REINA: Dual Frameworks in Machine Learning

• REINA is defined as two distinct frameworks in NLP: one augments training with BM25 retrieval of labeled examples, while the other employs information-theoretic policy learning for simultaneous translation.
• The retrieval-based approach enhances tasks like summarization, translation, and QA by efficiently indexing the training set and fine-tuning on augmented inputs.
• The SimulST framework uses a policy network to decide between reading additional audio or writing output tokens, with REINA-SAN and REINA-TAN improving temporal grounding.

Searching arXiv for papers titled or containing “REINA” to ground the article in current literature. REINA is an overloaded designation in recent arXiv literature. In machine learning, it denotes two unrelated frameworks: “REtrieving from the traINing datA”, a retrieval-augmented training and inference method that indexes only labeled training instances for NLU and NLG tasks, and “Regularized Entropy INformation Adaptation”, an information-theoretic policy-learning framework for simultaneous speech translation (SimulST) that decides when to READ more audio and when to WRITE the next token (Wang et al., 2022, Hirschkind et al., 7 Aug 2025). A subsequent SimulST line extends the latter with REINA-SAN and REINA-TAN, which introduce temporal grounding to reduce unstable “read loop” behavior (Liu et al., 10 Apr 2026). Outside NLP, the string also appears in unrelated contexts, including the Reina–Treves transformation in general relativity and Museo Reina Sofía in cultural data analysis (Astorino, 2019, Solà et al., 2023).

1. Terminological scope

The principal source of ambiguity is that the same acronym is used for different technical programs. The retrieval-based REINA addresses supervised NLP tasks such as summarization, machine translation, language modeling, and multiple-choice question answering, whereas the SimulST REINA addresses streaming decision-making under a latency–quality trade-off. They share neither objective functions nor model architectures; the commonality is nominal rather than methodological (Wang et al., 2022, Hirschkind et al., 7 Aug 2025).

Usage	Expansion	Core setting
REINA	REtrieving from the traINing datA	BM25 retrieval over the labeled training set for NLU/NLG
REINA	Regularized Entropy INformation Adaptation	READ/WRITE policy learning for SimulST
REINA-SAN / REINA-TAN	Temporal variants of SimulST REINA	Weak alignment supervision or timestep embeddings

A frequent misconception is to treat REINA as a single research program. The literature instead supports two distinct meanings: a retrieval-based augmentation scheme and an information-based streaming policy.

2. REINA as retrieval from labeled training data

The retrieval-based formulation starts from the observation that retrieval need not target a large external corpus. REINA indexes only the task’s labeled training set, arguing that the training distribution is closest to test inputs and that sparse retrieval over this smaller corpus reduces both indexing and runtime costs (Wang et al., 2022). The method differs from typical RAG in three ways stated explicitly in the paper: corpus scope is limited to the labeled training set, supervised fine-tuning is performed end-to-end on augmented inputs, and efficiency comes from sparse BM25 retrieval rather than dense dual-encoder training.

Formally, the training set is $T=\{(x_i,y_i)\}$, retrieval uses sparse BM25 ranking, and the retrieved set is

$$R(x)=\operatorname{TopK}\left\{(x_i,y_i)\in T : s_{\mathrm{BM25}}(x,x_i)\right\}.$$

The augmented input is

$$z(x)=\left[x;\operatorname{concat}_{(x_i,y_i)\in R(x)} g(x_i,y_i)\right].$$

During training, the current instance is excluded from retrieval to prevent leakage; during inference, no filtering is applied. For generation tasks, the objective is the standard negative log-likelihood $-\log p_\theta(y\mid z(x))$, while multiple-choice QA uses a choice-level softmax objective over per-choice augmented inputs (Wang et al., 2022).

A central technical feature is that the formatting function $g(x_i,y_i)$ is task-specific. For summarization, the index keys are full documents and the augmentation appends only retrieved summaries. For language modeling, text is chunked into sequences $C_i$, the value is $C_{i+1}$, and retrieved next chunks are prepended to the current chunk. For machine translation, the key is the source sentence and augmentation appends retrieved target sentences only. For multiple-choice QA, retrieval is conditioned on $[x;c_j]$, and the augmented input concatenates retrieved question–answer pairs; ConceptNet relations and Wiktionary definitions can optionally be integrated for some datasets (Wang et al., 2022).

This design situates REINA between classical retrieval augmentation and in-context learning. The paper explicitly characterizes it as using the same basic idea as attaching labeled exemplars, but with supervised training so that the model learns to exploit those exemplars. The method is implemented with BM25 via Lucene/PyLucene, supports long queries such as documents, and avoids FAISS- or DPR-style dense retrieval infrastructure (Wang et al., 2022).

3. REINA as information-theoretic policy learning for SimulST

In simultaneous speech translation, the system must decide at each state $(n,t)$ whether to READ more source audio or WRITE the next target token. REINA formulates this decision in terms of the information gain expected from waiting. With source audio $a=(a_1,\dots,a_T)$ and target sequence $$R(x)=\operatorname{TopK}\left\{(x_i,y_i)\in T : s_{\mathrm{BM25}}(x,x_i)\right\}.$$0, the information gain for the next token is defined as

$$R(x)=\operatorname{TopK}\left\{(x_i,y_i)\in T : s_{\mathrm{BM25}}(x,x_i)\right\}.$$1

Using mutual-information identities, the paper rewrites this as a difference of conditional entropies and then approximates it with teacher-forced token log-likelihoods under full versus partial audio: $$R(x)=\operatorname{TopK}\left\{(x_i,y_i)\in T : s_{\mathrm{BM25}}(x,x_i)\right\}.$$2 The idealized policy is to READ if this gain exceeds a threshold and WRITE otherwise (Hirschkind et al., 7 Aug 2025).

Inference cannot access future audio, so REINA trains a differentiable policy network $$R(x)=\operatorname{TopK}\left\{(x_i,y_i)\in T : s_{\mathrm{BM25}}(x,x_i)\right\}.$$3 to correlate with the batch-normalized information-gain surrogate. The core loss is a covariance-style objective,

$$R(x)=\operatorname{TopK}\left\{(x_i,y_i)\in T : s_{\mathrm{BM25}}(x,x_i)\right\}.$$4

supplemented by a monotonicity regularizer

$$R(x)=\operatorname{TopK}\left\{(x_i,y_i)\in T : s_{\mathrm{BM25}}(x,x_i)\right\}.$$5

and an $$R(x)=\operatorname{TopK}\left\{(x_i,y_i)\in T : s_{\mathrm{BM25}}(x,x_i)\right\}.$$6 term

$$R(x)=\operatorname{TopK}\left\{(x_i,y_i)\in T : s_{\mathrm{BM25}}(x,x_i)\right\}.$$7

The full policy objective is

$$R(x)=\operatorname{TopK}\left\{(x_i,y_i)\in T : s_{\mathrm{BM25}}(x,x_i)\right\}.$$8

At runtime, the decision rule is

$$R(x)=\operatorname{TopK}\left\{(x_i,y_i)\in T : s_{\mathrm{BM25}}(x,x_i)\right\}.$$9

The paper emphasizes that the policy is trained by standard backpropagation and that no reinforcement learning is used (Hirschkind et al., 7 Aug 2025).

The training pipeline has three stages. Stage 1 trains a strong non-streaming S2TT teacher with multitask cross-entropy on ASR, NMT, and S2TT data. Stage 2 fine-tunes the teacher on randomly truncated audio, with 80% truncated and 20% full inputs, to stabilize partial-audio likelihoods. Stage 3 freezes the base model and trains only the policy head with $$z(x)=\left[x;\operatorname{concat}_{(x_i,y_i)\in R(x)} g(x_i,y_i)\right].$$0. The reported architecture uses a Whisper Medium acoustic encoder, a 16-layer Transformer text decoder, and a 2-layer Transformer policy network of approximately $$z(x)=\left[x;\operatorname{concat}_{(x_i,y_i)\in R(x)} g(x_i,y_i)\right].$$1M parameters (Hirschkind et al., 7 Aug 2025).

