Build with Hedra

Talk to an engineerRead the research

Research expertise and GPU infrastructure for models you own.

© 2026 Hedra

Products

  • Inference
  • Model training
  • GPU Sandboxes
  • World Model Runtime
  • Pricing

Deployment

  • Deployment options
  • Sovereign deployment

Model Factory

  • Overview
  • Forward Deployed Researchers
  • Enterprise

Research

  • Overview
  • Recursive self-improvement
  • Mechanistic interpretability
  • Post-training
  • Inference
  • Blog

Applications

  • Physical AI
  • Defense
  • Media generation

Developers

  • Platform onboarding
  • SDKs & integrations
  • Documentation
  • Model APIs
  • MCP & CLI
  • Changelog

Company

  • About Hedra
  • Manifesto
  • Careers
  • Contact
  • Play
  • Play pricing
  • Play help
  • Status
  • LinkedIn
  • X
  • GitHub
  • Discord
  • Privacy Policy
  • Terms of Use
  • Acceptable Use
  • Cookie Policy
  • Biometric Data Policy
  • Security
Hedra
  • Pricing
Docs
Get started
Hedra
Get started
All posts

mINT: The First Mechanistic Interpretability Framework for World Models

Haiyang Liu, Sean Sun·November 18, 2026
Two mechanical reels with robotic arms, one white and one coral, against a dark teal background.

TL;DR: Hedra’s mINT (mechanistic interpretability) framework uses an understanding of world-model internals to guide inference-time steering. By locating the layers and attention heads that drive a specific behavior, we improve frozen MiniMax H3’s strict success rate on multi-machine instructions from 39/256 (15.2%) to 199/256 (77.7%), without updating model weights.


Original and steered MiniMax H3 examples with head 38.48 attention overlays for valve arms, Mars rovers and airlock hatches.
Figure 1. Original MiniMax H3 (top) and Hedra’s steered model (bottom), with measured cross-attention from head 38.48 in three multi-machine scenes.

World models predict how a physical scene changes over time. In Physical AI, teams use world models to simulate robot actions, multi-object interactions and physical outcomes. Yet precise instructions still break down: ask one robot arm to turn a valve while an identical arm stays still, and both may move. The standard fix is post-training across the full model. Online reinforcement learning (RL) works, but a single capability can take hundreds to over a thousand GPU-hours. Before committing to that training, we should identify the internal mechanisms that control the behavior.

Hedra’s mINT framework turns that understanding into control over world models. Mechanistic interpretability identifies which internal layers and attention heads drive a behavior. Our framework locates those components, tests their causal role and uses them to change the model’s output. This gives us training-free interventions at inference time. We’re making this process reusable for AI research agents, so mechanistic understanding can guide the next experiment.

MiniMax H3 moves machines it was instructed to keep still

MiniMax H3 failure examples with Mars rovers and robot arms. The non-target machine moves despite the stay-still instruction.
Figure 2. Failure examples from MiniMax H3: Mars rovers holding sample tubes (left) and blue robot arms on valve wheels (right). Red prompt text states the stay-still constraint; the detection box identifies the non-target machine.

We test the framework on MiniMax H3's ability to control identical machines independently. Across 256 clips in which only one machine was instructed to act, the base model succeeded in 39 cases (15.2%). The failures in Figure 2 show the non-target machine moving despite the stay-still instruction.

Success requires the target machine to complete its action while the others remain still.

The mINT framework starts with a behavior we can score. We search for internal components that change that behavior when we intervene on them, then use control experiments to test their causal contribution. Those findings guide how we steer the frozen model, with the correction evaluated against the original success criterion. A reusable agent skill packages the localization and steering procedure.

