Scaling Inherently Interpretable Language Models
Guide Labs Team, Andreas Madsen, Aya Abdelsalam Ismail, Giang Nguyen, Isaac Plant, Muawiz Chaudhary, Nathaniel Monson, Saqib Azim, Zhichen Guo, Julius Adebayo
cs.CL, cs.AI
2026-08-06
Steerling-8B bakes interpretability into training: a 33,000-concept additive bottleneck gives exact attribution, it lands within ~10% of peers trained on far more compute, and gets easier to interpret as it scales.
Today's language models are trained as black boxes and explained after the fact: with probes, feature attribution, sparse autoencoders (SAEs), and chain-of-thought. These tools help, but they share a structural flaw: the model was never trained to make the explanation a reliable interface. A probe shows information is decodable, not that the model uses it. Gradient attribution measures local sensitivity, not what happens if you remove a feature. An SAE reconstructs activations but need not correspond to a real unit of the model's computation. A chain of thought may give a plausible reason unconnected to the computation that produced the answer. This paper (Guide Labs) challenges the premise that interpretability must tax capability.
The core idea is to make interpretability a constraint of the training pipeline, optimized alongside data, architecture, and the objective, rather than added later. It lands in a model called Steerling-8B, built on three pillars.
First, a concept library. A model that decomposes outputs by concept needs a large, human-understandable concept library, and none existed. The Atlas pipeline extracts hundreds of millions of free-form tags from millions of documents, canonicalizes them into over 33,000 concepts (across technology, medicine, law, and more), and trains an annotator that labels arbitrary text at chunk level, annotating over 1 trillion tokens in total.
Second, the architecture. Steerling-8B is a diffusion language model with a causal attention mask. Why not autoregressive? AR generation goes left to right one token at a time, concepts span multiple tokens, and there is no natural "no information here" baseline, which faithful input attribution needs. The diffusion masking objective fixes this: [MASK] is a "no information" state the model saw during training, giving attribution a valid baseline, and multiple tokens are predicted jointly per denoising step, a natural handle for concept-level control. The team's Causal Diffusion keeps block attention (bidirectional within a block, causal across blocks) but drops Block Diffusion's clean-plus-noisy copy concatenation at training time, nearly halving training cost while retaining autoregressive-style KV caching at inference.
Third, the concept module. An additive bottleneck sits between the transformer backbone and the LM head, decomposing the hidden state into h-bar = k-hat plus u-hat plus epsilon: k-hat is a weighted sum of "known concepts" (the 33,000 labeled ones), u-hat of "unknown concepts" (learned, low-rank factored to save parameters), and epsilon is a residual. Because the LM head is linear, any output token's logit decomposes exactly into known plus unknown plus residual, with no approximation. Concept attribution reads off directly, and steering (adding a vector along a concept's direction) becomes a closed-form edit rather than a post-hoc estimate.
The training objective combines the diffusion language-modeling loss, a concept loss (chunk-level, positive-only), a reconstruction loss, and an independence loss (pushing known and unknown concepts apart), with a teacher-forcing schedule that feeds ground-truth concepts early and anneals toward the model's own predictions to prevent concept leakage. The concept module's parameter share falls fast with scale: 89% at 10M, 4% at 8B, under 1% at frontier scale.
Does this structure hobble capability? The authors run IsoFLOP sweeps (finding the best model size at each compute budget) across three orders of magnitude of compute and four model families (autoregressive AR, causal diffusion CDLM, and each with concepts) and fit scaling laws. The finding: adding the concept module shifts the compute-optimal scaling exponents by a small, fixed per-backbone offset, not a tax that grows with scale. More surprisingly, all four interpretability metrics (concept prediction accuracy, known/unknown separation, concept contribution share, concept alignment) improve with compute, so larger models are easier to understand. Predicting Steerling-8B's validation loss from small-scale fits lands within 0.11 nats.
Against same-scale open base models on seven benchmarks:
| Model | MMLU | GSM8K | ARC-C | HellaSwag | Avg |
| Steerling-8B | 46.4 | 44.4 | 52.3 | 70.3 | 51.6 |
| LLaMA2 7B | 45.9 | 13.1 | 46.3 | 76.0 | 48.2 |
| LLaMA3 8B | 65.4 | 48.7 | 53.1 | 79.1 | 60.0 |
| OLMo 2 7B | 63.7 | 67.5 | 79.8 | 83.8 | 67.3 |
Every peer was trained on 2 to 16 times Steerling's compute, yet it lands within about 10% of their average and ahead of models trained at comparable budgets. It is visibly weak on Math (8.0): the pretraining corpus (Nemotron-CC-HQ) is thin on math and code, which mid-training (150B tokens of code and math augmentation) addresses, lifting MMLU 17 points and GSM8K 30 points.
This paper shakes the assumption that capability requires a black box. If interpretability costs only a fixed offset, not a growing tax, and improves with scale, then baking it into the standard training recipe has an economic case. For alignment, safety, and auditing work, a model whose attributions are exact by construction and whose steering is a closed-form edit is far stronger than post-hoc probes or SAEs that merely "look right." The Atlas library itself (33,000 concepts, 1 trillion tokens annotated) is a reusable asset. The cold water: Steerling-8B's absolute capability still trails head same-scale models (Math 8.0, average 51.6 versus OLMo2's 67.3). It proves the paradigm is viable and the cost controllable, not that it replaces a frontier model today.
The authors draw several lines. Steerling-8B's "inherent interpretability" is not full mechanistic transparency: it guarantees faithful attribution for specified queries (input, concept, training data), not that every neuron is intelligible. Training-data attribution uses representational similarity retrieval and explicitly does not claim causal influence. The most concrete limitations surfaced during training: before mid-training, steering barely worked, with a sharp quality drop and about a third of low-frequency concepts never activating at any injection strength, because the model never saw injection during training, an out-of-distribution perturbation that a dedicated steering phase in mid-training had to fix. Thin math and code, an over-aggressive masking schedule, and late-training concept-head entanglement were all caught only after pretraining and patched in mid-training, showing the recipe is immature, with several hyperparameters exposing problems only at 8B scale. The peer comparison covers only base models and seven benchmarks, not the instruct-tuned setting most users actually deploy.