{"id":"56ab7507-de59-47ec-90d9-bd9e4162aca8","arxiv_id":"2604.24302","paper_version":1,"verdict":"UNVERDICTED","confidence":"LOW","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":1,"one_line_summary":"DFA learns a differentiable alignment to transfer circuits from small to large language models, recovering target faithfulness competitive with direct methods on Llama-3 1B to 3B but degrading with scale and architecture differences.","lead":"The paper introduces Differentiable Faithfulness Alignment (DFA), a method to transfer circuit information from smaller language models to larger ones by learning a differentiable mapping of node importance scores trained against a soft faithfulness objective. This approach aims to avoid expensive full circuit discovery on target models and shows competitive results on some Llama-3 scales but weaker performance on larger gaps or different architectures.","discovery_kind":"new_method","skeptic_critique":{"model":"grok-4.3","headline":"Learned mapping may discover high-faithfulness target circuits independently of source node scores, weakening the transfer interpretation.","rationale":"The reader's weakest assumption correctly flags the core premise that source scores can be aligned to serve as priors without full target discovery. The concern above sharpens this by identifying a specific mechanism (unconstrained optimization of the mapping) that could falsify the premise even if the alignment procedure runs. Because the reader reviewed only the abstract, this supplies a concrete, code-checkable angle on the same assumption; if the test passes, the claim strengthens and the verdict can move toward ACCEPT.","tokens_in":1735,"tokens_out":429,"duration_ms":48979,"concrete_test":"In the released code, modify the alignment training loop to replace the source importance vector with a constant (or random) vector while keeping the same faithfulness objective and optimizer; recompute faithfulness on the six tasks for the Llama-3 1B→3B setting. If the resulting scores remain within 15% of the reported DFA numbers, source circuit information is not required for the observed performance.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The central claim is that DFA transfers circuit information from a smaller source model to a larger target via a learned differentiable alignment of node importance scores, trained with a soft faithfulness objective, and that the resulting circuits are competitive with direct attribution on the target. For this to hold as transfer rather than target-only optimization, the projection must be constrained so that source scores remain load-bearing (e.g., via limited-capacity mapping, regularization that penalizes deviation from source relative importances, or architectural tying). If the mapping is a general differentiable function between source and target node sets, gradient descent on faithfulness alone can reassign importances to maximize the objective without preserving source structure. The abstract reports weaker recovery for larger gaps and on Qwen-2.5, consistent with cases where no faithful alignment exists and the optimizer defaults to target-only solutions. This makes the mechanistic-prior claim rest on unstated details of the alignment parameterization and any auxiliary losses.","agreement_with_reader":"partial"},"referee_report":{"model":"grok-4.3","summary":"The paper introduces Differentiable Faithfulness Alignment (DFA), a framework that transfers circuit information from a smaller source LM to a larger target LM by projecting source node importance scores into the target via a learned differentiable mapping trained against a soft faithfulness objective. This avoids full circuit discovery on the target. Evaluations on Llama-3 (1B→3B) and Qwen-2.5 models across factual retrieval, multiple-choice reasoning, and arithmetic tasks show DFA often competitive with or exceeding direct node attribution on smaller gaps, outperforming simple baselines, with zero-shot transfer effective in some cases; performance degrades for larger gaps and on Qwen-2.5.","tokens_in":1960,"tokens_out":630,"duration_ms":31039,"significance":"If the transfer interpretation holds, DFA offers a scalable route to mechanistic interpretability by using smaller models as mechanistic priors for larger ones, reducing the cost of circuit discovery. The empirical results on Llama-3 pairs and the code release are concrete strengths; the observed limits with scale/architectural gaps provide useful boundary conditions for the approach.","major_comments":[{"comment":"The central transfer claim requires that source node scores remain load-bearing after alignment. The manuscript does not specify the parameterization or capacity of the learned mapping (e.g., whether it is a low-rank or tied function) nor any auxiliary loss that penalizes deviation from source relative importances. Without such constraints, gradient descent on the faithfulness objective alone can reassign importances to maximize the target objective independently of the source, turning DFA into target-only optimization. This must be clarified with an explicit definition of the mapping and a control experiment replacing source scores with random or uniform values.","section":null},{"comment":"Recovery is reported as weaker for larger source-target gaps and substantially lower on Qwen-2.5. The manuscript should quantify this with per-task faithfulness scores, statistical significance tests, and an ablation showing whether the drop is due to architectural mismatch