From 2,800 attention heads to three

Animation of the 50-block search, randomized subsets, and fitted coefficients: head 37.2 (0.32), head 37.26 (0.23), and head 38.48 (0.14).
Figure 3. Internal-model visualization accompanying the localization procedure. The block search narrows 50 transformer blocks to blocks 35 to 39; subset probing then tests the 280 heads in that band. The fitted coefficients are 0.32 for head 37.2, 0.23 for head 37.26, and 0.14 for head 38.48.

H3 generates video through repeated denoising steps. Each step passes activations through 50 transformer blocks with 56 attention heads each, for 2,800 heads in total. We begin with contiguous groups of blocks, then examine individual heads within the responsive group.

Early denoising steps establish coarse spatial structure and motion trajectories; later steps refine visual detail. We therefore restrict the search to the first two denoising steps, when an intervention can change the scene's emerging motion.

The stay-still intervention isolates blocks 35 to 39

We first increase video-to-text attention on the prompt's stay-still clause. If a group of blocks contributes to suppressing motion in the non-target machines, amplifying that clause should change the generated trajectory.

Let KstillK_{\mathrm{still}}Kstill​ be the tokens in the stay-still clause, identified in each prompt by a vision-language model (VLM), as shown in Figure 2. Let pi=∑j∈KstillAi,jp_i = \sum_{j\in K_{\mathrm{still}}} A_{i,j}pi​=∑j∈Kstill​​Ai,j​ be the attention mass that future-frame video token iii assigns to them. Adding a positive bias β\betaβ to those tokens' pre-softmax attention logits changes the head output as follows:

oi′=oi+(eβ−1)∑j∈KstillAi,jvj1+(eβ−1)pio'_i = \frac{o_i + (e^\beta - 1)\sum_{j \in K_{\mathrm{still}}} A_{i,j}v_j}{1 + (e^\beta - 1)p_i}oi′​=1+(eβ−1)pi​oi​+(eβ−1)∑j∈Kstill​​Ai,j​vj​​

Here, oio_ioi​ is the original head output, and vjv_jvj​ is the value vector of token jjj. Heads that assign almost no attention to that clause change little; heads that already attend to it receive a stronger intervention.

For each decoded clip, we measure frame-to-frame pixel changes within the machines' bounding boxes. We check whether the non-target machine falls below the stillness threshold while the target completes its action.

With k=2k = 2k=2 denoising steps and β=3\beta = 3β=3, scanning contiguous five-block groups across all 50 blocks identifies blocks 35 to 39 as the only responsive band. Adjacent blocks, including 30 to 34, do not suppress the unwanted motion. Intervening on each of blocks 35 to 39 separately resolves only some failures, suggesting that multiple heads in the band contribute together.

Subset probing identifies contributions that single-head tests miss

An individual head can help suppress motion without pushing the non-target machine below the stillness threshold on its own. Single-head tests can therefore miss heads that contribute when combined with others. We activate a random half of the 280 candidate heads in each trial. This lets us estimate marginal contributions across all candidate heads in parallel, testing each head with different partners without exhaustively evaluating every combination.

Evaluating random subsets does not require finishing the remaining denoising steps or running the 3D VAE video decoder. We stop each probe run after the second denoising step and score its predicted clean video latent. Under the flow-matching convention used here, the velocity prediction gives the estimate x^0(k)=xk+σkvθ(xk,σk)\hat{x}_0^{(k)} = x_k + \sigma_k v_\theta(x_k,\sigma_k)x^0(k)​=xk​+σk​vθ​(xk​,σk​). We compare H3’s predicted clean latent for future frames with H3’s conditioning-frame latent inside the non-target machine's bounding box:

m(S)=1∣T∣ ∣Bstill∣∑t∈T∑u∈Bstill∥x^0(k)[t,u]−xcond[u]∥1m(S) = \frac{1}{|\mathcal{T}|\,|\mathcal{B}_{\mathrm{still}}|} \sum_{t\in\mathcal{T}} \sum_{u\in\mathcal{B}_{\mathrm{still}}} \left\| \hat{x}_0^{(k)}[t,u] - x_{\mathrm{cond}}[u] \right\|_1m(S)=∣T∣∣Bstill​∣1​t∈T∑​u∈Bstill​∑​​x^0(k)​[t,u]−xcond​[u]​1​

