Neural FOXP2 -- Language Specific Neuron Steering for Targeted Language Improvement in LLMs

Aman Chadha; Amitava Das; Anusa Saha; Tanmay Joshi; Vinija Jain

REVIEW 4 major objections 5 minor 29 references

A sparse set of internal features controls which language an LLM starts in, and steering those features can make Hindi or Spanish the default without retraining.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

T0 review · deepseek-v4-flash

2026-08-03 05:50 UTC pith:TP7XPQ2B

load-bearing objection A methodologically careful but empirically unproven steering recipe: the pipeline is new and the controls are thoughtful, but the headline numbers are reported without error bars, the main metric is the optimized objective, and the matched-prompt assumption is load-bearing and unchecked. the 4 major comments →

arxiv 2602.00945 v2 pith:TP7XPQ2B submitted 2026-02-01 cs.CL cs.AI

Neural FOXP2 -- Language Specific Neuron Steering for Targeted Language Improvement in LLMs

Anusa Saha , Tanmay Joshi , Vinija Jain , Aman Chadha , Amitava Das This is my paper

classification cs.CL cs.AI

keywords language defaultnesslanguage neuronssparse autoencodersactivation steeringlow-rank subspacemultilingual LLMsinference-time interventionmechanistic interpretability

verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

This paper argues that an LLM's tendency to default to English is not a diffuse side-effect of pretraining data imbalance but is governed by a sparse, low-rank control circuit of internal features, which the authors call language neurons. It introduces Neural FOXP2, a three-stage pipeline—localize, find steering directions, steer—that identifies these features in a sparse autoencoder basis, extracts the dominant low-rank directions of English-to-target language shift, and applies a signed, sparse activation edit within a chosen layer window. The claim is that this inference-time intervention shifts the model's language prior so that Hindi or Spanish becomes the default under weak prompting, while preserving task semantics across machine translation, QA, NLI, and summarization. If true, the paper demonstrates both a mechanistic fact about how multilingual ability is organized and a practical capability: targeted language improvement without retraining or fine-tuning, by reallocating access to capacity the model already has.

Core claim

The core claim is hypothesis H_LN: there exists a small set of internal units or circuitry whose intervention predictably changes the model's output language while preserving task semantics. The paper operationalizes this as controlling an early-step language advantage functional ΔM, the difference in probability mass the model assigns to target-language versus English token sets at the first few decoding steps. The discovery is that, on LLaMA-3 8B, this defaultness behaves like a localized, low-rank control signal: a compact support of SAE features that are both language-selective and causally lift target-language mass, a steering subspace recovered from SVD of matched English-to-target act

What carries the argument

The central object is the language-neuron set Nℓt: a sparse collection of dictionary features discovered by per-layer sparse autoencoders, where each feature is scored by the product of matched-pair selectivity (how much more it fires on target-language prompts than English prompts) and causal lift slope (how much a small push on that feature raises the early-step defaultness measure ΔM). The geometry is carried by the language-shift matrix ΔZ, built from paired feature-code differences between matched English and target prompts, and its layerwise SVD; the dominant right singular vectors define the steering subspace S and the eigengap/effective-rank spectra pick a contiguous intervention win

Load-bearing premise

The whole pipeline assumes that the parallel prompts in English, Hindi, and Spanish differ only in language, so that any activation difference or steering direction extracted from them reflects language identity rather than content, topic, register, or translation quality.

What would settle it

Take a set of meaning units and deliberately mismatch the translations so that the Hindi prompt describes different content than the English prompt; if the pipeline still produces the same defaultness gains and the SVD top directions remain sharp, the steering is exploiting content shift rather than a language circuit. Conversely, verifying with back-translation agreement and formality matching that gains persist only for semantically tight pairs would confirm the claim.

Watch this falsifier. Get emailed when new claim-graph text bears on it.

If this is right

If language defaultness is a sparse, low-rank control circuit, then multilingual models can serve multiple languages from a single frozen checkpoint by switching the edit, reducing the need for per-language fine-tuned models.
A one-scalar intensity knob can continuously dial from no-edit to strong defaultness within a calibrated safety envelope, giving practitioners a measurable trade-off between language preference and utility regression.
The causal tests—random-feature edits fail, ablating the discovered support collapses gains, editing outside the window is weaker, and cross-language leakage stays small—imply the effect is specific to the identified circuit rather than generic perturbation.
The method is best suited to mid-to-high resource languages already encoded in parametric memory but suppressed by the English prior; it does not create competence for low-resource languages.
The approach reframes some multilingual gaps as routing problems: a model may under-utilize an existing language channel, and inference-time steering can amplify that channel without continued pretraining.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

An editor inference: if the circuit claim extends beyond Hindi and Spanish, the same localization-and-steering recipe could be used to control other early-commitment defaults, such as register, formality, or output-format choices, wherever a matched-pair contrast can be constructed.
An editor inference: the matched-meaning assumption predicts a falsifiable diagnostic—if the SVD spectrum stays sharp when prompt pairs are mismatched in content, then the discovered directions encode topic rather than language; a careful reader could run that control cheaply before trusting the circuit interpretation.
An editor inference: the paper's separation of localization (where) from geometry (how) suggests a possible cross-lingual transfer test—whether a support and steering subspace discovered for Hindi can, after re-validation, partially transfer to typologically or script-related languages, which would strengthen the claim of a shared control mechanism.
An editor inference: the emphasis on first-token commitment implies that the technique is most robust for short-form, instruction-light interactions, and that long-form or domain-heavy prompts may dilute the effect; deployment evaluations should stratify by prompt length and domain.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Desk Editor's Note

A methodologically careful but empirically unproven steering recipe: the pipeline is new and the controls are thoughtful, but the headline numbers are reported without error bars, the main metric is the optimized objective, and the matched-prompt assumption is load-bearing and unchecked.

read the letter

You should know two things up front. First, the three-stage recipe is genuinely new: SAE-based feature localization, SVD of paired language shifts, and a signed sparse windowed edit, assembled this way, is not in LAPE or CAA. Second, the paper's central empirical claim is not supported by the text as written: Section 3 says FOXP2 “is expected to improve” the metrics, Table 1 has deltas with no error bars or significance tests, and no standard bench-mark scores (BLEU, EM, ROUGE) appear. The abstract states gains as fact, but the visible results section reads prospectively. That mismatch is the biggest problem.

