acceptodds
Under review as a conference paper at ICLR 2027

TRACE: Trajectory Selection for Parallel Scaling of Search Agents

Abstract

Parallel scaling improves search agents by generating multiple candidate rollouts for the same query, yet a correct answer may already be present in the pool and still be missed by final-answer voting. We formulate this consolidation stage as trajectory selection and introduce TRACE (Trajectory Ranking with Aggregated Cross-Rollout Evidence), a lightweight learned selector that ranks completed trajectories using the search evidence behind their answers. TRACE preserves individual query and evidence occurrences, connects rollouts through shared content or document identity, and propagates information across these relations. Each candidate answer then reads the updated states of its own trajectory, preserving retrieval provenance while incorporating evidence from related rollouts. Trained with answer-level supervision over frozen text embeddings, TRACE returns an existing answer without additional search or autoregressive aggregation. One selector per search setting transfers across rollout policies and agent backbones without agent-specific fine-tuning, improving over voting across six WebQA policies and six long-horizon dataset–backbone combinations at . On Qwen2.5-14B Base/SFT WebQA pools, TRACE achieves 45.2/49.2% EM, compared with 43.9/48.0% for the strongest Qwen3-32B generative aggregators. On long-horizon FRAMES, GAIA, and BrowseComp, it reaches 78.7% average accuracy, exceeding majority voting by 3.2 percentage points. On Base WebQA pools, TRACE with only 8 rollouts comes within 0.4 points of majority voting over 64. Beyond accuracy, TRACE is substantially more efficient than the three LLM-based aggregators, SolAgg, SummAgg, and AggAgent, achieving at least higher processing throughput across all seven WebQA benchmarks. These results show that reusing cross-rollout search evidence provides an effective and efficient alternative to heavyweight generative aggregation for parallel search. Code is available at https://anonymous.4open.science/r/TRACE-C718.

Primary area:
foundation or frontier models, including LLMs
Keywords:
test-time scaling, search agents, trajectory ranking, graph neural networks, evidence aggregation

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_Hb3COpenReview ↗
Rating
3(+1)
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes TRACE, a lightweight trajectory selector for parallel search agents. Instead of using majority voting or generative aggregation, TRACE constructs a query–evidence graph across completed search trajectories, propagates shared evidence through a GNN, and ranks candidate answers using an answer-conditioned readout. Experiments across WebQA and long-horizon search benchmarks demonstrate improvements over voting and competitive performance against LLM-based aggregators, with substantially lower post-rollout computation costs.

Critical strengths
  1. Efficient trajectory selection. TRACE avoids additional autoregressive generation and achieves 12–258× lower post-rollout GPU cost than LLM-based aggregation methods.
  2. Well-designed cross-rollout evidence sharing. The occurrence-preserving graph allows trajectories to exchange information while retaining their original retrieval context. The component ablations support the effectiveness of this design.
  3. Remaining gap to oracle selection. On long-horizon benchmarks, TRACE achieves 78.6% accuracy against a Pass@16 upper bound of 90.9%, leaving a 12.3-point gap. This raises questions about how effectively the selector identifies correct answers already present in the candidate pool.
Critical weaknesses
  1. Limited gains on stronger search policies. TRACE provides only marginal improvements on RL-trained agents (+0.3 points for Qwen2.5-14B RL). Given the high variance of search benchmarks, the lack of statistical significance analysis raises concerns about the reliability of these gains.
  2. Selection-only limitation. TRACE can only return an existing answer and cannot synthesize new answers from complementary evidence, limiting its applicability when no trajectory is correct.
  3. Unverified factual discrimination. TRACE learns to rank candidates using answer-level correctness labels and cosine similarity between learned representations, without explicit supervision for evidence verification. It remains unclear whether the selector can identify subtle factual errors rather than relying on semantic relevance or evidence overlap. Despite a Pass@16 of 90.9%, TRACE achieves only 78.6% selection accuracy, leaving a substantial gap that raises concerns about its ability to reliably identify correct answers
Questions for authors
  1. Statistical reliability. Could the authors report variance or statistical significance across repeated evaluations, particularly for the small gains on RL-trained agents?
  2. Selection versus synthesis. How does TRACE handle cases where no candidate is correct but complementary evidence across trajectories could support a correct answer?
  3. Factual discrimination. How does TRACE learn to distinguish factual correctness from semantic relevance using only answer-level supervision? Could the authors analyze why the selector still misses correct candidates despite the high Pass@16?
3 comments

Thank you for the detailed review. Shared controls are in the General Response.