4. Temporal grounding: REINA-SAN and REINA-TAN

A later paper identifies a failure mode in information-based SimulST policies: they can lack temporal context and enter a “read loop,” defined as the degenerate inference state where the system repeatedly predicts READ despite increasing $$z(x)=\left[x;\operatorname{concat}_{(x_i,y_i)\in R(x)} g(x_i,y_i)\right].$$2, causing it to ingest all audio before producing any tokens (Liu et al., 10 Apr 2026). Empirical analysis attributes this to temporal drift in the policy head: without an internal clock, the emission probability can remain too flat over time.

Two remedies are proposed while preserving the decoupled design and keeping Whisper Large V3 frozen. REINA-SAN augments REINA with weak monotonic supervision derived from LLM-generated text alignments and WhisperX forced alignment. The target emission boundary for token $$z(x)=\left[x;\operatorname{concat}_{(x_i,y_i)\in R(x)} g(x_i,y_i)\right].$$3 is $$z(x)=\left[x;\operatorname{concat}_{(x_i,y_i)\in R(x)} g(x_i,y_i)\right].$$4, and the supervision uses a soft label

$$z(x)=\left[x;\operatorname{concat}_{(x_i,y_i)\in R(x)} g(x_i,y_i)\right].$$5

with loss

$$z(x)=\left[x;\operatorname{concat}_{(x_i,y_i)\in R(x)} g(x_i,y_i)\right].$$6

and hybrid objective

$$z(x)=\left[x;\operatorname{concat}_{(x_i,y_i)\in R(x)} g(x_i,y_i)\right].$$7

The paper characterizes the BCE term as a global monotonic “guardrail” that reduces read loops markedly, especially at higher latency (Liu et al., 10 Apr 2026).

REINA-TAN instead injects explicit continuous time embeddings into the policy head, described as sinusoidal embeddings over elapsed audio time tuned for speech durations using base 100. These embeddings are broadcast-added to decoder hidden states,

$$z(x)=\left[x;\operatorname{concat}_{(x_i,y_i)\in R(x)} g(x_i,y_i)\right].$$8

so that the policy receives an internal clock. The paper reports that TAN yields more uniform and earlier emissions throughout the sentence and slightly better Pareto frontiers than SAN, while SAN is more robust against read loops (Liu et al., 10 Apr 2026).

The implementation remains lightweight. The backbone is Whisper Large V3, kept frozen, and the policy head is a 3-layer Transformer with dimension $$z(x)=\left[x;\operatorname{concat}_{(x_i,y_i)\in R(x)} g(x_i,y_i)\right].$$9 and feedforward dimension $-\log p_\theta(y\mid z(x))$0. Training runs for 4 epochs on $-\log p_\theta(y\mid z(x))$1A100-80G GPUs in approximately 5 hours. Inference uses streaming beam search with beam size $-\log p_\theta(y\mid z(x))$2, chunk size $-\log p_\theta(y\mid z(x))$3 ms, and threshold $-\log p_\theta(y\mid z(x))$4 to sweep the latency–quality operating point (Liu et al., 10 Apr 2026).

5. Metrics, Pareto frontiers, and empirical performance

The two REINA lines are evaluated in very different regimes. The retrieval-based REINA reports task metrics standard to summarization, MT, LM, and QA. It achieves state-of-the-art results on XSum and BigPatent and leaderboard No. 1 on CommonsenseQA. Reported results include approximately $-\log p_\theta(y\mid z(x))$5 Rouge-1/2/L on XSum with REINA+BART-large, approximately $-\log p_\theta(y\mid z(x))$6 on BigPatent, WikiText103 perplexity of approximately $-\log p_\theta(y\mid z(x))$7 versus $-\log p_\theta(y\mid z(x))$8 for BART-large, WMT16 en$-\log p_\theta(y\mid z(x))$9tr SacreBLEU of approximately $g(x_i,y_i)$0 versus $g(x_i,y_i)$1 for mBART, and CommonsenseQA test accuracy of $g(x_i,y_i)$2 (Wang et al., 2022). The reported gains are strongest when the training set is large enough that nearest neighbors contain useful label-bearing signal; on WikiText2, by contrast, REINA+BART-large does not improve over the baseline.