What the paper does well: the method is specified precisely enough to reimplement, the ablation design is scientifically sound (random-feature edits, out-of-window edits, necessity and sufficiency tests, cross-language leakage), and the limitations section is unusually honest — it names matched-pair leakage, token-set dependence, and the risk of style confounds. The reproducibility protocol is thorough, even if the current statement has an unresolved “(???)” and the reference list duplicates a few entries.

The soft spots are real but not all equal. The strongest concern is not the circularity of optimizing Delta M — that is true but partly mitigated by the utility and leakage guardrails. The load-bearing assumption is that the parallel prompts (x_en, x_hi, x_es) are meaning-equivalent. If the translations differ in topic, entities, or register, the SVD directions encode content, not language. The authors themselves flag this in L4 but provide no check (no back-translation agreement, no human parity, no residual epsilon analysis). That is a serious omission for a mechanistic claim. The missing error bars and standard benchmarks are secondary but still important; without them the magnitude of the effect is unverifiable.

Who this is for: readers working on multilingual steering, SAE-based interpretability, or activation editing will want to know this pipeline exists. The paper deserves a serious referee — the method is clearly specified and the authors have thought hard about controls — but only with major revisions: report actual results with CIs, run the matched-pair validation, add standard task metrics, and fix the prose and placeholder. My own verdict would be reject-and-resubmit, not accept.

I'd bring it to a reading group if the group cares about multilingual interventions; otherwise it can wait. I won't cite it in my own work until the empirical claims are solidified.

Referee Report

4 major / 5 minor

Summary. The paper proposes Neural FOXP2, a three-stage inference-time intervention for shifting an LLM's 'language defaultness' from English to a target language (Hindi or Spanish). Stage I trains per-layer SAE dictionaries and ranks features by matched-pair selectivity and causal logit-mass lift to identify a sparse 'language neuron' set N. Stage II forms English-to-target activation-difference matrices, performs per-layer SVD, and selects a low-rank steering subspace plus a contiguous intervention window W. Stage III applies a signed, sparse activation shift in dictionary space, pushing toward target-language directions and suppressing an English attractor. The abstract and introduction claim that on LLaMA-3 8B the method improves Hindi and Spanish performance across MT, QA, NLI, and summarization. The empirical content is a single table (Table 1) reporting deltas for defaultness channels, leakage, stability, and task utility, along with extensive but largely prospective discussion of methodology, limitations, and reproducibility protocols.

Significance. If the empirical claims were solid, the paper would be significant: it would provide evidence for a sparse, low-rank, causally controllable language-selection circuit in a large multilingual LM, with a practical inference-time knob for default-language control. The paper also ships a commendably detailed protocol for reproducibility—pinned hashes, hook-site identity tests, token-set sensitivity audits, bootstrap subspace stability, and qualitative failure taxonomy—which is a strength. However, the central empirical assertion is not actually supported by the manuscript as written. The headline gains are described prospectively ('expected to improve', Section 3), Table 1 reports no confidence intervals or significance tests, and the outcome metric ΔM is exactly the quantity the pipeline optimizes at several stages. The matched-meaning-unit assumption that underlies the language-attribution is explicitly acknowledged in the limitations as a potential source of content leakage, but no verification is provided. These issues are load-bearing for both the mechanistic hypothesis and the practical claim.

major comments (4)

[Section 3 / Table 1] The paper's central empirical claim is not supported by reported results. Section 3 states that 'Across metrics, FOXP2 is expected to improve both Δmass and Δlid' (prospective), and Table 1 presents only deltas with absolute means, without error bars, confidence intervals, or significance tests. No standard task-level scores (BLEU, EM, ROUGE, accuracy) are reported, despite the abstract's claim of improvement across translation, QA, NLI, and summarization. The 'TaskΔS' column is a single scalar with no task breakdown. As written, the table cannot be distinguished from placeholder or indicative numbers, and the abstract overstates what the body establishes.
[Section 2.1.2, 2.1.3, 2.3; Table 1] There is a circularity problem in the headline outcome. The defaultness gain ΔM is used in Stage I to score features (LiftSlope, Eq. in 2.1.2), to set the per-layer support size K by lift saturation (2.1.3), and in Stage III to select edit magnitudes λ, β by maximizing Δgain on Dweak (2.3). The main Table 1 then reports Δmass, Δlid, and Δgain as evidence. This is not an independent evaluation of the hypothesis; it is a fit of the pipeline to the same objective. At minimum, the paper would need a holdout objective (e.g., human-judged natural start, or downstream task metrics on unseen tasks/languages) that is not optimized during steering.
[Section 2.1 Init-3, Section 2.2 II-1, Section 5.2, L4] The matched meaning-unit equivalence is unverified and is load-bearing for the mechanistic claim. Both selectivity (Eq. 'Matched-pair selectivity') and the language-shift matrix ΔZ (II-1) are computed from parallel prompts (x_en, x_hi, x_es) assumed to differ only in language. The paper specifies no back-translation agreement, human parity check, or register/entity matching. Section 5.2's low-rank interpretation treats ε as noise ('ε_k') without reporting its magnitude, and Limitation L4 explicitly states that 'mismatched same meaning pairs leak task/content differences into ΔZ, corrupting the SVD directions.' If the pairs diverge in topic or content, the SVD directions and the entire steering subspace encode those spurious axes, so the empirical gains could be content-shift artifacts rather than language control. A concrete test would be to report residual norms ||ε|| relative to ||ΔZ||
[Reproducibility Statement, R12, Appendix references] The reproducibility protocols are extensive, but they do not compensate for the absence of actual results. The Reproducibility Statement contains an explicit placeholder '( ???)' and multiple references are incomplete (e.g., '?', 'Turner et al., 2024; ?'). More importantly, the paper does not report that any of the promised reproducibility checks (null-edit identity test, token-set sensitivity bands, entropy-matched controls, bootstrap CIs) were actually run. The frequently-asked-questions and troubleshooting sections describe what 'we would do' (e.g., FAQ A: 'If correlation is weak, we change what we headline') rather than reporting what was done. As a journal submission, this is a proposal or position paper lacking empirical validation, not a completed research report.