or failure of the alignment to preserve source structure (e.g., correlation between source and aligned scores before/after training).","section":null},{"comment":"The soft faithfulness objective is central to training but its precise formulation, including any post-hoc choices in evaluation metrics or circuit extraction thresholds, is not detailed enough to assess whether the reported competitiveness with direct attribution is robust or sensitive to those choices.","section":null}],"minor_comments":[{"comment":"The abstract and introduction would benefit from a clearer statement of the exact node sets being aligned (e.g., attention heads, MLPs, or residual streams) and how the projection handles differing layer counts between source and target.","section":null},{"comment":"Table or figure captions should explicitly state the number of runs, random seeds, and error bars for all reported faithfulness numbers to allow assessment of variability.","section":null}],"recommendation":"major_revision","confidential_remarks":"The citation pattern appears light on prior cross-model interpretability work; the authors should ensure they engage with relevant circuit transfer or model distillation literature even if the differentiable alignment angle is novel."},"author_rebuttal":{"model":"grok-4.3","summary":"We thank the referee for the constructive comments, which help clarify the transfer mechanism and improve the empirical rigor of the work. We address each major comment below and will incorporate revisions to strengthen the manuscript.","responses":[{"response":"We agree that an explicit definition of the mapping and a control experiment are necessary to substantiate that source scores remain load-bearing. The current manuscript describes the mapping as a learned differentiable projection of source node importance scores but does not detail its parameterization. In the revision we will add a precise definition (including capacity and any weight-tying) in the Methods section. We will also include the requested control experiment that replaces source scores with random or uniform values and reports the resulting drop in target faithfulness, thereby demonstrating that the alignment depends on source-derived information rather than independent target optimization.","revision_made":"yes","referee_comment":"The central transfer claim requires that source node scores remain load-bearing after alignment. The manuscript does not specify the parameterization or capacity of the learned mapping (e.g., whether it is a low-rank or tied function) nor any auxiliary loss that penalizes deviation from source relative importances. Without such constraints, gradient descent on the faithfulness objective alone can reassign importances to maximize the target objective independently of the source, turning DFA into target-only optimization. This must be clarified with an explicit definition of the mapping and a control experiment replacing source scores with random or uniform values."},{"response":"We will expand the results section to report per-task faithfulness scores for all model pairs and tasks. Where multiple runs are available we will add statistical significance tests. We will further include an ablation that computes the Pearson correlation between source node scores and the aligned scores both before and after training; this will help distinguish whether performance degradation stems from architectural mismatch or from the alignment failing to preserve source structure.","revision_made":"yes","referee_comment":"Recovery is reported as weaker for larger source-target gaps and substantially lower on Qwen-2.5. The manuscript should quantify this with per-task faithfulness scores, statistical significance tests, and an ablation showing whether the drop is due to architectural mismatch or failure of the alignment to preserve source structure (e.g., correlation between source and aligned scores before/after training)."},{"response":"We will provide the complete mathematical formulation of the soft faithfulness objective in the Methods section, together with the exact post-hoc choices used for evaluation metrics and any circuit extraction thresholds. This added detail will allow readers to assess robustness and will be accompanied by a brief sensitivity analysis where feasible.","revision_made":"yes","referee_comment":"The soft faithfulness objective is central to training but its precise formulation, including any post-hoc choices in evaluation metrics or circuit extraction thresholds, is not detailed enough to assess whether the reported competitiveness with direct attribution is robust or sensitive to those choices."}],"tokens_in":1537,"tokens_out":612,"duration_ms":37517,"standing_objections":[]},"desk_editor":{"model":"grok-4.3","letter":"The paper's main contribution is a learned mapping that takes node importance scores from a smaller source model and projects them into a larger target, then optimizes the mapping parameters with a soft faithfulness objective on the target. This is meant to supply mechanistic priors without running full circuit discovery on the big model. They test it on Llama-3 1B-to-3B and Qwen-2.5 pairs across factual, reasoning, and arithmetic tasks, with code released at the GitHub link in the abstract.","headline":"DFA gives a workable differentiable route for aligning node scores across model sizes, but the transfer story rests on unshown details of how much the source actually constrains the