Research

A signal-latent world model for aligned action.

Keplen’s research goal is a world model that treats signals, not only tokens, as the substrate for physical intelligence.

The model direction is multimodal, predictive, uncertainty-aware, and runtime-connected: it should learn shared latent structure across text/code, audio, visual/video, sensor streams, state traces, and future action evidence.

Why this matters

Physical AI needs more than instruction following.

It needs a model that can hold state, anticipate change, represent uncertainty, compare possible actions, and abstain when the evidence is not enough.

Research pillars

Four commitments.

These are research commitments, not product claims. They define how a Keplen model should behave before it is allowed to matter.

  1. 01

    Shared latent field

    A common representation for signals across modalities, with modality-specific encoders and decoders at the edges.

  2. 02

    Prediction before action

    The model should support forecasting, counterfactuals, and uncertainty before a policy commits to external effects.

  3. 03

    No-op as valid behavior

    Abstention, pause, escalation, and internal-only updates should be first-class outcomes, not failure states.

  4. 04

    Runtime-grounded alignment

    Model behavior should be evaluated through governed loops, traces, denials, approvals, state commits, and replayable evidence.

Signal scope

The signal-latent world model direction should include:

  • Text
  • Code
  • Audio
  • Visual / video
  • Sensor streams
  • State traces
  • Action proposals
  • Action outcomes
  • Human review
  • Runtime denials
  • Runtime approvals
  • Fleet telemetry

Alignment process

One loop, before broader autonomy is claimed.

Alignment is a process with explicit steps, not a property asserted after the fact. Each step produces evidence the next step depends on.

  1. 01

    Define the context

    What is the agent allowed to perceive, decide, delegate, and request?

  2. 02

    Register the evidence

    Datasets, model artifacts, training recipes, assumptions, licenses, leakage checks, and evaluation limits must be explicit.

  3. 03

    Train and compare

    No quality claim without baselines, ablations, failure cases, and clear parameter/data/compute accounting.

  4. 04

    Run in a governed loop

    The model proposes. Runtime boundaries verify. Actions produce trace-linked outcomes.

  5. 05

    Review failures

    Unsupported claims, contradictions, unsafe proposals, overbroad delegation, calibration errors, and no-op behavior are measured.

  6. 06

    Publish limits

    Every release should state what it is for, what it is not for, what failed, and what remains untested.

Evaluation topics

What we intend to measure.

  • Prediction quality
  • Uncertainty calibration
  • Counterfactual stability
  • Trace consistency
  • Action proposal safety
  • No-op correctness
  • Delegation boundaries
  • Runtime denial handling
  • Fleet coordination failures
  • State handoff failures
  • Human review outcomes

Collaborate on aligned physical AI.

Keplen is inviting research partners across world models, robotics, simulation, runtime systems, evaluation, datasets, and value alignment.