minor comments (5)

[Figure 2 caption] The y-axis label reads 'Mean M M Across Target Languages' — likely a typo for 'Mean ΔM'.
[Throughout] The method is described as using SAEs in the abstract and Stage I, but Section 2.1 repeatedly refers to a 'V AE/autoencoder' and 'V AE dictionary'. The notation is inconsistent and should be unified.
[Table 1] The table is visually dense and hard to parse. Consider separating defaultness, leakage, stability, and utility into sub-tables, and adding bootstrap confidence intervals or at least standard deviations.
[Section 1 / Figure 3] Figure 3 is illustrative and schematic, but the three panels are not referenced in the main text in a way that connects them to specific empirical claims. Clarify whether these are actual data projections or conceptual diagrams.
[References] Several references have inconsistent year/venue formatting (e.g., Tang et al. 2024a/2024b, Turner et al. 2023/2024, Gao et al. duplicate Cunningham et al.). Please reconcile.

Circularity Check

0 steps flagged

No load-bearing circularity: the pipeline optimizes Δgain, but the reproducibility protocol separates the optimization set (Dweak) from the headline evaluation set (Dneutral), and external task/utility and ablation checks provide independent content.

full rationale

Stage I selects features by LiftSlope on ΔM over Dweak (Sec 2.1.2), sets support size by lift saturation on the same quantity (Sec 2.1.3), and Stage III tunes λ on a dev set to maximize Δgain (Sec 2.3). If the headline Δgain were measured on the same prompts, it would be the training objective reported as a prediction. However, the paper's reproducibility protocol explicitly separates Dweak (used for causal lift and gain) from Dneutral (used for headline reporting), and R3.2 states: 'We store explicit train/val/test splits and guarantee id-disjointness across roles.' Taken at face value, the reported Table 1 gains are out-of-sample, not fitted. The more serious scientific concern is the unverified matched meaning-unit equivalence (Init-3; L4/R8: 'mismatched same meaning pairs leak task/content differences into ΔZ, corrupting the SVD directions'). That is a confound that could make the steering directions encode content axes, but it is a validity threat rather than a circular reduction: the paper does not define the result as identical to its input by construction. There is no self-citation chain, no imported uniqueness theorem, and no citation-smuggled ansatz; the SAE/dictionary and LAPE references are external. The paper's own extensive limitations and troubleshooting sections further specify conditions under which the claims fail, which is inconsistent with a derivation that is circular by design. Score 1 reflects the unresolved ambiguity about whether the reported Δgain is computed on the same distribution used for tuning, not a demonstrated circular step.

Axiom & Free-Parameter Ledger

5 free parameters · 5 axioms · 2 invented entities

The central claim rests on several pulled-from-data quantities: support size, rank, window, and edit magnitudes are all selected by optimizing on the very Delta M metric family that later serves as the headline outcome. The measurement itself depends on hand-set token-set thresholds (Reproducibility Table 6) that the authors call a potential hidden lever. The existence and atomicity of 'language neurons' as a sparse low-rank circuit inherits its justification from cited SAE work, and within this paper it is assumed, not yet demonstrated.

free parameters (5)

Edit magnitudes lambda_l, beta_l = grid search: lambda in {0, lambda, 2lambda}, rho in {0, 0.5, 1}
Tuned on a held-out dev set to maximize Delta gain under regression constraints (Section 2.3, 'Choosing magnitudes').
Top-K language-neuron budget per layer = set by lift saturation on Dweak
Greedy addition until marginal gains in E[Delta M] plateau (Section 2.1.3); the support size is fit to the same Delta M quantity used as the headline outcome.
Steering rank r_l = effective rank + eigengap heuristic
Chosen by spectral diagnostics on the same dZ matrix (Section 2.2); data-dependent model selection on the discovery set.
Intervention window W = contiguous band maximizing sum Mass*Stab
Selected from per-layer strength/stability spectra (Section 2.2, 'Selecting the intervention window').
Token-set construction thresholds = e.g., tau_hi=0.6, tau_spec=2.0/1.5, tau_H=1.5, f_min=50
Hand-set hyperparameters in Reproducibility Table 6 that directly determine the headline Delta-mass values; the authors themselves call token sets a potential 'hidden lever' (FAQ B).

axioms (5)

domain assumption Language identity and defaultness are readable in residual-stream activations at the last prompt token and the first 1-3 decoding steps.
Central to the measurement target (Init-2, Fig. 7); if language choice is not committed at the probed site and horizon, localization and evaluation both fail.
domain assumption Matched meaning-unit prompts (same meaning in English/Hindi/Spanish) isolate language identity from topic and content.
Used to construct selectivity scores and the dZ matrix (Section 2.1, 2.2); the cross-lingual equivalence of the parallel prompts is asserted, never verified.
domain assumption SAE/VAE dictionary features are more monosemantic and causally stable intervention targets than raw neurons.
Borrowed from cited prior work (Bricken et al. 2023; Cunningham et al. 2023); the entire edit-in-z-space design (Section 2.1) rests on it.
ad hoc to paper English-to-target activation differences across matched meanings share a common low-dimensional subspace (dz_tilde ≈ sum alpha_i v_i + epsilon).
Stated in Section 5.2 ('Low-rank interpretation'); the SVD stage only recovers a language-control mode if this low-rank generative model holds, otherwise the top singular directions are arbitrary linear combinations of noise axes.
standard math Layerwise SVD and effective-rank/eigengap diagnostics identify the dominant control subspace.
Standard linear algebra used in Stage II (Section 2.2); the interpretation of the leading singular vectors as a language-selection mechanism is the domain assumption above.

invented entities (2)

