysis