acceptodds
Under review as a conference paper at ICLR 2027

BRAG: BAYESIAN RETRIEVAL AUGMENTED GENERATION

Abstract

Retrieval-augmented generation (RAG) ranks memory chunks by semantic similarity, but similarity is not what matters to the consumer of a retrieved chunk: what matters is whether the chunk actually gets used to answer the query. We introduce BRAG, a Bayesian stochastic filter that maintains, for every chunk in a RAG database, a Beta posterior over its latent utility, the probability that it will be retrieved and used. From the same stream of retrieval events, the filter also rotates each chunk's semantic keys toward the queries it actually answers, modeled as a mixture of von Mises–Fisher components fit by online EM. Utility relaxes toward a low-value, high-uncertainty equilibrium in the absence of evidence, in the tradition of Bayesian demand models from library science and rational-analysis models of human memory. The filter is closed form, stores two scalars per chunk, and costs per query. In a simulation built from a 20,640-chunk corpus accumulated over nine months of daily use of a personal AI agent, BRAG raises Recall@5 on new queries targeting chunks that had been retrieved before from 0.73 to 0.97 and reduces the mean rank of the correct chunk from 12.5 to 3.8, with no significant loss on a control set of chunks that were never queried during training.

Primary area:
probabilistic methods (Bayesian methods, variational inference, sampling, UQ, etc.)
Keywords:
Agentic AI, RAG, Memory Systems, Bayesian Models

Reviews

Among:
Shape:
1234this paper, Top 32%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.50 over 2 reviews, above 68% of the others.
Reviewer_oBS3OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper adapts retrieval using feedback about which chunks are useful. It combines per-chunk Beta beliefs with temporal decay and multiple semantic keys fitted to successful query embeddings through a von Mises–Fisher mixture. On a 20,640-chunk personal corpus, a simulation with 1,405 synthetic queries reports Recall@5 improving from 0.73 to 0.97.

Critical strengths
  1. Adapting retrieval to recurring information needs is a useful problem.

  2. Queries are scored before their feedback is applied, which supports the evaluation of adaptation across different queries.

Critical weaknesses
  1. The novelty of adapting document vectors from query feedback is not sufficiently established. NUDGE (ICLR 2025) already adjusts document embeddings using labeled queries, but is neither discussed nor compared against.

  2. The main experiment does not demonstrate a benefit from combining Bayesian beliefs with adapted keys. In Table 1, Keys alone and full BRAG both achieve 0.97 Recall@5 and 3.8 average rank.

  3. The Bayesian formulation needs a clearer technical contribution. The conjugate updates and vMF mixture fitting are standard, while Appendix B.4 ultimately adopts the prediction recursion as a definition after approximations to the OU dynamics.

  4. The evaluated feedback does not match the stated notion of utility closely enough. Utility is defined through actual downstream use, whereas Appendix A.6 labels whether a chunk contains information answering the query. Retrieval scores alone therefore do not establish better generated answers or better estimates of actual use.

Questions for authors

see weakness

4 comments
AuthorsOpenReview ↗

Response to Reviewer oBS3 (1/2)

Response to Reviewer oBS3 (1/2)

Thank you for pointing us to NUDGE.

W1. "The novelty of adapting document vectors from query feedback is not sufficiently established. NUDGE (ICLR 2025) already adjusts document embeddings using labeled queries, but is neither discussed nor compared against."

BRAG is a generative framework that clarifies the conditions under which previous approaches are optimal and suggests new ones. For example, NUDGE can be interpreted as BRAG with one key per chunk (K = 1), infinite utility lifetime, and a hard prior on the embedding moves. What BRAG adds is the overall generative framework, including a step that grows with evidence, several keys beside the original embedding, online updates, and a utility that fades with time. We made this clearer in the revised related work section.

We compared against NUDGE on three of its seven text benchmarks, using its code, splits and one of its embedders (BGE-small), with NDCG@10 (Table 3 of the revised paper).

NUDGE's step size is chosen on validation, so we tuned BRAG's settings the same way and read the test split once. BRAG is ahead of the better NUDGE variant on FEVER (+0.013) and NQ (+0.015), and the two are tied on NFCorpus (+0.005). The NUDGE rows on the two public benchmarks are in Table 2 of the revised paper.

W2. "The main experiment does not demonstrate a benefit from combining Bayesian beliefs with adapted keys. In Table 1, Keys alone and full BRAG both achieve 0.97 Recall@5 and 3.8 average rank."

The revised paper contains a new experiment with public datasets that investigates the joint contribution of the system dynamics and the adaptation of the observation model.

Table 2 (Section 6.2) runs keys only and full BRAG side by side on LongMemEval-M and PersonaMem-v2. On queries asked again the two tie (0.858 on LongMemEval-M, 0.659 and 0.658 on PersonaMem-v2). The belief matters when a new chunk makes the information in an old one obsolete (a fact updated on LongMemEval-M, a forget request on PersonaMem-v2). There, adding the system dynamics raised Recall@5 of the new chunk from 0.711 to 0.860 and from 0.669 to 0.801, and BRAG ranks the new chunk above the old one in 0.557 and 0.567 of cases, against 0.047 and 0.006 for keys only.

On the original paper's corpus, the rerun of Table 1 (labels applied only within the top 100) shows the belief at the top of the ranking. Recall@1 rises by +0.027 [0.009, 0.048] over keys alone, while Recall@5 and average rank barely move (+0.001 and −0.05).

AuthorsOpenReview ↗

Response to Reviewer oBS3 (2/2)

Response to Reviewer oBS3 (2/2)

W3. "The Bayesian formulation needs a clearer technical contribution. The conjugate updates and vMF mixture fitting are standard, while Appendix B.4 ultimately adopts the prediction recursion as a definition after approximations to the OU dynamics."

The revised version makes the main contribution of the paper clearer. It is a generative model, inspired by, but quite different from, classic library and human memory models. The model illuminates optimality conditions for inference and control. The approximation methods are not novel. They are the Laplace bridge, to move between a logit-Gaussian and a Beta belief, and a factorized posterior with one Beta per chunk. However, these methods had not been used in the classic library and human-memory literature, and they are essential for real-time practical use.

The wording of the paper was confusing regarding the relationship between the OU dynamics in logit space and the Beta recursion. We made this clearer in the revised submission, which now reads "The discrete-time OU process on the logit of utility defines the dynamics of the model. The Laplace bridge approximation turns it into a recursion on the Beta parameters (e, n). Once the bridge is applied, the filter no longer needs the underlying logit process and propagates the belief directly in the Beta family."

To measure the approximations we simulated data from BRAG's own generative model (30 or 50 chunks, 2 keys each, 200 steps, 5 seeds, 55 runs) and compared BRAG with a joint particle filter with 65.5K particles (doubling to 131K moved posterior means by less than 0.003). Since the exact posterior is calibrated under the model, the distance between BRAG's belief and that posterior measures both the approximation error and the loss of calibration.

The simplified recursion. BRAG's beliefs stay within 0.02-0.09 of the exact posterior. Factorizing the posterior across chunks adds a mean error of 0.002-0.007, at the particle filter's own noise floor. The Beta projection adds 0.02 nats KL, and the Laplace bridge another 0.01 nats, with top-3 agreement with the exact filter of 0.975 (0.988 without it). Recall@3 stays within ±0.015 of the exact filter.

W4. "The evaluated feedback does not match the stated notion of utility closely enough. Utility is defined through actual downstream use, whereas Appendix A.6 labels whether a chunk contains information answering the query. Retrieval scores alone therefore do not establish better generated answers or better estimates of actual use."

Yes, this was confusing. The revised paper defines utility through the generative model, not through use. Each query q_t is generated by one chunk, and U_t = i means q_t was drawn from p(q | i). A chunk's utility X_i is then the rate at which it generates queries, its demand. The curator's label ("does this chunk contain the information asked for?") is the observable evidence about which chunk generated the query. When two chunks both contain the answer, both labels are noisy evidence for the same latent cause, which the responsibility γ splits between them. The feedback we evaluate is therefore an observation of the quantity the model defines.

To test answers as well as retrieval, we gave the same LLM the chunks retrieved by plain search or by BRAG and graded its answers with LongMemEval's official judge prompt. Accuracy rose from 0.552 to 0.702 on LongMemEval-M (+0.150 [+0.118, +0.183]) and from 0.457 to 0.553 on PersonaMem-v2 (+0.096 [+0.086, +0.107]).

Dear Reviewer oBS3,

Thank you for your time and effort in reviewing our paper. As the author–reviewer discussion period is coming to a close, we would be happy to answer any remaining questions you may have. We appreciate your engagement.

AuthorsOpenReview ↗

Follow-up before the discussion closes tonight

Dear Reviewer oBS3,

The discussion period closes tonight (November 18, 11:59 pm AoE), so we wanted to follow up once more. In our response we compared BRAG with NUDGE on three of NUDGE's own benchmarks, using its code and splits, and added NUDGE to the two public datasets. We also showed where the Bayesian beliefs add to the adapted keys, which is when a new chunk makes an old one obsolete, and clarified the generative model behind the formulation. Every change is marked in blue in the revised PDF.

Could you let us know whether these results address your concerns? If they do, we would be grateful if you would consider updating your score. If anything is still unclear, we will answer right away.

Best regards, The Authors

Reviewer_ftfmOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

BRAG uses interaction feedback to improve retrieval from long-lived agent memories. Each chunk maintains a Beta belief about its future utility, while its retrieval keys adapt toward queries that the chunk successfully answered. Retrieval combines utility estimates with semantic alignment. On a 20,640-chunk personal-agent corpus and a 44-day simulation, Recall@5 increases from 0.73 to 0.97 on the experimental group. Key adaptation alone reaches the same result.

Critical strengths

The problem is worth studying. Previous queries provide useful information about how users ask for stored content, which a static document embedding does not fully capture. Adapting the index without retraining the encoder is a practical direction, and the retrieval–feedback loop is easy to follow.

The ablations are informative. The authors separate utility updates from key adaptation and openly discuss their non-additive gains. Scoring queries before the day’s updates avoids evaluating a query against an index already updated with that same query. The never-trained control group is also a sensible initial check for interference.

Critical weaknesses

The experiments establish the benefit of feedback-driven key adaptation more clearly than the value of the Bayesian framework. Keys-only and full BRAG both achieve Recall@5 of 0.97 and average rank of 3.8. Although acknowledged, the main comparison remains against a static retriever that cannot exploit feedback. Historical-query retrieval, regularized query centroids, and simple frequency reranking would help separate the value of feedback from the proposed modeling choices. The distinction from feedback-based approaches such as LADER also needs clarification.

The evidence comes from one personal corpus with several generated queries per selected chunk; the 14 runs change query order rather than the corpus or query source. More importantly, the feedback protocol is ambiguous. The main text restricts labels to retrieved chunks, while the appendix describes precomputed labels applied through labels-only updates. If missed targets still receive positive updates, the simulation uses information that may be unavailable during deployment.

Questions for authors

Are positive updates restricted to chunks actually retrieved by each method? What happens when the correct chunk is absent from its candidate set?

How does BRAG compare with a historical-query index and a single regularized query centroid under the same feedback budget and comparable storage?

Are results available on naturally occurring subsequent user queries, rather than generated variants targeting the same selected chunks?

5 comments
AuthorsOpenReview ↗

Response to Reviewer ftfm

Response to Reviewer ftfm

Thank you for the precise questions.

W1. "… Historical-query retrieval, regularized query centroids, and simple frequency reranking would help separate the value of feedback from the proposed modeling choices. The distinction from feedback-based approaches such as LADER also needs clarification."

BRAG is a generative framework that clarifies the conditions under which previous approaches are optimal and suggests new ones. For example, averaging queries corresponds to using an uninformative prior, infinite utility lifetime, and a single vMF observation likelihood model. LADER can be seen as BRAG with uninformative prior, infinite utility lifetime and a kernel density likelihood model, i.e., K→∞ in the BRAG framework. We modified the previous work section of the paper to clarify the similarities and differences with previous approaches.

Table 2 of the revised paper (Section 6.2) compares BRAG with a historical-query cache, a regularized centroid, Rocchio, frequency and recency reranking, NUDGE and LADER on LongMemEval-M and PersonaMem-v2, all learning from the same labels. On queries asked again, the cache, the centroid, Rocchio, NUDGE and LADER all match BRAG within 0.006. Frequency reranking ties on LongMemEval-M (0.858) but falls to 0.476 on PersonaMem-v2.

In this experiment, the system dynamics (which carry the belief forward in time before each query) mattered most when newer chunks made the information in older, similar chunks obsolete (the update and forget columns of Table 2). BRAG found the new fact in 0.86 and 0.80 of reads, against 0.71 and 0.67 for every method that learns only from past feedback.

BRAG+ adds two fixes. A gate applies learned state only to queries close to a chunk's learned keys, and a curator LLM transfers the keys and utility of the obsolete chunk to the new chunk. Both are described in the revised paper.

W2 (protocol) / Q1. "The main text restricts labels to retrieved chunks, while the appendix describes precomputed labels applied through labels-only updates. … Are positive updates restricted to chunks actually retrieved by each method? What happens when the correct chunk is absent from its candidate set?"

In the submitted paper, each generated query's target chunk was labeled useful even when it was not retrieved. This corresponds to a curator who widens the search when the answer is missing. In our deployed system (13 users, up to six months of use), the curator widens up to k = 100. The revised paper states this protocol in the main text and reruns Table 1 with labels gated at k = 100. Absolute results changed by about 2 percentage points, and the ranking of methods stayed the same.

On the public benchmarks we also ran the strict version, where each method learns only from chunks in its own top 10. Recall@5 on PersonaMem-v2 falls by 0.053 for BRAG and by 0.033-0.037 for keys only, Rocchio and the cache. It did not change on LongMemEval-M. A target that is never retrieved gets no label, so no method learns it.

Q2. "How does BRAG compare with a historical-query index and a single regularized query centroid under the same feedback budget and comparable storage?"

Both are in Table 2 (see W1 above), learning from the same labels as BRAG.

Regarding memory and computation, BRAG scales linearly with the number of chunks. Per chunk it adds two scalars (the utility belief) and, at most, K learned key vectors next to the original embedding. New keys are allocated only for chunks that receive feedback, so unused chunks cost no more than in plain RAG. The average cost therefore depends on how much of the memory is used. In our PersonaMem-v2 runs BRAG used on average 1,322 real-valued numbers per memory, against 1,004 for the regularized centroid and 988 for the historical-query cache.

Q3. "Are results available on naturally occurring subsequent user queries, rather than generated variants targeting the same selected chunks?"

BRAG is most useful when chunks are retrieved many times, and of no use when a chunk is used only once in the lifetime of the database. The statistics of utilization are likely to change dramatically for different scenarios.

We ran an experiment on NQ (Natural Questions), whose queries are real, anonymized Google searches. On NQ, BRAG reaches 0.474 NDCG@10 against 0.212 for standard RAG, 0.386 for NUDGE-M and 0.459 for NUDGE-N.

We added six public benchmarks with their own corpora and queries, LongMemEval-M and PersonaMem-v2 (Table 2), NFCorpus, FEVER and NQ (Table 3), and TripClick.

On TripClick (Rekabsaz et al., SIGIR 2021), real click logs of a health search engine that LADER was tested on, with the 20-candidate read BRAG uses in deployment, BRAG+ and LADER are within 0.014 NDCG@10 (0.311 / 0.303 / 0.308 vs 0.306 / 0.316 / 0.322). With 100 or more candidates LADER leads by 0.02–0.04, while BRAG+ gains little beyond 20.

Dear Reviewer ftfm,

Thank you for your time and effort in reviewing our paper. As the author–reviewer discussion period is coming to a close, we would be happy to answer any remaining questions you may have. We appreciate your engagement.

AuthorsOpenReview ↗

Follow-up before the discussion closes tonight

Dear Reviewer ftfm,

The discussion period closes tonight (November 18, 11:59 pm AoE), so we wanted to follow up once more. In our response we added the baselines you asked for (a historical-query index, a regularized query centroid and frequency reranking) together with NUDGE and LADER on two public datasets. We also ran the strict protocol in which each method learns only from chunks in its own top 10, and added benchmarks with naturally occurring queries, including TripClick's real click logs. Every change is marked in blue in the revised PDF.

Could you let us know whether these results address your concerns? If they do, we would be grateful if you would consider updating your score. If anything is still unclear, we will answer right away.

Best regards, The Authors

Comments on the paper

5 comments
AuthorsOpenReview ↗

General response (1/2)

General response (1/2)

We thank the reviewers. All three note that in Table 1 key adaptation alone matches full BRAG, and that simple feedback baselines are missing. We agree, and we ran the requested experiments.

  1. Public benchmarks. Under BRAG's protocol, Recall@5 rises from 0.697 to 0.858 on LongMemEval-M (301 test questions, +0.162 [+0.129, +0.197]) and from 0.426 to 0.658 on PersonaMem-v2 (3,288 needs, +0.232 [+0.220, +0.245]). The gain holds with a second embedder.
  2. Named baselines. On re-asked needs, a regularized centroid, Rocchio, a historical-query cache, LADER and NUDGE all match BRAG within 0.006.
  3. NUDGE's benchmarks. With BRAG's settings tuned on validation, as NUDGE's step size is, BRAG reaches 0.963, 0.462 and 0.474 NDCG@10 on FEVER, NFCorpus and NQ, against 0.950, 0.457 and 0.459 for the better NUDGE variant.
  4. Answers. Answer accuracy rises by +0.150 (LongMemEval-M) and +0.096 (PersonaMem-v2).
  5. Labels and approximations. Labels restricted to each method's own top 10 cost 0.053 on PersonaMem-v2. Random label noise up to 30% barely matters, while labels from an LLM curator cost 0.176. Both approximations keep Recall@5 on re-asked needs within 0.003 of the exact versions.