Language neurons (language-specific SAE feature coordinates) no independent evidence
purpose: Sparse control set whose steering shifts the model's default language
An operationalized re-naming of prior 'language-specific neurons' (Tang 2024; Kojima 2024). The paper proposes causal necessity/sufficiency tests as evidence (Section 2.3, checks 1-5), but describes the outcomes as expected rather than reporting measured results.
Language defaultness (early-step mass advantage Delta M) no independent evidence
purpose: Operational construct and optimization target for steering
A new construct proxied by token-mass and LID channels. The FAQ claims human-preference calibration (rho=0.62/0.58) without a described protocol execution, annotator counts, or CIs; the construct's validity outside the paper is not established.

pith-pipeline@v1.3.0-alltime-deepseek · 53923 in / 19804 out tokens · 214453 ms · 2026-08-03T05:50:34.028051+00:00 · methodology

0 comments

read the original abstract

LLMs are multilingual by training, yet their lingua franca is often English, reflecting English language dominance in pretraining. Other languages remain in parametric memory but are systematically suppressed. We argue that language defaultness is governed by a sparse, low-rank control circuit, language neurons, that can be mechanistically isolated and safely steered. We introduce Neural FOXP2, that makes a chosen language (Hindi or Spanish) primary in a model by steering language-specific neurons. Neural FOXP2 proceeds in three stages: (i) Localize: We train per-layer SAEs so each activation decomposes into a small set of active feature components. For every feature, we quantify English vs. Hindi/Spanish selectivity overall logit-mass lift toward the target-language token set. Tracing the top-ranked features back to their strongest contributing units yields a compact language-neuron set. (ii) Steering directions: We localize controllable language-shift geometry via a spectral low-rank analysis. For each layer, we build English to target activation-difference matrices and perform layerwise SVD to extract the dominant singular directions governing language change. The eigengap and effective-rank spectra identify a compact steering subspace and an empirically chosen intervention window (where these directions are strongest and most stable). (iii) Steer: We apply a signed, sparse activation shift targeted to the language neurons. Concretely, within low to mid layers we add a positive steering along the target-language dominant directions and a compensating negative shift toward the null space for the English neurons, yielding controllable target-language defaultness.

Figures

Figures reproduced from arXiv: 2602.00945 by Aman Chadha, Amitava Das, Anusa Saha, Tanmay Joshi, Vinija Jain.

**Figure 2.** Figure 2: English is the lingua franca by default. Mean targetlanguage defaultness ∆M = Mℓ − MEn for 4 languages: Hindi, Spanish, French, and German. Across all targets, ∆M remains biased toward English (negative mass advantage), indicating that defaultness behaves as a prior rather than a capability limit. This motivates FOXP2s language-specific steering: shifting logit mass toward the target language. We formal… view at source ↗

**Figure 1.** Figure 1: Neural FOXP2: VAE/autoencoder-based discovery of language-selective neurons (Hindi or Spanish). The workflow proceeds in three stages: (a) Sample activations from a language model by running language-specific prompts and extracting intermediate activations h (ℓ) (x) at selected layers ℓ (left panel; Embedding → Transformer Blocks). (b) Learn a feature dictionary using an autoencoder (here, a VAE/SAE-style… view at source ↗

**Figure 3.** Figure 3: Geometric Interpretation of FOXP2 as Language Steering in Latent Space. Across these panels, FOXP2 appears as defaultness by motion and direction, not by understanding. Rather than reshaping the models conceptual topology, FOXP2 applies a low-rank, signed activation shifta small set of language vectors that nudges default language without touching beliefs or semantics. It teaches the model which language t… view at source ↗

**Figure 4.** Figure 4: Reproducibility protocol. Scan-friendly, end-to-end requirements to reproduce FOXP2: pinned versions (weights/tokenizer/detector), immutable manifests (data/token sets), exact hook sites, dictionary training recipe, localization and geometry extraction rules, signed edit schedules with hard guardrails, and deterministic aggregation that regenerates camera-ready tables. manifests with SHA256. • models/ – m… view at source ↗

**Figure 5.** Figure 5: Reproducibility checklist. A one-page audit sheet: pin every dependency (model/tokenizer/detectors/data), freeze all intermediate stage outputs (N , S, W), and ensure logs + a pinned command deterministically regenerate the camera-ready tables and figures. Step 7: Check style confounds that masquerade as language control. If the edit increases politeness markers or refusal templates that correlate with ce… view at source ↗

**Figure 6.** Figure 6: Troubleshooting fast-path. The five checks most likely to eliminate gibberish without altering the FOXP2 pipeline stages. F3. Improve dictionary training before blaming steering. If RelRecon(ℓ) is high in W: increase width m, tune λsparse, apply column-norm constraints on Wℓ, and train until reconstruction and sparsity stabilize. Feature dictionaries are only useful when they provide a stable coordinate sy… view at source ↗

**Figure 7.** Figure 7: Shared notation and measurement target. Stage I: Localize language neurons. I-1 (Dictionary training). Learn (Wℓ, bℓ) by reconstruction + sparsity: min Wℓ,bℓ Ex∼Dmix h kh (ℓ) (x) − Wℓz (ℓ) (x)k 2 2 + λsparsekz (ℓ) (x)k1 i . I-2 (Selectivity). For target ℓt ∈ {hi, es} using matched pairs (x (k) en , x (k) ℓt ): Sel(ℓ,ℓt) j = Ek[z (ℓ) j (x (k) ℓt )] − Ek[z (ℓ) j (x (k) en )], Sel f(ℓ,ℓt) j = Sel(ℓ,ℓt) j Stdk… view at source ↗

**Figure 8.** Figure 8: Stage I (Localize): sparse support discovery in dictionary coordinates. 46 [PITH_FULL_IMAGE:figures/full_fig_p046_8.png] view at source ↗

**Figure 9.** Figure 9: Stage II (Steer): low-rank geometry and intervention window. Stage III: Signed sparse edit and consolidation metrics. III-1 (Signed edit components). For each edited layer ℓ ∈ W, construct: δz(ℓ),+ ℓt (x) = λ (ℓ) + ΠS (ℓ) ℓt ΠN (ℓ) ℓt ∆z (ℓ,ℓt) (x) , δz(ℓ),− ℓt (x) = −λ (ℓ) − Π( S (ℓ) ℓt )⊥ ΠN (ℓ) en ∆z (ℓ,en)(x) . III-2 (Apply sparse code update). z (ℓ) (x) ← z (ℓ) (x) + ΠN (ℓ) ℓt δz(ℓ),+ ℓt (x)… view at source ↗