solution.","tokens_in":2413,"tokens_out":181,"would_cite":false,"duration_ms":23679,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"grok-4.3","headline":"A differentiable alignment transfers node importance scores from smaller language models to larger ones by optimizing a soft faithfulness objective.","keywords":["differentiable faithfulness alignment","cross-model circuit transfer","mechanistic interpretability","language model circuits","node importance alignment","faithfulness objective","circuit discovery","model scaling"],"falsifier":"If DFA-aligned circuits on a new source-target pair produce substantially lower faithfulness scores or task accuracy than circuits found by direct attribution on the target, the transfer method would not succeed.","tokens_in":2652,"feed_emoji":"","tokens_out":774,"duration_ms":42678,"temperature":0.7,"pith_summary":"The paper presents Differentiable Faithfulness Alignment as a way to move circuit information between language models without running full circuit discovery on the larger target. It learns a mapping that takes importance scores from a small source model and projects them onto the target, then tunes the mapping so the transferred circuits keep the target model's behavior on tasks like factual retrieval and arithmetic. This avoids the high cost of direct methods on big models and uses the source as a prior. The evaluations show the approach works best when the source and target are close in size and architecture, such as Llama-3 1B to 3B. It beats simple baselines and sometimes matches or exceeds direct attribution in faithfulness.","feed_headline":"Differentiable alignment transfers circuits from small to large language models","feed_subtitle":"DFA learns a mapping of node importance scores and tunes it with a soft faithfulness loss, recovering target circuits competitively on Llama","key_machinery":"Differentiable Faithfulness Alignment (DFA), a learned mapping from source node importance scores to the target model that is optimized by a soft faithfulness loss to preserve task performance.","core_discovery":"We introduce Differentiable Faithfulness Alignment (DFA), a framework that transfers circuit information from a smaller source model to a larger target model through a learned differentiable alignment. DFA projects source-model node importance scores into the target model and trains this mapping with a soft faithfulness objective, avoiding full circuit discovery on the target model. We evaluate DFA on Llama-3 and Qwen-2.5 across six tasks spanning factual retrieval, multiple-choice reasoning, and arithmetic. The strongest results occur on Llama-3 1B to 3B, where aligned circuits are often competitive with direct node attribution and zero-shot transfer remains effective.","pith_inferences":["The method could lower the cost of circuit discovery for very large models by reusing smaller ones as starting points.","It suggests mechanistic priors might be chained across successive model scales as training progresses.","Architecture differences appear to limit transfer, pointing to a need for alignment methods that account for model family structure.","The soft faithfulness objective might extend to transferring other interpretability signals such as activation patterns beyond node importance."],"forward_implications":["On Llama-3 1B to 3B, DFA circuits are competitive with or stronger than direct node attribution in faithfulness for the tested tasks.","Zero-shot transfer of the learned alignment works without additional training in some settings.","Transfer performance drops for larger size gaps and is substantially weaker on Qwen-2.5 than on Llama-3.","DFA outperforms simple baselines across factual retrieval, reasoning, and arithmetic tasks.","In some cases the method recovers target-model circuits whose faithfulness matches or exceeds that of direct attribution."],"fun_headline_variants":["DFA projects source circuits into larger target models","Learned alignment tunes faithfulness for circuit recovery","Circuit transfer succeeds on Llama-3 1B to 3B with DFA","Soft faithfulness objective aligns node importance scores"],"cache_read_input_tokens":64,"weakest_assumption_plain":"Node importance scores from a smaller source model contain information that can be aligned differentiably to serve as useful mechanistic priors for a larger target without full circuit discovery on the target.","fun_headline_variants_meta":{"raw":{"variants":["DFA projects source circuits into larger target models","Learned alignment tunes faithfulness for circuit recovery","Circuit transfer succeeds on Llama-3 1B to 3B with DFA","Soft faithfulness objective aligns node importance scores"]},"model":"grok-4.3","cost_usd":0.009862,"raw_usage":{"total_tokens":4346,"prompt_tokens":749,"num_sources_used":0,"completion_tokens":61,"cost_in_usd_ticks":98615500,"prompt_tokens_details":{"text_tokens":749,"audio_tokens":0,"image_tokens":0,"cached_tokens":64},"completion_tokens_details":{"audio_tokens":0,"reasoning_tokens":3536,"accepted_prediction_tokens":0,"rejected_prediction_tokens":0}},"tokens_in":749,"tokens_out":61,"duration_ms":46501,"temperature":1.0,"reasoning_tokens":3536,"cache_read_input_tokens":64,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-05-08T03:38:03.546675+00:00","model_set":{"reader":"grok-4.3"},"falsifier":"If DFA-aligned circuits on a new source-target pair produce substantially lower faithfulness scores or task accuracy than circuits found by direct attribution on the target, the transfer method would not succeed.","supporting_citations":[],"review_version":1}