What we agree with, and what we changed.

  1. The evaluation was too narrow. The submitted paper used a single user's private memory. That is realistic, but it means one corpus, one embedder and generated queries. The revision adds public benchmarks with many users (LongMemEval-M, PersonaMem-v2), three of NUDGE's seven text corpora, run with its own code, a second embedder, and the baselines the reviewers asked for.
  2. The relationship to previous models was unclear, and we missed two important papers, NUDGE and LADER. Both belong to the same family as BRAG, which interprets them within an optimality framework. NUDGE is BRAG with one key per chunk and no dynamics, and LADER is the many-key (kernel density) limit with no dynamics. ERM (Hu et al., 2026), which the revision now also cites, is BRAG with feedback-gated key updates and no utility dynamics. The revised related-work section says this.
  3. The label protocol was ambiguous. The main text and the appendix described it differently. The revision states the curator's rescue step (widening the search up to k = 100) and reruns Table 1 with labels gated at k = 100. Recall@5 changes by at most 0.021, and the ranking of methods is unchanged.
AuthorsOpenReview ↗

General response (2/2)

General response (2/2)

What this shows. The common feedback methods are special cases of BRAG with fixed parameters, so on re-asked questions they match it, and the gain there comes from key adaptation. BRAG pulls ahead when the utility of chunks changes over time. In two scenarios a new chunk makes an old one obsolete. On LongMemEval-M a stored fact is updated, and on PersonaMem-v2 the user asks to forget a preference. We measure Recall@5 of the new chunk at the next ask and how often it ranks above the old one (higher is better).

MethodLME-M, update arrives later (Recall@5 of new fact)PMv2, forget request arrives later (Recall@5 of forget request)LME-M, new ranked above oldPMv2, new ranked above old
Plain search0.7110.6730.5280.236
Regularized centroid0.7110.6730.3060.236
Rocchio0.7110.6690.0470.007
Historical-query cache0.7110.6690.0470.005
Recency rerank0.7150.0220.553–
Frequency rerank0.7110.6730.5280.236
NUDGE-M0.7110.6690.3570.005
NUDGE-N0.7110.6690.1320.005
LADER0.7110.6690.4850.005
Keys only (no belief)0.7110.6690.0470.006
BRAG0.8600.8010.5570.567
BRAG+0.8510.8500.8430.540

– Recency rerank retrieves neither the old preference nor the forget request (each is in its top 5 in under 3% of reads), so the ranking measure does not apply.

Methods that learn only from past feedback cannot lift a fact that just arrived, so they find it about as often as plain search. Only BRAG's belief lifts it.

BRAG+. The revision adds BRAG+, which is BRAG with two fixes. A gate applies learned state only to queries close to a chunk's learned keys, which cuts the loss on never-asked needs. A curator LLM detects when a new chunk makes an older, similar chunk obsolete and transfers the old chunk's learned state to the new one, at a cost of 0.009 in Recall@5 on unchanged facts. Every table reports BRAG and BRAG+.

Revised PDF. Changes are marked in blue. The main changes are a new Experiments section (public benchmarks with the requested baselines, changing-facts scenarios, answer accuracy, label noise, and comparisons with NUDGE and LADER), the label protocol stated in the main text with Table 1 rerun, BRAG+, utility defined through the generative model in Sections 2–5, and a rewritten related-work section that places NUDGE and LADER within BRAG's framework.

AuthorsOpenReview ↗

Response to Reviewer oLBV

Response to Reviewer oLBV

Thank you for the careful review.

W1 / Q1. "… Given the same historical feedback, how much does BRAG improve over query-vector averaging, single-key adaptation, and frequency- or recency-based reranking? Under what conditions does full BRAG significantly outperform Keys?"

We ran the proposed baselines plus NUDGE and LADER on two public datasets, LongMemEval-M and PersonaMem-v2 (Table 2 of the revised paper, Section 6.2).