**Figure 10.** Figure 10: Neural FOXP2: complete pipeline (Initialization → Stage I–III). Initialization ( [PITH_FULL_IMAGE:figures/full_fig_p047_10.png] view at source ↗

**Figure 11.** Figure 11: Horizon sweep for Hindi defaultness (target). Each curve reports the mean (with bootstrap CI band) as a function of T ∈ {1, 2, 3, 5, 10} for the Hindi-tuned configuration: (a) E[∆Mhi T ] (mass channel), (b) E[∆LIDhi] (script-agnostic validator), (c) E[DEFAULTHI] (joint success requiring agreement), and (d) leakage proxy E[DEFAULTES] under the Hindi-tuned edit. This figure is the primary evidence that the … view at source ↗

**Figure 12.** Figure 12: Horizon sweep for Spanish defaultness (non-target regression control). Same protocol as [PITH_FULL_IMAGE:figures/full_fig_p069_12.png] view at source ↗

**Figure 13.** Figure 13: Cross-metric agreement map across horizons. Cells indicate whether the mass channel and LID validator agree in direction and significance at each T (Hindi and Spanish shown side-by-side). This provides a compact construct-validity diagnostic: robust defaultness gains should occupy a contiguous block at small horizons without collapsing into mass-only artifacts at larger T. Sensitivity bands. For each meth… view at source ↗

**Figure 14.** Figure 14: Token-set sensitivity bands over (Vℓ, w). For each method, we report the median and dispersion of (a) ∆hi mass and (b) DEFAULTHI over script-only, diagnostic, and transliteration-aware token sets. Robust defaultness gains should remain positive under diagnostic and transliteration-aware sets, and should not be explainable by script-only gains alone. on a distributed structure rather than a small brittle l… view at source ↗

**Figure 15.** Figure 15: Dropout robustness on Vℓ. Curves show ρ 7→ E[∆Mhi T ] and ρ 7→ E[DEFAULTHI] (with CI bands) under token-set dropout. Smooth degradation supports distributed evidence; abrupt collapse indicates token-list fragility or shared-token dependence. Shared-token inflation stress test. We rerun the mass computation under an adversarial setting where shared tokens are not downweighted (effectively increasing w(u) f… view at source ↗

**Figure 16.** Figure 16: Window sweep over layers. We shift the intervention window across depth and plot defaultness gain (DEFAULTHI and/or ∆hi mass) versus window center. This figure localizes where FOXP2 is causally effective, separating where the signal lives from where the output is produced [PITH_FULL_IMAGE:figures/full_fig_p072_16.png] view at source ↗

**Figure 17.** Figure 17: Defaultness gains stratified by KL drift. We bucket prompts by KLT (x) under the Hindi-tuned edit and report DEFAULTHI (and optionally continuous channels) in each bucket. Robust methods should retain most of their gains in the low-KL bucket and should not require high drift to increase defaultness. 72 [PITH_FULL_IMAGE:figures/full_fig_p072_17.png] view at source ↗

**Figure 18.** Figure 18: Domain-transfer retention curves. Retention is computed as DEFAULTHIOOD/DEFAULTHIID (with CIs) for multiple OOD prompt families and OOD task domains. We also report leakage and ∆S per OOD slice to ensure the retained gains are not achieved via broad regressions [PITH_FULL_IMAGE:figures/full_fig_p073_18.png] view at source ↗

discussion (0)

Reference graph

Works this paper leans on

29 extracted references · 1 canonical work pages

[1]

XQuAD : A cross-lingual question answering dataset

Artetxe, M., Ruder, S., and Yogatama, D. XQuAD : A cross-lingual question answering dataset. In Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics (ACL), 2020. URL https://arxiv.org/abs/1910.11856

Pith/arXiv arXiv 2020
[2]

Towards monosemanticity: Decomposing language models with dictionary learning

Bricken, T., Templeton, A., Batson, J., Chen, B., Jermyn, A., Conerly, T., Turner, N., Anil, C., Denison, C., Askell, A., Lasenby, R., Wu, L., Kravec, S., Schiefer, N., Maxwell, T., Joseph, N., Hatfield-Dodds, Z., Tamkin, A., Nguyen, K., McLean, C., Burke, L., Hume, T., Carter, S., Henighan, T., and Olah, C. Towards monosemanticity: Decomposing language m...

2023
[3]

R., Schwenk, H., and Stoyanov, V

Conneau, A., Rinott, R., Lample, G., Williams, A., Bowman, S. R., Schwenk, H., and Stoyanov, V. XNLI : Evaluating cross-lingual sentence representations. In Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing (EMNLP), 2018. URL https://arxiv.org/abs/1809.05053

Pith/arXiv arXiv 2018
[4]

Cunningham, H., Ewart, A., Riggs, L., Huben, R., and Sharkey, L. D. Sparse autoencoders find highly interpretable features in language models. arXiv preprint arXiv:2309.08600, 2023. doi:10.48550/arXiv.2309.08600. URL https://arxiv.org/abs/2309.08600

Pith/arXiv arXiv doi:10.48550/arxiv.2309.08600 2023
[5]

Llm.int8(): 8-bit matrix multiplication for transformers at scale

Dettmers, T., Lewis, M., Belkada, Y., and Zettlemoyer, L. Llm.int8(): 8-bit matrix multiplication for transformers at scale. arXiv preprint arXiv:2208.07339, 2022. doi:10.48550/arXiv.2208.07339. URL https://arxiv.org/abs/2208.07339

Pith/arXiv arXiv doi:10.48550/arxiv.2208.07339 2022
[7]

Toy models of superposition

Elhage, N., Hume, T., Olsson, C., Schiefer, N., Henighan, T., Askell, A., Bai, Y., Chen, A., Conerly, T., DasSarma, N., Drain, D., Ganguli, D., Hatfield-Dodds, Z., Hernandez, D., Jones, A., Kernion, J., Lovitt, L., Nanda, N., Ndousse, K., Joseph, N., Martin, B., McCandlish, S., Olah, C., Shlegeris, B., Thorisson, K., and Kaplan, J. Toy models of superposi...

