Safin-1.5: A safety-native architecture and verifiable reasoning for recursive safety improvement
Safin-1.5: Safety-Native Architecture and Verifiable Reasoning for Recursive Safety Improvement
Shanghai AI Laboratory proposes Safin-1.5: safety architecture + recursive safety improvement + formal reasoning harness; the authors report H-CoT ASR dropping from 82.00% to 14.00%.
The paper is not an attack paper and has no explicit threat model. Safety goals: resistance to attacks within the red-team pool (resistance, ASR on WildJailbreak/HarmBench) and to unseen attack mechanisms outside the pool (resilience, H-CoT hijacking of the safety reasoning chain, X-Teaming adaptive multi-turn attacks); and excision of WMDP-Bio/Cyber dangerous knowledge via REX while preventing its recovery in subsequent updates.
- In recursive self-improvement (RSI), model updates introduce new risks and weaken existing safety measures, while existing self-improvement methods mainly optimize benchmark performance and do not incorporate safety into the evolution loop; the authors, following the AI-45° Law, argue that safety should co-evolve with capability.
- Safin-1.5 consists of three parts: a safety architecture combining MARCH persistent safety memory with REX dangerous knowledge decoupling (migrating dangerous knowledge into removable experts); ReSI, a three-stage discover-evaluate-repair recursive safety improvement loop; and POGB, an inference-time formal reasoning harness that uses semantic bridging to fill formalization gaps and produce checkable proofs.
- Compared with Safin-1, H-CoT ASR drops from 82.00% to 14.00%, X-Teaming from 64.78% to 32.08%, and WMDP-Bio from 82.56% to 43.68%; XSTest benign compliance drops from 88.40% to 86.80%; POGB achieves an average accuracy of 91.80% on four reasoning benchmarks.
- The authors plan to examine in the future safety preservation over longer improvement trajectories and formalized safety decisions under explicit rules. The experiments report only one round of ReSI; the frontier models in the safety comparison are 4; the paper does not report the judge model or the number of repetitions used for ASR determination.
- Evaluated models
- Safin-1.5
- Benchmarks
- WildJailbreakHarmBenchH-CoTX-TeamingXSTestMMLU-ProIFEvalWMDP-BioWMDP-CyberFOLIOProofWriterPrOntoQAProverQA-Hard
- Metrics
- ASRbenign complianceMMLU-Pro accuracyIFEval loose prompt-level accuracyWMDP multiple-choice accuracyreasoning accuracy
Safin-1.5 adopts a safety-native architecture and recursive safety improvement training, paired with a formal reasoning component at inference time, and experiments claim its safety and reasoning performance outperform frontier models. The architecture combines a persistent safety state and dangerous knowledge decoupling; recursive improvement iterates in a "discover-evaluate-fix" loop, accepting updates only when safety improves, general capabilities are maintained, and excessive refusal is limited; the formal component uses symbolic reasoning, verification, and semantic bridging to generate checkable proofs.
Paper analysis
What problem does this paper attempt to solve?
To address the problem that safety does not grow in step with capability in RSI—safety knowledge being overwritten and dangerous knowledge being hard to isolate—the authors propose Safin-1.5.
Core problem: In recursive self-improvement (RSI), the continuous growth of model capability does not automatically bring about synchronized growth in safety, and existing self-improvement methods mainly optimize benchmark performance without systematically incorporating safety into the evolution loop.
- Problem setting: The authors cite prior work stating that a single fine-tuning can weaken existing alignment, and that repeated updates in recursive evolution may introduce new vulnerabilities; the system cards of GPT-5.6 and Claude 4 have begun evaluating capability alongside risks such as misuse and prompt injection, echoing the AI-45° Law previously proposed by the authors (safety should co-evolve with capability).
- Two shortcomings of existing architectures (per the paper): First, safety knowledge resides in shared parameters, and subsequent updates overwrite it, so repaired risks are forgotten; second, dangerous knowledge and general knowledge are mixed in shared parameters, so parameter-level interventions may harm general capability, and subsequent fine-tuning can restore suppressed dangerous knowledge.
- Shortcomings of empirical alignment (per the paper): The model learns safe responses from risk data, but does not necessarily establish the basis for why the responses satisfy safety requirements; existing neuro-symbolic methods may lose implicit premises or semantic relations when formalizing natural language, preventing the solver from deriving the expected conclusion.
- Proposed approach: Safin-1.5, comprising three parts—the safety architecture (MARCH persistent safety memory + REX dangerous knowledge decoupling), the recursive safety improvement training loop ReSI, and the inference-time formal reasoning harness POGB; its predecessor Safin-1 was a preliminary exploration of Safety from Within.
- Protection goals: The two goals of resisting known attacks in the red-team pool (resistance) and unseen risks outside the pool (resilience) (attributed by the paper to R2 AI), while maintaining utility, limiting over-refusal, and excising WMDP-Bio/Cyber dangerous knowledge in the deployed model and preventing it from being restored by updates.
What related research exists?
Related work covers four groups: RSI and safety, parameter-level unlearning, alignment methods, and neuro-symbolic reasoning; the baselines are A3 and RMU.
RSI and the AI-45° Law
- The RSI paradigm (Chen et al. 2026, Duan et al. 2026, Dream-RSI, etc.)—the model becomes the main driver of its own iterative improvement; the authors state that these works do not incorporate safety into the evolution loop.
- SafeWork-R1 (2025)—the authors previously proposed the AI-45° Law, advocating co-evolution of safety and capability; R2 AI (2025/2026) proposed the two safety goals of resistance and resilience, which this paper's ReSI follows.
- GPT-Red (OpenAI, 2026)—self-play-trained red teams generate attacks to improve the prompt injection robustness of subsequent models, which the authors cite as an example of continuous discovery-repair.
Fragility of Alignment and Dangerous Knowledge Removal
- Qi et al. (2024, 2025)—the authors cite that fine-tuning breaks existing safety alignment and that alignment resides only at the level of a few tokens.
- RMU (WMDP, 2024) and NPO (2024)—parameter-level unlearning methods that modify shared parameters in place; the authors state that more thorough removal comes at the cost of general capability, and that subsequent fine-tuning can restore suppressed knowledge (citing Doshi and Stickland 2024).
Alignment and Red-Team Methods
- A3 (2026)—pairs harmful requests with pattern-similar benign requests for SFT; SInternal (2026)—fine-tunes a model to reproduce an expert's critique of its own responses and predict safety judgments; AlphaAlign (2025)—uses reinforcement learning to reward safe responses without over-refusing benign requests. The three form ReSI's alignment method pool.
- Red-team and evaluation resources: WildJailbreak (2024), HarmBench (2024), H-CoT (2025, hijacking safety reasoning chains), X-Teaming (2025, adaptive multi-turn attacks), XSTest (2023), WMDP (2024), DARWIN and JailbreakSkill (red-team pool).
Neuro-Symbolic Reasoning
- Logic-LM (2023) and Draft, Sketch, and Prove (2023)—use symbolic solvers to assist LLM reasoning; the authors state that natural language formalization loses implicit premises or semantic relations, and that wording changes cause inconsistent formalization, preventing the solver from establishing the expected conclusion.
Baselines and Positioning: The directly compared baselines are the fixed alignment method A3 (Table 5) and the same RMU recipe applied directly to the original model (R and R-SFT in Table 4). The authors position this paper as an extension of Safin-1: co-designing safety memory, dangerous knowledge decoupling, continual training, and formal guarantees into a full-stack framework.
How does the paper address this problem?
Three components: the MARCH safety memory + REX excisable-expert architecture, the three-stage ReSI training loop, and the POGB formal reasoning harness.
Overall approach: at the architecture level, use MARCH to persistently store safety experience and REX to migrate dangerous knowledge into physically excisable experts; at the training level, use ReSI to loop through discovering and repairing vulnerabilities; at the inference level, use POGB to combine language understanding with symbolic solving and produce machine-checkable proofs (Figure 2).
MARCH: Safety Memory (inherited from Safin-1)
- State anchors: The base mixer continuously updates the recurrent state, and MARCH archives anchors at boundaries every C text tokens, with each anchor saving the accumulated prefix state; anchors are read out via a shared embedding and generate routing keys.
- Persistent safety state: The memory bank additionally contains a learned state Psafe, which each layer assembles into a Safety State, available from the first token and occupying no input position; Safin-1 trains it with token-level cross-entropy under a frozen backbone.
- Joint routing readout: The candidate set contains context anchors, the safety state, and an empty option (allowing bypassing retrieval); after routing scoring, the readout is added to the native recurrent path by weight, without changing the base recurrent update.
REX: Dangerous Knowledge Decoupling
- Expert expansion: Source experts are selected by the difference in the frequency with which experts are selected into the top-K on target data versus general data; their parameters and routing rows are copied to form excisable experts, and the source experts are scaled by a fixed factor γ.
- Selective gradient masking: On target data, only the excisable experts and routing are updated; on general data, only the original experts and routing are updated, and general tokens mask the excisable experts' logits before top-K selection; the authors state that this makes the original experts alone carry general capability, while existing dangerous knowledge fades for lack of training signal.
- Routing differentiation and objective: A mass loss constrains the upper and lower bounds of the excisable experts' routing mass on target/general tokens, and a rank loss makes the excisable experts' maximum logit on general tokens lower than the original experts' top-K boundary; the total objective is LREX = LLM + λ1 Lmass + λ2 Lrank, and during training the attention, embeddings, normalization, shared experts, and LM head are frozen.
- Excision and dual-form delivery: The excisable experts are excised and a lightweight RMU cleaning is performed to obtain the public base A2; subsequent updates first reassemble the retained experts with A2 into the full structure, train on the full structure, and excise again at deployment, yielding a full model for authorized users and a public model with no excisable-expert parameters.
- Safin-1.5 instantiation: Applied to Safin-1-35B-A3B, a single group of excisable experts carries both Bio and Cyber; per layer, 8 source experts are selected per domain and their union is padded to 16, each layer is expanded from 256 to 272 experts, γ=0.3; migration uses about 40M tokens of target data (Bio:Cyber about 4:1) and about 27M tokens of general data; cleaning uses a 32M token forgetting budget and updates only layers 7–9 MLP; after reassembly, standard post-training uses 250M tokens.
ReSI: Recursive Safety Improvement
- Recursive formalization: Each round treats the safety update as an automated research problem, searching for training recipes starting from the current target model Mr and the attacks that succeeded against it; if a recipe passes the Pareto gate, the update with the largest safety gain is promoted to the next round's target, otherwise Mr is retained and the process terminates.
- Three stages per round: Stage 1 uses the red-team method pool to construct attacks against Mr and retains the input-response pairs that elicit unsafe responses; Stage 2 runs low-budget pilot trials for each attack source × alignment method (using only the current safety data and default hyperparameters), screens by the safety gain on the held-out validation set gr(ω) = ASR(Mr) − ASR(Mrω), and performs conditional extended probing for sources with large attack volumes, shortlisting at most K_pilot recipes; Stage 3 trains at full scale and revises the recipe based on gating feedback.
- Pareto gate: Requires a positive safety gain, and benign compliance and instruction following no lower than the initial model M0 minus tolerance ε; the authors state that anchoring to M0 prevents small per-round losses from accumulating across rounds.
- Feedback revision (S1–S4): Expand the combination of attack sources, construct A3-style benign pairs, add general replay or lower the learning rate, and for qualified SFT recipes continue with AlphaAlign-style safety GRPO (DAPO-style optimization, with the revision counter reset); each path allows at most N_refine revisions.
- Applied to the safety architecture: ReSI trains the full model, but red-teaming, validation, and gating all act on its excised counterpart; each round measures dangerous knowledge accuracy on the excised checkpoint; the replay weight λ_rep works with MARCH to reduce forgetting.
POGB: Formal Reasoning Harness
- Applicability switch: Enabled only for logical reasoning tasks of the form "whether the conclusion follows from the premises"; when enabled, the LLM performs grounding autoformalization, translating the problem into first-order logic and maintaining a grounding table that records the type and natural language information of each formal element.
- Proof-obligation-guided search: The symbolic solver searches for proofs of the query and of its negation separately; unresolved obligations are decomposed by existential/universal/conjunction/disjunction, with type information constraining the permissible instantiations; when still unresolved, Algorithm 2 constructs candidate semantic bridges over support sets of at most two literals, which are filtered by semantic similarity and NLI (threshold τ_NLI=0.80) and then judged for acceptance by an LLM Gate, and residual obligations are derived by logical abduction and solved recursively.
- Termination and soundness: An ancestor-obligation check rejects circular reasoning; Appendix A.2 gives a soundness proof—the successfully returned bridge set both enables the solver to complete the proof and is entirely accepted by the Gate; the final output is Entailed/Contradicted/Unknown/Conflicted according to the proof results for the two goals.
What experiments did the paper conduct?
Compares the safety of Safin-1 with 4 frontier models, the excision persistence of REX vs RMU, ReSI vs A3, and the reasoning accuracy of POGB vs 8 models.
Experimental Setup
- Models under test: Safin-1.5 and Safin-1 (35B-A3B); the frontier models in the safety comparison are GPT-5.6-Luna, Claude Opus 5, GLM-5.3, and Kimi K3 (Table 2); the reasoning comparison additionally includes Gemini3-Flash, DeepSeek-V3.2, Qwen3.8-Flash, and GPT-5.5 (Table 3, each with Direct and CoT).
- ReSI instantiation: The safety data pool is initialized with the WildJailbreak training set; the red-team pool contains DARWIN and JailbreakSkill; the alignment method pool contains A3, SInternal, and AlphaAlign; general data is taken from Dolci-Instruct-SFT; all updates use LoRA (rank 32, α=64); the validation set is taken from a development panel consisting of WildJailbreak, H-CoT, XSTest, and IFEval.
- Metrics: resistance uses ASR on WildJailbreak and HarmBench (lower is better); resilience uses ASR on H-CoT and X-Teaming outside the red-team pool; benign compliance uses the full compliance rate on XSTest benign requests; general capability uses MMLU-Pro and IFEval (loose prompt-level accuracy); dangerous knowledge uses WMDP-Bio/Cyber four-choice accuracy (chance level 25%).
- Evaluation protocol: Safin-1.5 and the REX ablation variants are all evaluated after excision; the paper does not report how ASR is judged (rule-based/human/LLM judge), the sample size, or the number of repetitions.
- POGB settings: Each question samples at most 3 formalizations, stopping once a valid formal proof is found, and falling back to CoT if all fail; semantic similarity uses all-MiniLM-L6-v2, NLI uses roberta-large-mnli, τ_NLI=0.80; the four reasoning benchmark evaluation sets total 1,804 cases (FOLIO 204, ProofWriter 600, PrOntoQA 500, ProverQA-Hard 500).
Main Results
According to Table 1 (Safin-1 vs Safin-1.5, %) and Table 2 (frontier models, %):
Swipe horizontally to read the table
Model WJB↓ HB↓ H-CoT↓ X-Teaming↓ XSTest↑ GPT-5.6-Luna 1.90 0.00 8.00 56.69 100.00 Claude Opus 5 9.25 3.50 38.00 84.85 98.80 GLM-5.3 5.35 3.25 24.00 76.73 99.20 Kimi K3 4.50 3.75 18.00 77.80 99.60 Safin-1 2.10 0.25 82.00 64.78 88.40 Safin-1.5 1.20 1.25 14.00 32.08 86.80 - The authors state that Safin-1.5 has the lowest ASR in the table on X-Teaming and WildJailbreak, and on H-CoT is lower than Claude Opus 5, GLM-5.3, and Kimi K3 but higher than GPT-5.6-Luna; its XSTest benign compliance is the lowest in the table.