Here, xcondx_{\mathrm{cond}}xcond​ is the conditioning-frame latent, T\mathcal{T}T indexes future latent frames and Bstill\mathcal{B}_{\mathrm{still}}Bstill​ covers the non-target machine. Lower drift indicates less unwanted motion. On 112 full validation videos, the early-latent ranking agreed 99% of the time with whether the non-target machine moved in the final decoded video. Using this proxy accelerated subset evaluation by more than 3×: stopping after the first two steps avoids the remaining denoising steps and the 3D variational autoencoder (VAE) video decoder.

For each subset, we normalize mean drift between the all-candidate intervention and the base model. We then fit a linear model across the candidate heads, using each subset's head-inclusion indicators to predict normalized drift. Forward stepwise least squares accounts for the selected heads before testing whether another head explains the remaining variation. Scores and head-inclusion indicators are centered within each batch. We stop when no remaining head clears the 95th-percentile threshold across random permutations of the residual:

ym=m‾(Sm)−m‾(H)m‾(∅)−m‾(H),y~≈−∑h∈AβhX~:,h\begin{aligned} y_m &= \frac{\overline{m}(S_m) - \overline{m}(\mathcal{H})}{\overline{m}(\varnothing)-\overline{m}(\mathcal{H})}, \\[0.5em] \tilde{y} &\approx -\sum_{h\in\mathcal{A}}\beta_h\tilde{X}_{:,h} \end{aligned}ym​y~​​=m(∅)−m(H)m(Sm​)−m(H)​,≈−h∈A∑​βh​X~:,h​​

The candidate pool is H\mathcal{H}H, each sampled subset is SmS_mSm​, and A\mathcal{A}A is the set of heads selected by the fit. The indicator Xm,hX_{m,h}Xm,h​ equals 1 when head hhh is in subset SmS_mSm​. A normalized score of 0 corresponds to intervening on all candidate heads; 1 corresponds to the base model. A positive fitted coefficient βh\beta_hβh​ means that activating head hhh predicts lower drift after accounting for the other selected heads.

We test 108 random half-subsets across the 280 candidate heads and use 200 random permutations for the noise test. The linear fit selects three heads above the noise floor:

Attention head

Fitted coefficient, βh\beta_hβh​

Block 37, head 2

0.32

Block 37, head 26

0.23

Block 38, head 48

0.14

Subsets that exclude all three heads remain close to the base model, with normalized drift of 0.88, even though they amplify roughly half the candidate pool. Subsets containing all three reach 0.16, recovering nearly the full candidate-pool effect.

The block intervention narrows 2,800 heads to 280 candidates, early-latent scoring skips the remaining denoising steps and video decoding, and randomized subsets and the linear fit identify contributions across head combinations. Together, they turn an internal-model analysis into a repeatable localization procedure.

Three heads steer MiniMax H3 without retraining

Inspecting the three heads reveals why H3 fails to keep the non-target machines still: it does not attend strongly enough to the text prompt. The identical machines begin in poses associated with action, and video-to-video self-attention favors motion across those visually similar objects. The heads associate relevant prompt words with the corresponding machines: in head 38.48, attention to the stay-still tokens is 10 times stronger on non-target machines than on the target. Yet video tokens allocate only 2.4% to 7.2% of their total attention mass to the prompt.

Our interpretation is that the stay-still signal is correctly grounded but too weak to overcome the visual motion pattern. This diagnosis motivates an adaptive intervention that increases prompt attention and scales the output of the three located heads.