Pith/arXiv arXiv doi:10.48550/arxiv.2209.10652 2022
[8]

Gao, L. et al. Sparse autoencoders find highly interpretable features in language models. arXiv preprint arXiv:2309.08600, 2023

Pith/arXiv arXiv 2023
[9]

S., Mubasshir, K., Li, Y.-F., Shahriyar, R., Hossain, M

Hasan, T., Bhattacharjee, A., Islam, M. S., Mubasshir, K., Li, Y.-F., Shahriyar, R., Hossain, M. S. A., and Iqbal, A. XL - Sum : Large-scale multilingual abstractive summarization for 44 languages. In Findings of the Association for Computational Linguistics: ACL-IJCNLP 2021 , 2021. URL https://arxiv.org/abs/2106.13822

Pith/arXiv arXiv 2021
[10]

M., Kocisk \'y , T., Grefenstette, E., Espeholt, L., Kay, W., Suleyman, M., and Blunsom, P

Hermann, K. M., Kocisk \'y , T., Grefenstette, E., Espeholt, L., Kay, W., Suleyman, M., and Blunsom, P. Teaching machines to read and comprehend. Advances in Neural Information Processing Systems, 28, 2015. URL https://proceedings.neurips.cc/paper/2015/hash/afdec7005cc9f14302cd0474fd0f3c96-Abstract.html

2015
[11]

On the multilingual ability of decoder-based pre-trained language models: Finding and controlling language-specific neurons

Kojima, T., Okimura, I., Iwasawa, Y., Yanaka, H., and Matsuo, Y. On the multilingual ability of decoder-based pre-trained language models: Finding and controlling language-specific neurons. In Proceedings of the 2024 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies (Volume 1: Long Paper...

doi:10.18653/v1/2024.naacl-long.384 2024
[12]

On the multilingual ability of decoder-based pre-trained language models: Finding and controlling language-specific neurons, 2024 b

Kojima, T., Okimura, I., Iwasawa, Y., Yanaka, H., and Matsuo, Y. On the multilingual ability of decoder-based pre-trained language models: Finding and controlling language-specific neurons, 2024 b

2024
[13]

Locating and editing factual associations in gpt

Meng, K., Bau, D., Andonian, A., and Belinkov, Y. Locating and editing factual associations in gpt. In Advances in Neural Information Processing Systems, 2022. doi:10.48550/arXiv.2202.05262. URL https://arxiv.org/abs/2202.05262

Pith/arXiv arXiv doi:10.48550/arxiv.2202.05262 2022
[14]

Meta llama 3

Meta AI . Meta llama 3. Blog post, 2024 a . URL https://ai.meta.com/blog/meta-llama-3/

2024
[15]

Meta-llama-3-8b model card

Meta AI . Meta-llama-3-8b model card. Hugging Face model card, 2024 b . URL https://huggingface.co/meta-llama/Meta-Llama-3-8B

2024
[16]

Panickssery, N. et al. Steering L lama 2 via contrastive activation addition. arXiv preprint arXiv:2312.06681, 2023

Pith/arXiv arXiv 2023
[17]

Squad: 100,000+ questions for machine comprehension of text

Rajpurkar, P., Zhang, J., Lopyrev, K., and Liang, P. Squad: 100,000+ questions for machine comprehension of text. In Proceedings of the 2016 Conference on Empirical Methods in Natural Language Processing (EMNLP), pp.\ 2383--2392. Association for Computational Linguistics, 2016. doi:10.18653/v1/D16-1264. URL https://aclanthology.org/D16-1264/

doi:10.18653/v1/d16-1264 2016
[18]

Rimsky, N. et al. Steering L lama 2 via contrastive activation addition, 2024

2024
[19]

Trust region policy optimization

Schulman, J., Wolski, F., Dhariwal, P., Radford, A., and Klimov, O. Trust region policy optimization. In International Conference on Machine Learning (ICML), 2015. arXiv:1502.05477

Pith/arXiv arXiv 2015
[20]

and Basirat, A

Shani, N. and Basirat, A. Language dominance in multilingual large language models. In Proceedings of the 8th BlackboxNLP Workshop: Analyzing and Interpreting Neural Networks for NLP, pp.\ 137--148, Suzhou, China, November 2025. Association for Computational Linguistics. ISBN 979-8-89176-346-3. doi:10.18653/v1/2025.blackboxnlp-1.7. URL https://aclantholog...

work page doi:10.18653/v1/2025.blackboxnlp-1.7 2025
[21]

D., Ng, A., and Potts, C

Socher, R., Perelygin, A., Wu, J., Chuang, J., Manning, C. D., Ng, A., and Potts, C. Recursive deep models for semantic compositionality over a sentiment treebank. In Proceedings of the 2013 Conference on Empirical Methods in Natural Language Processing (EMNLP), pp.\ 1631--1642, 2013. URL https://aclanthology.org/D13-1170/

2013
[22]

Language-specific neurons: The key to multilingual capabilities in large language models, 2024 a

Tang, L., Laban, P., and Ravichandran, D. Language-specific neurons: The key to multilingual capabilities in large language models, 2024 a

2024
[23]

Language-specific neurons: The key to multilingual capabilities in large language models

Tang, T., Luo, W., Huang, H., Zhang, D., Wang, X., Zhao, X., Wei, F., and Wen, J.-R. Language-specific neurons: The key to multilingual capabilities in large language models. In Ku, L.-W., Martins, A., and Srikumar, V. (eds.), Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pp.\ 5701--5715, ...

doi:10.18653/v1/2024.acl-long.309 2024
[24]

R., Cross, J., C elebi, O., Elbayad, M., Heafield, K., He, T., Guzm \'a n, F., et al

Team, N., Costa-juss \`a , M. R., Cross, J., C elebi, O., Elbayad, M., Heafield, K., He, T., Guzm \'a n, F., et al. The FLORES -200 evaluation benchmark for low-resource and multilingual machine translation. arXiv preprint arXiv:2207.04672, 2022. URL https://arxiv.org/abs/2207.04672

Pith/arXiv arXiv 2022
[25]