W1 & Q1. RL pools and significance. All methods share questions and pools, so we use question-level paired tests (exact McNemar, paired bootstrap). All six WebQA pools are in Appendix K.

TRACE vs. MVMV → TRACEGain [95% CI]p
WebQA 14B Base42.6 → 45.2+2.66 [+1.76, +3.55]6e-9
WebQA 14B SFT47.2 → 49.2+1.92 [+1.22, +2.66]2e-7
WebQA 7B RL45.1 → 45.8+0.67 [+0.29, +1.06]0.0015
WebQA 14B RL47.6 → 47.9+0.26 [−0.19, +0.70]0.32
WebQA RL, both pools46.4 → 46.8+0.46 [+0.16, +0.77]0.003
Long-horizon average75.5 → 78.6+3.07 [+1.64, +4.54]1e-5

The small RL gain reflects how little there is to select from, not the strength of the policy:

  • Answer collapse. In 57–61% of RL questions all 16 rollouts give the same answer (2–9% for Base), and every selection method then equals MV. Reweighting the Base/SFT gains to the RL distribution of distinct answers accounts for most of the difference.
  • Parallel scaling saturates on RL pools. MV over 16 rollouts adds 0.7–1.1 points over one rollout (13.1 on 14B Base).
  • The best parallel-scaling results use TRACE: SFT rollouts with TRACE reach 49.2 (14B) and 48.3 (7B), above every method on the RL pools.

W2 & Q2. Selection vs. synthesis. TRACE targets questions where a correct answer is already in the pool but voting misses it: on long-horizon search Pass@16 is 90.9% and MV 75.5%, so 15.4 points are available to selection alone. When no candidate is correct, the generative aggregators recover only 0.3–1.4 points of total accuracy, mostly by shortening a verbose candidate or from their own knowledge, not by combining evidence across trajectories (Appendix J.4). TRACE still has the higher overall accuracy (Tables 1–2).

W3 & Q3. Factual discrimination. All candidates in a pool answer the same question from evidence retrieved for it, so semantic relevance is nearly constant within a pool. The ranking losses contrast candidates within a pool, so only features that separate a supported answer from an unsupported one can lower them. On the 2,168 questions whose pool holds the correct answer and a wrong answer of the same type (another date, number, person or place):

SubsetQuestionsTRACEMV
All2,16874.267.1
Correct answer tied with top wrong answer15571.043.9
Correct answer in the minority60925.50.0

Example (Bamboogle): "Who was the father of the father of psychoanalysis?" Gold: Jacob Freud. MV returns Schlomo Freud (6 votes), TRACE Jacob Freud (5 votes). Both groups read the same passage, which says Jacob is Sigmund's father and Schlomo is Jacob's. TRACE scores the Jacob rollouts 0.09 to 0.13 and the Schlomo rollouts −0.40 to −0.21; evidence overlap cannot explain this (more examples: Appendix J.5).

The gap to Pass@16. Pass@16 is the ceiling of every selection method. When a correct candidate exists, TRACE returns one in 86.5% of cases (MV 83.1%), and a missed correct answer is usually ranked second (65–75%). Of the remaining gap, 5.0 points are questions where only one or two of the 16 rollouts are correct, which is also where TRACE gains most over voting (single correct rollout: 20.0 → 38.8).

Thank you again for your detailed review. As the author-reviewer discussion closes on November 18, we would like to check whether our response has addressed your concerns. A short summary of our answers to your three questions:

  • Q1, statistical reliability. Paired question-level tests (Appendix K): the gain over majority voting is significant on the Base/SFT WebQA pools (p ≤ 2e-7) and on the long-horizon average (+3.07, 95% CI [+1.64, +4.54]). The RL gain is small because in 57–61% of RL questions all 16 rollouts give the same answer, so every selection method equals voting.
  • Q2, selection vs. synthesis. On long-horizon search Pass@16 is 90.9% and voting 75.5%, so 15.4 points are available to selection alone. When no candidate is correct, the generative aggregators recover only 0.3–1.4 points of total accuracy (Appendix J.4).
  • Q3, factual discrimination and the gap to Pass@16. On the 2,168 questions whose pool holds the correct answer and a wrong answer of the same type, TRACE reaches 74.2% against 67.1% for voting, and 25.5% against 0.0% when the correct answer is in the minority. When exactly one rollout is correct, TRACE selects it in 38.8% of long-horizon cases against 20.0% for voting.

If any concern still stands, please let us know and we will address it before the deadline. If our response has resolved them, we would be grateful if you could consider updating your score.

Thank you for your rebuttal. I have no other questions. Raised my rating to weak acceptance.

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

TRACE frames post-rollout consolidation as selecting one of K completed search trajectories rather than generating a new answer. It builds a graph of Subquery and Evidence occurrences, linking rollouts through identical chunks (WebQA) or shared Doc nodes (long-horizon browsing). A 4-layer relational GraphSAGE processes frozen Qwen3-Embedding-8B features, and an answer-conditioned cross-attention readout over each candidate’s own trajectory produces its score. Training combines BCE, listwise, and hard-ranking losses, with one checkpoint per search setting. TRACE reports gains over majority voting across 6 WebQA policies and 3 long-horizon benchmarks with 2 backbones, performing competitively with Qwen3-32B aggregators at 12–258× lower post-rollout cost on WebQA.

Critical strengths
  • Keeping each occurrence separate is well motivated.
  • Doc nodes are justified by the data: 74.3% of same-document cross-rollout pairs have different content.
  • TRACE beats majority voting on all 7 WebQA datasets (14B Base and SFT) and on all 6 long-horizon combinations.
  • Strong efficiency of 0.27 GPU-h vs. 3.3–70 GPU-h (for aggregators).
  • TRACE at K=8 roughly matches majority voting at K=64.
Critical weaknesses
  • TRACE trains on ~100k questions but compares only against heuristics and zero-shot aggregators.
  • Removing cross-rollout links (72.7) or Doc/content links (71.1) performs worse than majority voting (75.5), potentially exaggerating the graph's contribution.
  • Training on Qwen2.5-14B rollouts from NQ/HotpotQA overlaps with evaluation settings.
  • The unspecified long-horizon training backbone further weakens cross-model generalization claims.
  • TRACE removes 27.2% of 7B Base trajectories, potentially changing candidate pools across methods (K=1 score (28.5) trails single-rollout performance (29.5)).
  • SolAgg outperforms TRACE on 3/6 WebQA policies, while multi-hop datasets often favor existing aggregators (gains concentrate on NQ and PopQA).
  • Small ablation gains (0.3–1.7 pp) and GAIA improvements (3/103 questions) may reflect noise.
  • Comparisons omit LLM-based index selection, overlap-weighted voting, and ParallelMuse, leaving TRACE's advantage over simpler alternatives unclear.
  • AggAgent trails majority voting (68.9 vs. 75.5), reports identical FRAMES scores across backbones, and uses estimated runtime.
  • Reported efficiency excludes ~1.7M training trajectories, labeling costs, and separate selectors.
  • Experiments focus on short-answer exact-match tasks and fixed-corpus retrieval.
  • Conflicting accuracy/gain figures (78.7/3.2 vs. 78.6/3.1), an incorrect all-settings voting claim, missing K=64 results, duplicated tables, and selectively presented comparisons weaken clarity.
Questions for authors
  • How often does TRACE override the majority answer, and how accurate is it in those cases? Qualitative examples would help.
  • How does TRACE perform without the vote feature?
2 comments

Thank you for the thorough review. We answer in order; pairs are WebQA / long-horizon.

W1. Learned baselines. A graph-free supervised MLP with the same questions, labels, embeddings and losses reaches 43.6 / 75.7 (MV 42.6 / 75.5, TRACE 45.2 / 78.6). Supervision explains 1.0 of the 2.6 points on WebQA and 0.2 of the 3.1 on long-horizon; the rest comes from the graph.

W2. Ablations below MV. The graph's contribution is measured against MV and graph-free learned selectors: long-horizon MV 75.5, w/o GNN 75.5, TRACE 78.6. The link-removal variants keep the GNN and cut specific relations to show which ones propagation depends on; the headline gain does not rest on them.

W3. Train/eval overlap. Training uses the NQ and HotpotQA training splits with 14B Base rollouts only; evaluation questions are disjoint. Five of seven datasets and five of six policies are unseen. On 14B Base the five unseen datasets gain +2.6 on average, against +2.8 for NQ and HotpotQA. Unseen policies gain +3.0 (7B Base), +2.2 (7B SFT) and +1.9 (14B SFT).

W4. Long-horizon training. The selector is trained on the public OpenResearcher dataset: deep-research questions, each with 16 gpt-oss-120B trajectories over an offline corpus. None of the three benchmarks is in it. OpenResearcher-30B-A3B rollouts are unseen, and TRACE improves over MV on all three benchmarks with this backbone (+2.6, +2.2, +2.9). Section 4.1 now states this.

W5. Filtering; K = 1. The validity filter (final answer and retrieved evidence) is applied once, and voting, the aggregators and TRACE share the retained pool, so it favors no method. Only the "Single rollout" reference row is unfiltered (Appendix A.1), hence 28.5 at K = 1 on the retained pool.

W6. SolAgg; multi-hop. SolAgg uses a 32B LLM, larger than both rollout models, at 17× the post-rollout cost. TRACE is higher on 7B SFT, 14B Base and 14B SFT and within about a point elsewhere. Its largest gains over MV are on multi-hop 2Wiki (+4.0) and MuSiQue (+3.0). On long-horizon search it is best or tied in five of six pools and 2.0 points above SolAgg on average.

W7. Noise. Paired tests (G3): the gain over MV is significant on the four Base/SFT WebQA pools (p ≤ 2e-7) and on the long-horizon average (+3.07, 95% CI [+1.64, +4.54]), which pools 2,866 questions; on BrowseComp-Plus with gpt-oss-120B it is +8.3 (CI [+6.1, +10.6]). The long-horizon ablation gaps in Table 3 are 3.0 to 5.9 points.

W8. Simpler alternatives (G1). Overlap-weighted voting: 42.0 / 75.6. Qwen3-32B index selection: 41.3 / 71.8 from answers, 42.7 / 73.7 with trajectory summaries. Neither exceeds MV by more than 0.1. ParallelMuse changes rollout generation and so yields a different pool; its aggregation stage, which compresses each trajectory into a report before integration, is the strategy SummAgg covers.

W9. AggAgent. All aggregators use the same Qwen3-32B, and the single-pass SolAgg and SummAgg exceed MV (76.6, 77.0). AggAgent instead needs the LLM to act as an agent, reading 16 long-horizon trajectories through tool calls, which a 32B model does not do reliably. Its gap to MV grows with trajectory length (1.1–3.6 points on WebQA, 6.6 on long-horizon), and it depends strongly on the aggregator (25.4 → 41.5 on WebQA Base from Qwen2.5-14B to Qwen3-32B, Appendix F). The two FRAMES scores coincide but come from different questions (discordant with TRACE: 89 / 19 and 96 / 20). Table 5 now lists each run's LLM calls and tokens: 28,054 calls and 817M prompt tokens for AggAgent (69.93 GPU-hours), none for TRACE (0.19 GPU-hours; its text encoder now runs on vLLM, like the aggregators).

W10. Training cost. Table 5 is the per-query cost. Training is one-time (Appendix H): 4.2 minutes on one H20 for the WebQA selector. With 10% of the training questions it still reaches 44.4 on WebQA (MV 42.6). WebQA labels are exact match against gold answers and the long-horizon data is public: no added annotation. One selector per search protocol is reused across all datasets, policies and backbones.

W11. Scope. Long-horizon agents browse with Search/Open/Find and return free-form responses judged for correctness, not string-matched. BrowseComp-Plus and FRAMES use fixed corpora; GAIA runs on the live web.