The inference-time steering rule uses the prompt as a whole. Let KpromptK_{\mathrm{prompt}}Kprompt​ contain all prompt tokens and ph,i=∑j∈KpromptAh,i,jp_{h,i} = \sum_{j\in K_{\mathrm{prompt}}} A_{h,i,j}ph,i​=∑j∈Kprompt​​Ah,i,j​ be the prompt-attention share for video token iii in head hhh. During the first two denoising steps, we set p~h,i=max⁡(ph,i,αh)\tilde{p}_{h,i} = \max(p_{h,i},\alpha_h)p~​h,i​=max(ph,i​,αh​), then scale the head's output by γh\gamma_hγh​. The rule preserves relative attention weights within the prompt and within the non-prompt tokens:

oh,i′=γh(p~h,i oprompt,h,i+(1−p~h,i)oh,i−ph,i oprompt,h,i1−ph,i),p~h,i=max⁡(ph,i,αh)\begin{aligned} o'_{h,i} &= \gamma_h \left(\tilde{p}_{h,i}\,o_{\mathrm{prompt},h,i} + (1-\tilde{p}_{h,i})\frac{o_{h,i}-p_{h,i}\,o_{\mathrm{prompt},h,i}}{1-p_{h,i}}\right),\\[0.5em] \tilde{p}_{h,i} &= \max(p_{h,i}, \alpha_h) \end{aligned}oh,i′​p~​h,i​​=γh​(p~​h,i​oprompt,h,i​+(1−p~​h,i​)1−ph,i​oh,i​−ph,i​oprompt,h,i​​),=max(ph,i​,αh​)​

In this expression, oprompt,h,io_{\mathrm{prompt},h,i}oprompt,h,i​ is the attention-weighted value average within the prompt. The floor changes the balance between prompt and non-prompt contributions; γh\gamma_hγh​ controls how strongly the resulting head output is written into the residual stream.

Tokens whose prompt attention already meets the floor pass through the clamp unchanged before output scaling. Unlike the localization probe, the final rule does not need stay-still token lists, spatial masks or prompt parsing. The localization and evaluation stages still use the clause and bounding boxes described above.

We fit the intervention parameters on 64 calibration videos, obtaining αh=0.9\alpha_h = 0.9αh​=0.9 for all three heads and γh=(1,2,2)\gamma_h = (1,2,2)γh​=(1,2,2) for heads 37.2, 37.26 and 38.48. We then freeze those parameters for evaluation across 256 videos spanning 16 unseen machine types.

Evaluation criterion

Base MiniMax H3

Steered MiniMax H3

Strict selective-motion success

39/256 (15.2%)

199/256 (77.7%)

Success allowing slight non-target motion

73/256 (28.5%)

226/256 (88.3%)

For MiniMax H3, strict selective-motion success rises from 15.2% to 77.7%, an increase of 62.5 percentage points. The weights remain frozen throughout. Head discovery and calibration identify the intervention; the inference-time rule applies it to new scenes.