Activation addition: Steering language models without fine-tuning

Turner, A., Thiergart, L., Leech, G., Udell, M., and Olah, C. Activation addition: Steering language models without fine-tuning. arXiv preprint arXiv:2404.04871, 2024. URL https://arxiv.org/abs/2404.04871

Pith/arXiv arXiv 2024
[26]

Turner, A. et al. Activation addition: Steering language models without optimization. arXiv preprint arXiv:2308.10248, 2023

Pith/arXiv arXiv 2023
[27]

Williams, A., Nangia, N., and Bowman, S. R. A broad-coverage challenge corpus for sentence understanding through inference. In Proceedings of the 2018 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies (NAACL-HLT), pp.\ 1112--1122. Association for Computational Linguistics, 2018. doi:10.1...

doi:10.18653/v1/n18-1101 2018
[28]

J., and Wan, A

Yu, M., Wang, D., Shan, Q., Reed, C. J., and Wan, A. The super weight in large language models. arXiv preprint arXiv:2411.07191, 2024. doi:10.48550/arXiv.2411.07191. URL https://arxiv.org/abs/2411.07191

Pith/arXiv arXiv doi:10.48550/arxiv.2411.07191 2024
[29]

and Nanda, N

Zhang, F. and Nanda, N. Towards best practices of activation patching in language models: Metrics and methods. arXiv preprint arXiv:2309.16042, 2023. doi:10.48550/arXiv.2309.16042. URL https://arxiv.org/abs/2309.16042

Pith/arXiv arXiv doi:10.48550/arxiv.2309.16042 2023
[30]

write newline

" write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 global.max substring 't := if while FUNCTION format.date year duplicate empty "emp...

[1] [1]

XQuAD : A cross-lingual question answering dataset

Artetxe, M., Ruder, S., and Yogatama, D. XQuAD : A cross-lingual question answering dataset. In Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics (ACL), 2020. URL https://arxiv.org/abs/1910.11856

Pith/arXiv arXiv 2020

[2] [2]

Towards monosemanticity: Decomposing language models with dictionary learning

Bricken, T., Templeton, A., Batson, J., Chen, B., Jermyn, A., Conerly, T., Turner, N., Anil, C., Denison, C., Askell, A., Lasenby, R., Wu, L., Kravec, S., Schiefer, N., Maxwell, T., Joseph, N., Hatfield-Dodds, Z., Tamkin, A., Nguyen, K., McLean, C., Burke, L., Hume, T., Carter, S., Henighan, T., and Olah, C. Towards monosemanticity: Decomposing language m...

2023

[3] [3]

R., Schwenk, H., and Stoyanov, V

Conneau, A., Rinott, R., Lample, G., Williams, A., Bowman, S. R., Schwenk, H., and Stoyanov, V. XNLI : Evaluating cross-lingual sentence representations. In Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing (EMNLP), 2018. URL https://arxiv.org/abs/1809.05053

Pith/arXiv arXiv 2018

[4] [4]

Cunningham, H., Ewart, A., Riggs, L., Huben, R., and Sharkey, L. D. Sparse autoencoders find highly interpretable features in language models. arXiv preprint arXiv:2309.08600, 2023. doi:10.48550/arXiv.2309.08600. URL https://arxiv.org/abs/2309.08600

Pith/arXiv arXiv doi:10.48550/arxiv.2309.08600 2023

[5] [5]

Llm.int8(): 8-bit matrix multiplication for transformers at scale

Dettmers, T., Lewis, M., Belkada, Y., and Zettlemoyer, L. Llm.int8(): 8-bit matrix multiplication for transformers at scale. arXiv preprint arXiv:2208.07339, 2022. doi:10.48550/arXiv.2208.07339. URL https://arxiv.org/abs/2208.07339

Pith/arXiv arXiv doi:10.48550/arxiv.2208.07339 2022

[6] [7]

Toy models of superposition

Elhage, N., Hume, T., Olsson, C., Schiefer, N., Henighan, T., Askell, A., Bai, Y., Chen, A., Conerly, T., DasSarma, N., Drain, D., Ganguli, D., Hatfield-Dodds, Z., Hernandez, D., Jones, A., Kernion, J., Lovitt, L., Nanda, N., Ndousse, K., Joseph, N., Martin, B., McCandlish, S., Olah, C., Shlegeris, B., Thorisson, K., and Kaplan, J. Toy models of superposi...

Pith/arXiv arXiv doi:10.48550/arxiv.2209.10652 2022

[7] [8]

Gao, L. et al. Sparse autoencoders find highly interpretable features in language models. arXiv preprint arXiv:2309.08600, 2023

Pith/arXiv arXiv 2023

[8] [9]

S., Mubasshir, K., Li, Y.-F., Shahriyar, R., Hossain, M

Hasan, T., Bhattacharjee, A., Islam, M. S., Mubasshir, K., Li, Y.-F., Shahriyar, R., Hossain, M. S. A., and Iqbal, A. XL - Sum : Large-scale multilingual abstractive summarization for 44 languages. In Findings of the Association for Computational Linguistics: ACL-IJCNLP 2021 , 2021. URL https://arxiv.org/abs/2106.13822

Pith/arXiv arXiv 2021

[9] [10]

M., Kocisk \'y , T., Grefenstette, E., Espeholt, L., Kay, W., Suleyman, M., and Blunsom, P

Hermann, K. M., Kocisk \'y , T., Grefenstette, E., Espeholt, L., Kay, W., Suleyman, M., and Blunsom, P. Teaching machines to read and comprehend. Advances in Neural Information Processing Systems, 28, 2015. URL https://proceedings.neurips.cc/paper/2015/hash/afdec7005cc9f14302cd0474fd0f3c96-Abstract.html

2015

[10] [11]

On the multilingual ability of decoder-based pre-trained language models: Finding and controlling language-specific neurons

Kojima, T., Okimura, I., Iwasawa, Y., Yanaka, H., and Matsuo, Y. On the multilingual ability of decoder-based pre-trained language models: Finding and controlling language-specific neurons. In Proceedings of the 2024 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies (Volume 1: Long Paper...