The SimulST REINA line is organized around the latency–quality frontier. The original paper evaluates with BLEU, AL or LAAL, and the proposed Normalized Streaming Efficiency (NoSE),

$g(x_i,y_i)$3

NoSE measures how much of the non-streaming BLEU is retained over a common latency interval $g(x_i,y_i)$4. The paper reports that REINA improves the latency–quality trade-off by as much as $g(x_i,y_i)$5 compared to prior approaches, normalized against non-streaming baseline BLEU scores, and reports state-of-the-art streaming results for comparable-sized models trained on open-source or synthetically generated data (Hirschkind et al., 7 Aug 2025).

The temporal-awareness extensions shift evaluation to LAAL, BLEU, XComet-XL, and NoSE on Whisper Large V3. Across FLEURS and Europarl for de/fr/es$g(x_i,y_i)$6en, REINA-TAN yields the strongest Pareto frontier and REINA-SAN remains close behind. The paper reports NoSE gains up to $g(x_i,y_i)$7 over competitive baselines, with examples on FLEURS of de$g(x_i,y_i)$8en improving from SeamlessStreaming $g(x_i,y_i)$9 to Whisper-REINA-TAN $C_i$0 within latency bounds $C_i$1 s, fr$C_i$2en improving from $C_i$3 to $C_i$4 within $C_i$5 s, and es$C_i$6en improving from $C_i$7 to $C_i$8 within $C_i$9 s (Liu et al., 10 Apr 2026).

These results support a consistent interpretation across the SimulST papers: explicit policy learning can improve streaming efficiency without modifying the core translation backbone, and temporal grounding is the main ingredient missing from purely information-based policies.

6. Limitations, operating regimes, and unrelated homographs

The retrieval-based REINA is deliberately simple, but its own paper records several risks. Retrieving labeled examples can raise memorization and privacy concerns; contamination can inflate evaluation if retrieved labels overlap heavily with gold outputs; domain shift can make nearest neighbors misleading; and the added retrieval plus longer context windows impose some computational overhead. The paper also notes that small datasets can weaken the neighbor signal, as illustrated by the lack of improvement on WikiText2 (Wang et al., 2022).

The SimulST REINA family has a different set of limitations. The original framework depends on a strong teacher, especially under partial audio, making truncated-audio fine-tuning critical. The policy threshold $C_{i+1}$0 or $C_{i+1}$1 must be swept empirically to expose the Pareto frontier. Information-based policies can still exhibit rare read loops under weak temporal signals or conservative thresholds. In the 2026 extensions, REINA-SAN becomes sensitive to alignment quality from WhisperX and Qwen3-32B, while REINA-ALL can degrade the frontier because TAN and SAN may introduce conflicting inductive biases. The study is also limited to fr/de/es$C_{i+1}$2en for the Whisper setting, with explicit caution that broader language coverage and domain shifts may require retuning thresholds and loss weights (Hirschkind et al., 7 Aug 2025, Liu et al., 10 Apr 2026).

Beyond NLP, “Reina” functions as a homograph rather than an acronymic family. It appears as a coauthor name in noncommutative geometry, as in the analysis of a quantum instanton bundle due to Landi, Pagani and Reina (Zegers et al., 2020); as the Reina–Treves transformation in axisymmetric Einstein–Maxwell theory (Astorino, 2019); as Museo Reina Sofía in quantitative museum studies, where the institution is reported to have a collection lag of approximately $C_{i+1}$3 years (Solà et al., 2023); and as Reina-Campos in astrophysical work comparing JWST globular-cluster photometry with EMP-Pathfinder simulations (Harris et al., 2024). These usages are etymologically unrelated to the NLP acronym.

In current arXiv usage, therefore, REINA is best understood not as a single framework but as a technically overloaded label. In NLP it names either a training-set retrieval paradigm or an information-theoretic SimulST policy, and the distinction is substantive: one augments inputs with retrieved labeled exemplars, while the other learns when streaming systems should wait for more source information before emitting target tokens.