Steering three heads outperforms prompt engineering in MiniMax H3

4A. Prompt changes
58
Three-head edit
Original prompt
17
“remains motionless”
22
“keeps its position”
29
“never moves”
21
No stay-still clause
8
Resting state only
18
Stay-still twice
17
Extra emphasis
25
Extra held-object
7
Still before action
7
Prompt before scene
6
Three-head edit
58
016324864
Successful clips out of 64
Chart data
4A. Prompt changes
Prompt conditionSuccessful clips out of 64
Original prompt17
“remains motionless”22
“keeps its position”29
“never moves”21
No stay-still clause8
Resting state only18
Stay-still twice17
Extra emphasis25
Extra held-object7
Still before action7
Prompt before scene6
Three-head edit58
4B. Other layers and heads
58
Heads 37.2, 37.26, 38.48
No edit (base)
17
Layers 0 to 9
16
Layers 10 to 19
16
Layers 20 to 49 (1)
29
Layers 20 to 49 (2)
18
Layers 35 to 36, 39 to 40
19
Layers 37 to 38 (1)
18
Layers 37 to 38 (2)
7
Layers 37 to 38 (3)
21
Top still-word heads
19
Top prompt heads
14
Heads 37.2, 37.26, 38.48
58
016324864
Successful clips out of 64
Chart data
4B. Other layers and heads
Attention-head conditionSuccessful clips out of 64
No edit (base)17
Layers 0 to 916
Layers 10 to 1916
Layers 20 to 49 (1)29
Layers 20 to 49 (2)18
Layers 35 to 36, 39 to 4019
Layers 37 to 38 (1)18
Layers 37 to 38 (2)7
Layers 37 to 38 (3)21
Top still-word heads19
Top prompt heads14
Heads 37.2, 37.26, 38.4858
4C. Position and scale
Base H3Steered H3
Mirrored
Base H3 13
Steered H3 51
Shifted
Base H3 21
Steered H3 54
Scaled up
Base H3 10
Steered H3 56
Scaled down
Base H3 42
Steered H3 58
Smaller + shifted
Base H3 46
Steered H3 55
016324864
Successful clips out of 64
Chart data
4C. Position and scale
Layout conditionBase H3Steered H3
Mirrored1351
Shifted2154
Scaled up1056
Scaled down4258
Smaller + shifted4655
Figure 4. Ablations with 64 clips per bar. Prompt changes and control head triples do not match the selected three-head intervention. The same steering parameters also improve results across changes in object position and scale.

We run three sets of ablations on a 64-clip subset. For evaluation, we combine sampled keyframes, enlarged first/middle/last crops of stationary machines, pixel-difference maps and per-frame box-motion curves into a composite sheet for a model judge. Our evaluation used Opus 5.5 and matched human inspection on 64 of 64 checked clips.

Prompt changes. Across ten variations, including stronger stay-still wording and resting-state descriptions, none produces more than 29 successful clips out of 64. Moving instructions before the scene description reduces success to 6/64. With the original prompt, the three-head intervention improves success from 17/64 to 58/64.

Head controls. Applying the same steering parameters to ten alternative head triples, including heads with the highest raw prompt attention, produces 7 to 29 successful clips out of 64, compared with 58/64 for the selected heads. Blocking the selected heads from reading the prompt reduces success from 17/64 to 0/64; blocking random triples leaves it unchanged.

Position and scale. The intervention improves success across mirrored, shifted and rescaled layouts (Figure 4C). With the images held fixed and the instruction switched to the other machine, base H3 succeeded on 33 of 64 clips and steered H3 on 58.

The selected heads give us a causal control over this failure: strengthening their prompt signal improves selective motion, while blocking it eliminates successful clips in the ablation set. The highest-attention controls also show why an attention map alone is insufficient. We have to test which heads change the output. The position and scale tests show that the resulting intervention survives changes to the scene layout.

Autonomous research needs mechanistic diagnosis

To avoid repeating manual exploration for every new failure mode, we packaged the mINT framework’s localization and steering procedure into a mechanistic interpretability skill: an instruction document for coding agents, with no hardcoded layer or head IDs, paired with a preregistered search script.

For the selective-motion failure above, the automated pipeline ran from scratch without a handwritten word list. It independently localized blocks 35 to 39 and recovered the primary head found in our investigation. This run tested whether the localization procedure could be repeated automatically.

We separately tested whether the skill helps coding agents make appropriate research decisions. Seven design problems included choosing a localization approach for an ignored instruction versus unwanted subtitles, and identifying stable heads across subset batches. Across four agent models (Fable, Opus, Sonnet and Haiku), average success rose from 75.0% to 97.4% with the skill.

Agent performance on research-design problems
92 to 100%
Success rate with the skill
Without skillWith skill
Fable
Without skill 82%
With skill 100%
Opus
Without skill 72%
With skill 100%
Sonnet
Without skill 85%
With skill 97%
Haiku
Without skill 62%
With skill 92%
020406080100
Success rate (%)
Chart data
Agent performance on research-design problems
Agent modelWithout skillWith skill
Fable82%100%
Opus72%100%
Sonnet85%97%
Haiku62%92%
Figure 5. Success rates on seven mechanistic interpretability research-design problems with and without the skill.

