Can We Understand How LLMs Reason? — ACM, Causal Abstraction, and J-Space

Large language models write essays, prove lemmas, and ship code — but we still cannot fully explain how.

That is the opening of Communications of the ACM, July 7, 2026: engineering is ahead of science. Stanford philosopher-computer scientist Thomas Icard and the mechanistic interpretability community are trying to close the gap with causal abstraction — asking when a billion-parameter network implements a higher-level algorithm you would recognize from logic or cognitive science.

Six days earlier, Anthropic published J-space — a privileged internal workspace in Claude, instrumented with the J-lens, with swap interventions that change answers. Same scientific program, different venue: academia on Llama, frontier lab on Claude.

This post maps the ACM story, embeds explainx.ai's J-lens and J-space deep dives, and states what builders should do with partial understanding.

TL;DR — three layers of "understanding"

Icard's bottom line: interpretability may never yield a closed-form equation for GPT-class models — but hidden algorithms can become partly legible.

Causal abstraction — ideal gas for neural nets

Icard's analogy (via ACM): physicists abstract colliding molecules to pressure and temperature. Neural nets can likewise be described above matrix multiplies — with much messier overlap, because one neuron can participate in many functions at once.

The precision tool is causal abstraction:

When do two causal models in different languages describe the same underlying reality — one more abstract than the other?

Applied to LLMs: you hypothesize a high-level algorithm (logical inference, cyclic arithmetic), then test whether internal states factor like that algorithm's variables. Not "it got the answer right" — the mechanism matches.

Atticus Geiger (Icard's former Ph.D. student, now a leading interpretability researcher at Goodfire) helped build this framework. The field's progress since 2021 spans training-data reflection, post-training reshaping, and links between internal abstractions and generalization — the same arc Mandy Lu's "why does AI work?" debate highlights from the theory side.

BERT learned logic (2021)

In a 2021 study, Icard, Geiger, and collaborators showed a BERT-based language model internally implements pieces of a logical reasoning system — including inferences with quantifiers (every, some, not) and negation.

That was early proof the question is not only philosophical ("does it mimic reasoning?") but empirical: you can align internal computation to symbolic steps.

"Arithmetic in the Wild" — August + 6 = February

Goodfire's recent paper (Geiger-led, cited by Icard in ACM) asks: What month is six months after August?

Naive expectation: reason directly on a 12-month cycle.

Observed mechanism in a Llama-based model:

August → internal month index 8
       → decimal addition: 6 + 8 = 14
       → map 14 back: 14 = 12 + 2 → February

The same template appears for weekdays and clock time — one general strategy, multiple domains, never explicitly trained as "use decimal first."

Icard: "A beautiful example of using causal abstraction methods to uncover how a large language model reasons about cyclic concepts."

explainx.ai note: This is why private enterprise evals must include weird edge cases — models can pass benchmarks while running surprising internal routes that break under distribution shift.

Open weights vs closed frontier — who can look inside?

From ACM:

Icard: commercial models are off-limits, but Anthropic and Google DeepMind have dedicated interpretability groups — "we regularly talk with both."

That asymmetry feeds Nadella's trust boundary argument: vendors see more of the model than you do; you must own evals, traces, and monitoring on your side.

Embedded guide — What is the J-lens?

Full post: What Is the J-Lens? Anthropic's Jacobian Lens for Reading Claude's Silent Thoughts

The J-lens (Jacobian lens) is Anthropic's July 2026 instrument for the same scientific goal as causal abstraction — but operationalized for frontier Claude:

Forward:  activations → layers → logits → (maybe) token
J-lens:   which activation direction pushes logits toward word W later?

For each vocabulary token, find the internal pattern that increases future probability of saying that word. Stack layer readouts → a live list of silent "words on Claude's mind."

From Anthropic's global workspace post:

"For every word in Claude's vocabulary, the J-lens finds the internal activity pattern that makes Claude more likely to say that word at some point in the future."

Causal-abstraction parallel: Goodfire explains month arithmetic as a high-level two-step algorithm; J-lens explains France capital questions as workspace variables you can surgically replace — same epistemology, different granularity.