- Compared with Safin-1, Safin-1.5's HarmBench ASR rises from 0.25% to 1.25%; IFEval rises from 85.58% to 87.06%, and MMLU-Pro drops from 77.10% to 76.80% (Table 1).
- Dangerous knowledge: WMDP-Bio drops from 82.56% to 43.68%, and WMDP-Cyber from 58.33% to 33.17% (Table 1, evaluated after excision).
REX vs RMU Ablation (Table 4)
- Setup: On Safin-1-35B-A3B, R applies the same RMU operator, hyperparameters, and forgetting budget directly to the original model B; the three routes A3 (Safin-1+REX), A3-direct (trained directly on the excised A2), and R-SFT use the same 250M token full-parameter SFT, the same data order, and the same random seed.
- Excision effect: After 250M token post-training, the full model F3 is 82.8%/57.4% on WMDP-Bio/Cyber (B is 82.6%/58.3%), and its excised counterpart A3 drops to 40.8%/32.5%, with a 2.4 percentage point difference in MMLU.
- Persistence: The kept share (the proportion of removal retained after post-training) is 90.0% for A3 and 3.4% for R-SFT on Bio, and 89.9% and 43.6% on Cyber; A3-direct's WMDP-Bio rebounds to 64.0% (kept share 40.0%). The authors state that the gap between R-SFT and A3-direct comes from migration and excision, and the gap between A3-direct and A3 comes from training on the full structure and excising at deployment.
ReSI vs A3 Ablation (Table 5)
- Two starting points: On Safin-1, ReSI reduces WildJailbreak ASR from 2.10% to 0.35% and H-CoT from 82.00% to 6.00%, with HarmBench staying at 0.25%; on Safin-1+REX (+ReSI is Safin-1.5), WildJailbreak drops from 2.10% to 1.20%, H-CoT from 64.00% to 14.00%, and HarmBench changes from 1.00% to 1.25%.
- Preservation cost: After ReSI, XSTest drops by 2.80 percentage points on Safin-1 and by 8.40 percentage points on Safin-1+REX; MMLU-Pro changes by +2.30 and −2.40 percentage points respectively, and IFEval by −1.29 and −5.55 percentage points respectively.
- Comparison with A3: The authors state that on Safin-1, ReSI is better than A3 on all metrics; on Safin-1+REX, ReSI and A3 are on par on HarmBench and H-CoT, with ReSI better on WildJailbreak, XSTest, and IFEval, but lower on MMLU-Pro (A3 is 79.80%, ReSI is 76.80%).
POGB Reasoning Evaluation (Table 3)
- Results: Safin-1.5 achieves 88.73% on FOLIO (0.98 percentage points higher than the strongest baseline, Claude Opus 5 Direct), 99.50% on ProofWriter (0.33 percentage points lower than the best 99.83%), 100.00% on PrOntoQA (on par with multiple baselines), and 75.60% on ProverQA-Hard; the weighted average over 1,804 cases is 91.80%.
- Exception: The authors state that this average accuracy exceeds all evaluated baseline configurations except Claude Opus 5 (Direct 93.68%, CoT 92.29%); on ProverQA-Hard, Claude Opus 5 Direct is 84.00%, higher than Safin-1.5.
- Authors' interpretation: When the formal proof search succeeds, it produces a machine-checkable proof, and the conclusion can be independently verified within the scope of the formalized premises and accepted semantic bridges.
Other Ablations and Analyses
- After one round of ReSI, the excised model maintains 43.68%/33.17% on WMDP-Bio/Cyber (Table 1); the authors state that REX's removal effect carries over past this round of ReSI.