The skill captures how to choose an intervention suited to a failure, estimate head contributions across combinations and validate candidate heads against permutation and random-head controls. This gives coding agents a reusable procedure for investigating the next world-model failure.

Steering gravity and contact inside MiniMax H3

Four held-out physics scenes comparing base and steered MiniMax H3, including a blue ball released over kinetic sand, with activation profiles across 50 blocks.
Figure 6. Qualitative comparisons on four held-out scenes, with base MiniMax H3 above and the steered model below. The fourth scene shows a blue ball released over kinetic sand. The 50-block activation-difference spectrum identifies blocks 30 to 49 as the intervention region.

We also explored release-and-drop and collision scenes. The prompts specify normal gravity and rigid surfaces, but the base model can produce sluggish falls or incorrect contact.

We run frozen H3 on three reference scenes with contrasting environment outcomes and average the activation differences for each of the 50 blocks. This analysis isolates a physical-environment direction in blocks 30 to 49. At inference time, we apply this direction to blocks 30 to 49 without updating the model's weights. In the illustrated held-out scenes, the steered rollouts show a paper wad dropping into water, a basketball bouncing from a crate edge onto a table, a mug dropping onto a stone block, and a grabber-held blue ball dropping onto kinetic sand.

The intervention changes with the failure. Selective motion calls for strengthening three attention heads; these gravity and contact examples use an activation direction across a band of blocks. The common step is to locate an internal representation relevant to the failure and test whether changing it improves the rollout. This is how mechanistic analysis can guide a choice of intervention across different behaviors.

Mechanistic diagnosis should come before another training run

Our self-critiquing video systems research used internal signals in a verifier to identify candidate failures for human review. Here, we intervene inside the video generator itself and measure the change in its output.

At Hedra, we believe mechanistic diagnosis should come before another expensive training run. A failed output alone does not identify which internal components need to change. In selective motion, locating and steering three attention heads improved instruction following without updating model weights. The mINT framework gives us a concrete starting point for model improvement: an intervention we can test and refine.

Our ambition is for AI research agents to make these decisions through causal experiments: diagnose a failure, locate the internal components involved, test an intervention and carry what they learn into the next experiment.

Build with Hedra Model Factory

Hedra Model Factory brings together forward deployed researchers, agents, GPU compute and optimized inference to help solve your problem with a multimodal model. For teams developing world models, we work with you to define the behavior you need, investigate failures, run post-training and evaluate results. That work can extend through data preparation and deployment, with production data informing the next version.

Bring us the problem you need to solve. We help you build, run and keep improving a model you own.

Talk to our team about your model and the behavior you need to change.

References

MiniMax H3: An Open Model Breaking the Boundaries Between Tasks and Modalities.

Building self-critiquing video systems, Hedra. Related research on verification, separate from the generator interventions reported here.

DiffusionNFT: Online RL with Forward Process.

Flow-GRPO: Training Flow Matching Models via Online RL.

RewardDance: Reward Scaling in Visual Generation.

Interpreting Physics in Video World Models. Physical-variable probing and attention ablations in video encoders.

Causal Physics Steering in Video World Models via Concept Activation Vectors. Weight-frozen interventions that shift VideoMAE's physical-plausibility judgments.

The Invisible Hand of Physics: When Video Diffusion Models Know More Than They Show. Physical representations in video diffusion models examined through recovered latent trajectories.

More from Hedra

Building self-critiquing video systems

November 12, 2026

Optimizing Krea 3 on AMD Instinct MI355X with AI researchers

November 2, 2026

The Hedra Developer API: models, SDKs, MCP and the console in one place

October 6, 2026
Read more from Hedra