  • Asked again. Every feedback method lifts Recall@5 above plain search (0.697 to about 0.858 on LongMemEval-M, 0.426 to about 0.66 on PersonaMem-v2). The regularized centroid, Rocchio, the historical-query cache, NUDGE and LADER all match BRAG within 0.006. Frequency and recency reranking fall behind on PersonaMem-v2 (0.476 and 0.583, against 0.658 for BRAG).
  • Never asked. Every learning method falls below plain search (0.444). BRAG falls to 0.320, and BRAG+'s gate recovers most of the loss (0.424).
  • When full BRAG beats Keys. When a new chunk makes an old one obsolete (a fact updated on LongMemEval-M, a forget request on PersonaMem-v2). Keys only and every other method that learns only from past feedback stay at plain-search level (Recall@5 of the new chunk 0.711 and 0.669–0.673), while BRAG reaches 0.860 and 0.801. BRAG ranks the new chunk above the old one in 0.557 and 0.567 of cases, against 0.047 and 0.006 for keys only.

BRAG+ adds two fixes. A gate applies learned state only to queries close to a chunk's learned keys, and a curator LLM transfers the keys and utility of the obsolete chunk to the new chunk. Both are described in the revised paper.

W2 / Q2. "The experiments use a single user's corpus, one embedding model, and agent-generated queries. Labels are assigned by an LLM without human verification. … Do the main improvements hold on real query logs, other users, or independent corpora? Do retrieval improvements translate into higher final answer accuracy?"

Other users. Addressed in W1 / Q1 above.

Another embedding model. With a second embedder (Qwen3-Embedding-8B, BRAG not retuned), the gain over plain search holds (+0.187 Recall@5 on LongMemEval-M and +0.243 on PersonaMem-v2), no smaller than with our default embedder.

Independent corpora. On three of NUDGE's seven text benchmarks (NFCorpus, FEVER, and NQ with real anonymized Google searches), with NUDGE's code, splits and embedder, BRAG reaches 0.462, 0.963 and 0.474 NDCG@10, against 0.457, 0.950 and 0.459 for the better NUDGE variant (Table 3 of the revised paper).

Answers. Answer accuracy rises by +0.150 [+0.118, +0.183] (LongMemEval-M) and +0.096 [+0.086, +0.107] (PersonaMem-v2), and falls by 0.078 on never-asked questions, graded with LongMemEval's official judge prompt.

Labels. Labels here come from gold evidence, not an LLM. Random label noise matters little. With 30% extra wrong labels Recall@5 on PersonaMem-v2 falls by 0.002, and with 30% of labels missing by 0.032. With a 35B LLM curator instead of gold labels (50 personas), every feedback method loses about the same Recall@5 (−0.167 to −0.184, Appendix Table 5), so label quality matters for all feedback methods alike, not for BRAG in particular.

Real query logs. On TripClick (Rekabsaz et al., SIGIR 2021), the click logs of a health search engine that LADER was tested on, with the 20-candidate read BRAG uses in deployment, BRAG+ and LADER are within 0.014 NDCG@10 (0.311 / 0.303 / 0.308 vs 0.306 / 0.316 / 0.322). With 100 or more candidates LADER leads by 0.02–0.04, while BRAG+ gains little beyond 20.

W3 / Q3. "… Under the default parameters, what are the approximation errors and posterior calibration of the simplified prediction recursion? Do the results change when the nearest-key approximation is replaced with the full mixture likelihood?"

To measure the approximations we simulated data from BRAG's own generative model (30 or 50 chunks, 2 keys each, 200 steps, 5 seeds, 55 runs) and compared BRAG with a joint particle filter with 65.5K particles (doubling to 131K moved posterior means by less than 0.003). Since the exact posterior is calibrated under the model, the distance between BRAG's belief and that posterior measures both the approximation error and the loss of calibration.

The simplified recursion. BRAG's beliefs stay within 0.02-0.09 of the exact posterior. Factorizing the posterior across chunks adds a mean error of 0.002-0.007, at the particle filter's own noise floor. The Beta projection adds 0.02 nats KL, and the Laplace bridge another 0.01 nats, with top-3 agreement with the exact filter of 0.975 (0.988 without it). Recall@3 stays within ±0.015 of the exact filter.

Nearest-key vs full mixture. Taking the max over keys and averaging over them give the same Recall@3 (−0.004 ± 0.008). On PersonaMem-v2, replacing the nearest key with the full mixture changes Recall@5 by −0.003 [−0.004, −0.001].

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.