- A2 (after REX cleaning, before post-training) is 36.2% on WMDP-Bio, 20.7 percentage points lower than R with the same RMU recipe; A2 retains more than 96% of B's MMLU accuracy (Table 4).
What are some points that can be further explored?
The authors plan to examine safety preservation over longer improvement trajectories and formal reasoning for rule-based safety decisions; the experiments ran only one round of ReSI.
Limitations noted by the authors and future directions
- The authors plan to examine safety preservation over longer improvement trajectories (Conclusions).
- The authors plan to explore formal reasoning for safety decisions governed by explicit rules (Conclusions).
- The paper states that the applicability of formal guarantees depends on how faithfully the formal representation captures the semantics of the original problem, and that the scope of the guarantees is limited to the formal premises and the accepted semantic bridge (Introduction, Section 4).
Experimental coverage
- Models tested: the safety comparison includes Safin-1, Safin-1.5 and 4 frontier models (Table 2); the reasoning comparison includes Direct/CoT configurations of 8 models (Table 3).
- ReSI reports results for only one round: one round of safety updates starting from F3 (Section 5.3.1); the paper does not report experiments on multiple recursive rounds.
- The red-team pool contains two methods, DARWIN and JailbreakSkill; the resilience evaluation uses two attack mechanisms outside the pool, H-CoT and X-Teaming (Section 5.1).
- Dangerous knowledge removal covers only two domains, WMDP-Bio and WMDP-Cyber, carried by a single set of removable experts (Section 2.2).
- The paper does not report how ASR is judged (judge model or rules), the sample size of each benchmark, the number of repetitions, or statistical significance tests.
Summarize the main content of the paper
Safin-1.5 combines safety memory, dangerous knowledge excision, recursive repair, and verifiable reasoning into a full-stack framework, and the authors state that both safety and reasoning exceed frontier models.
Safin-1.5 from Shanghai AI Lab targets the problem that in RSI, safety does not grow in sync with capability, and per the AI-45° Law treats safety as an endogenous capability that co-evolves with capability.
- Method: The safety architecture = MARCH persistent safety state + REX, which migrates Bio/Cyber dangerous knowledge into excisable experts (each layer expanded from 256 to 272 experts, physically excised at deployment); ReSI repairs vulnerabilities in a three-stage loop of red-team evaluation, pilot screening, and full refinement, with a Pareto gate anchored to the initial model to limit over-refusal and capability loss; POGB uses semantic bridging at inference time to fill formalization gaps, with a symbolic solver producing checkable proofs.
- Safety: Compared with Safin-1, H-CoT ASR drops from 82.00% to 14.00%, X-Teaming from 64.78% to 32.08%, and WildJailbreak from 2.10% to 1.20%; the authors state that X-Teaming and WildJailbreak ASR are lower than all evaluated frontier models, but XSTest benign compliance of 86.80% is the lowest in Table 2, and HarmBench ASR rises from 0.25% to 1.25%.
- Dangerous knowledge: After excision, WMDP-Bio drops from 82.56% to 43.68% and WMDP-Cyber from 58.33% to 33.17%; after 250M token post-training, the kept share for Bio under the REX route is 90.0%, versus only 3.4% for R-SFT with the same RMU recipe (Table 4).
- Reasoning: POGB achieves an average accuracy of 91.80% on 1,804 cases across FOLIO, ProofWriter, PrOntoQA, and ProverQA-Hard, which the authors state exceeds all evaluated Direct/CoT baseline configurations except Claude Opus 5.
- Authors' conclusion: The results demonstrate progress toward the AI-45° Law goal (coordinated development of safety and capability); the authors plan to examine in the future safety preservation over longer improvement trajectories and formalized safety decisions under explicit rules.
Related papers
- On the Efficiency-Safety Dilemma in Large Reasoning ModelsAnalysis / interpretabilityarXiv