doi:10.18653/v1/2024.naacl-long.384 2024

[11] [12]

On the multilingual ability of decoder-based pre-trained language models: Finding and controlling language-specific neurons, 2024 b

Kojima, T., Okimura, I., Iwasawa, Y., Yanaka, H., and Matsuo, Y. On the multilingual ability of decoder-based pre-trained language models: Finding and controlling language-specific neurons, 2024 b

2024

[12] [13]

Locating and editing factual associations in gpt

Meng, K., Bau, D., Andonian, A., and Belinkov, Y. Locating and editing factual associations in gpt. In Advances in Neural Information Processing Systems, 2022. doi:10.48550/arXiv.2202.05262. URL https://arxiv.org/abs/2202.05262

Pith/arXiv arXiv doi:10.48550/arxiv.2202.05262 2022

[13] [14]

Meta llama 3

Meta AI . Meta llama 3. Blog post, 2024 a . URL https://ai.meta.com/blog/meta-llama-3/

2024

[14] [15]

Meta-llama-3-8b model card

Meta AI . Meta-llama-3-8b model card. Hugging Face model card, 2024 b . URL https://huggingface.co/meta-llama/Meta-Llama-3-8B

2024

[15] [16]

Panickssery, N. et al. Steering L lama 2 via contrastive activation addition. arXiv preprint arXiv:2312.06681, 2023

Pith/arXiv arXiv 2023

[16] [17]

Squad: 100,000+ questions for machine comprehension of text

Rajpurkar, P., Zhang, J., Lopyrev, K., and Liang, P. Squad: 100,000+ questions for machine comprehension of text. In Proceedings of the 2016 Conference on Empirical Methods in Natural Language Processing (EMNLP), pp.\ 2383--2392. Association for Computational Linguistics, 2016. doi:10.18653/v1/D16-1264. URL https://aclanthology.org/D16-1264/

doi:10.18653/v1/d16-1264 2016

[17] [18]

Rimsky, N. et al. Steering L lama 2 via contrastive activation addition, 2024

2024

[18] [19]

Trust region policy optimization

Schulman, J., Wolski, F., Dhariwal, P., Radford, A., and Klimov, O. Trust region policy optimization. In International Conference on Machine Learning (ICML), 2015. arXiv:1502.05477

Pith/arXiv arXiv 2015

[19] [20]

and Basirat, A

Shani, N. and Basirat, A. Language dominance in multilingual large language models. In Proceedings of the 8th BlackboxNLP Workshop: Analyzing and Interpreting Neural Networks for NLP, pp.\ 137--148, Suzhou, China, November 2025. Association for Computational Linguistics. ISBN 979-8-89176-346-3. doi:10.18653/v1/2025.blackboxnlp-1.7. URL https://aclantholog...

work page doi:10.18653/v1/2025.blackboxnlp-1.7 2025

[20] [21]

D., Ng, A., and Potts, C

Socher, R., Perelygin, A., Wu, J., Chuang, J., Manning, C. D., Ng, A., and Potts, C. Recursive deep models for semantic compositionality over a sentiment treebank. In Proceedings of the 2013 Conference on Empirical Methods in Natural Language Processing (EMNLP), pp.\ 1631--1642, 2013. URL https://aclanthology.org/D13-1170/

2013

[21] [22]

Language-specific neurons: The key to multilingual capabilities in large language models, 2024 a

Tang, L., Laban, P., and Ravichandran, D. Language-specific neurons: The key to multilingual capabilities in large language models, 2024 a

2024

[22] [23]

Language-specific neurons: The key to multilingual capabilities in large language models

Tang, T., Luo, W., Huang, H., Zhang, D., Wang, X., Zhao, X., Wei, F., and Wen, J.-R. Language-specific neurons: The key to multilingual capabilities in large language models. In Ku, L.-W., Martins, A., and Srikumar, V. (eds.), Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pp.\ 5701--5715, ...

doi:10.18653/v1/2024.acl-long.309 2024

[23] [24]

R., Cross, J., C elebi, O., Elbayad, M., Heafield, K., He, T., Guzm \'a n, F., et al

Team, N., Costa-juss \`a , M. R., Cross, J., C elebi, O., Elbayad, M., Heafield, K., He, T., Guzm \'a n, F., et al. The FLORES -200 evaluation benchmark for low-resource and multilingual machine translation. arXiv preprint arXiv:2207.04672, 2022. URL https://arxiv.org/abs/2207.04672

Pith/arXiv arXiv 2022

[24] [25]

Activation addition: Steering language models without fine-tuning

Turner, A., Thiergart, L., Leech, G., Udell, M., and Olah, C. Activation addition: Steering language models without fine-tuning. arXiv preprint arXiv:2404.04871, 2024. URL https://arxiv.org/abs/2404.04871

Pith/arXiv arXiv 2024

[25] [26]

Turner, A. et al. Activation addition: Steering language models without optimization. arXiv preprint arXiv:2308.10248, 2023

Pith/arXiv arXiv 2023

[26] [27]

Williams, A., Nangia, N., and Bowman, S. R. A broad-coverage challenge corpus for sentence understanding through inference. In Proceedings of the 2018 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies (NAACL-HLT), pp.\ 1112--1122. Association for Computational Linguistics, 2018. doi:10.1...

doi:10.18653/v1/n18-1101 2018

[27] [28]

J., and Wan, A

Yu, M., Wang, D., Shan, Q., Reed, C. J., and Wan, A. The super weight in large language models. arXiv preprint arXiv:2411.07191, 2024. doi:10.48550/arXiv.2411.07191. URL https://arxiv.org/abs/2411.07191

Pith/arXiv arXiv doi:10.48550/arxiv.2411.07191 2024

[28] [29]

and Nanda, N

Zhang, F. and Nanda, N. Towards best practices of activation patching in language models: Metrics and methods. arXiv preprint arXiv:2309.16042, 2023. doi:10.48550/arXiv.2309.16042. URL https://arxiv.org/abs/2309.16042

Pith/arXiv arXiv doi:10.48550/arxiv.2309.16042 2023

[29] [30]

write newline

" write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 global.max substring 't := if while FUNCTION format.date year duplicate empty "emp...