Embedded guide — Anthropic's J-space global workspace

Full post: Anthropic's J-Space: A Global Workspace Inside Claude

On July 6, 2026, Anthropic argued Claude has a strikingly similar divide to human conscious access: most processing is automatic; a small privileged channel holds reportable, controllable concepts — the J-space.

TL;DR — J-space at a glance

Global workspace theory hook

Anthropic tests five properties from global workspace theory (Dehaene, Naccache, Baars):

External validation: Dehaene and Naccache commented; Neel Nanda (Google DeepMind) replicated on open weights.

Safety — why partial understanding matters

May's NLAs read eval awareness; July's J-lens intervenes — the progression ACM's article predicts for safer, more reliable models.

Companion: Is Claude conscious? — access vs phenomenal consciousness; what Anthropic refuses to claim.

ACM promises and limits — mapped to product reality

Icard sees promise for:

Major challenges (ACM): scale to largest models; reduce human insight bottlenecks.

explainx.ai translation for teams: Monitoring, not full alignment — interpretability is a dashboard slice, paired with human review on high-stakes evals, not a substitute for governance.

One stack — July 2026 interpretability timeline

2021  BERT + logical inference (Icard / Geiger)
2025  NLAs — natural language feature descriptions (Anthropic)
2026  Arithmetic in the Wild — cyclic Llama reasoning (Goodfire / Geiger)
Jul 6 J-space + J-lens — workspace + swaps (Anthropic)
Jul 7 ACM — causal abstraction survey (Icard quoted)

Academia and frontier labs are converging on causality, not just saliency maps. Enterprises sit on the consumption side: you get behavior from APIs, occasionally research blog visibility, rarely internal graphs.

explainx.ai read

1. "Can we understand reasoning?" — Partially, and improving fast; not completely
2. Causal abstraction — The right standard (mechanism match, not answer match)
3. J-lens / J-space — Frontier proof that interventions beat post-hoc stories
4. Open weights — Science runs on Llama; production runs on Claude/GPT — plan for opacity
5. Enterprise moat — Own evals + traces; treat vendor interpretability as R&D signal

Science is catching up to engineering. Your job is not to wait for the full theory — it is to log, test, and intervene on your workflows while the field maps the rest.

How to read these claims without overreaching

The three layers in the table above are easy to blur, so it helps to keep them apart when you cite this research internally.

1. A correct answer is behavior, not mechanism. The ACM piece holds up causal abstraction because it tests whether internal states factor like the proposed algorithm, not whether the output was right.
2. The Llama month example is one model and one task family. The decimal-then-modulo route was observed in a Llama-based model for months, weekdays, and clock time. It does not tell you how Claude or any closed model handles the same question.
3. The BERT result dates to 2021. It shows alignment to symbolic steps is possible, not that today's frontier models reason this way.
4. J-space findings come from Anthropic's own models and a replication on open weights. Treat them as a strong research signal, not a guarantee about your deployment.
5. Icard's own bottom line is partial legibility. No closed-form equation is promised.

If you write about this for your team, state which layer a claim sits on, which model it was shown on, and whether it was observed or only hypothesized. That one habit keeps interpretability from being quoted as proof of reliability.

Related on explainx.ai

• What is the J-lens? — full Jacobian lens guide
• Anthropic J-space — global workspace deep dive
• Is Claude conscious? — J-space companion
• What are NLAs?
• Anthropic NLAs deep dive (May 2026)
• Why AI works — Mandy Lu theory gap
• AI interpretability monitoring for teams
• How to build enterprise AI benchmarks
• Nadella Reverse Information Paradox
• AI alignment introduction for product teams

Sources: Communications of the ACM — Can We Understand How Large Language Models Reason?, Jul 7 2026 · Anthropic — A global workspace in language models, Jul 6 2026 · Goodfire — Arithmetic in the Wild (Geiger et al.)

Research citations and model access reflect publication-week status, July 2026. Open-weight tooling and commercial model internals change frequently — verify repos and terms before production monitoring designs.