W12. Presentation. All corrected: 78.6 / +3.1 throughout, "over majority voting", MV at K = 64 (43.6) in Appendix C, Table F.1 comparing the two aggregator backbones, and a pointer to Table E.1 for all six policies.

Q1. Overrides. TRACE leaves the majority answer on 20.3% of questions. There it is correct in 60.3% of long-horizon cases (MV 39.4%) and 18.9% of WebQA cases (MV 8.6%); over all questions where exactly one is right, it wins 413 to 157. Examples: Appendix J.5.

Q2. No vote feature. 43.1 / 76.1: slightly above MV from evidence alone, yet differing from it on 39% of WebQA questions. A vote-only scorer equals MV exactly; Full TRACE combines both (G1).

After carefully reading the authors' comments addressing my queries and those of the fellow reviewers, I have decided to raise the rating to a weak acceptance.

Reviewer_5ZntOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper looks at the last step of parallel scaling for search agents: you have K finished search rollouts, so which answer do you return? Instead of majority voting or asking another LLM to write a new answer, the authors treat this as trajectory selection and propose TRACE. TRACE builds a graph over all rollouts, keeping each search query and each retrieved piece of evidence as its own node, and linking rollouts that saw the same chunk (WebQA) or the same document (long-horizon browsing). A relational GNN passes information across these links. Each candidate answer then attends over its own trajectory’s updated nodes and gets a score, and the top-scoring existing answer is returned. The selector is trained on answer-correctness labels over frozen embeddings, so it can be reused across different rollout generators.

Critical strengths
  1. Returning one of the existing answers, rather than generating a new one, makes sense when the right answer is often already in the pool. A single selector trained once works across different generators without fine-tuning them. It needs 0.27 GPU-hours, compared with 3.3–70 GPU-hours for the LLM aggregators, and is still competitive or better on accuracy. This matters because the cost of parallel scaling is a real bottleneck, and the method is cheap enough to actually use.
Critical weaknesses
  1. It’s not clear why TRACE works: is it really reasoning over evidence, or just a smarter form of voting? The model only sees frozen sentence embeddings, so it’s hard to believe it checks whether a passage actually supports an answer. Several results hint that the gain may simply be “evidence-level consensus”: Without the GNN, long-horizon accuracy equals majority voting (75.5). All ablations keep the answer-frequency (vote) feature. RL pools, where rollouts overlap the most (Jaccard ~0.85), show the smallest gains (+0.3–0.7). Adding more links (Doc + content) does worse than Doc alone. The key controls are missing: a no-learning “evidence-overlap voting” baseline, structure-only (random embeddings), random rewiring, removing the vote feature, and a win/loss breakdown against majority voting. Without these, the main claim (graph-based cross-rollout evidence reasoning) isn’t really supported, and the contribution could reduce to something much simpler.
Questions for authors

How much of the gain comes from structure rather than content? it is unclear to me. Could you try to solve my weakness comment and do some analysis?

2 comments

Thank you for listing the controls that would settle this question. We ran all of them; the combined table is in the General Response (G1). Numbers are WebQA / long-horizon; MV is 42.6 / 75.5 and TRACE 45.2 / 78.6 (+2.6 / +3.1).

1. Simpler selectors do not reproduce the gain. Against MV, the controls in G1 give +0.0 for the vote feature alone, −0.6 / +0.1 for no-learning evidence-overlap voting, +0.5 / +0.3 for structure only (random embeddings) and at most +1.1 / +0.2 for content without the graph, so evidence-level consensus does not account for the gain. The last control is why the w/o-GNN variant equals MV on long-horizon: a learned selector without propagation adds nothing there, and the +3.1 appears only with the GNN.

2. The gain needs content and evidence-identity links together, and evidence complements the vote.

  • Random rewiring (G1; embeddings kept; degree-preserving swaps of cross-rollout links, 29% of true links survive): 44.3 / 76.4, 0.9 / 2.2 below TRACE. A rewired link still brings in evidence that another rollout retrieved for the same question, so it acts as a coarse summary of the pool, which keeps this control above removing cross-rollout links (Table 3). Linking by evidence identity adds the rest.
  • No vote feature (G1; retrained): 43.1 / 76.1, slightly above MV from evidence alone while returning a different answer on 39% of WebQA questions. Full TRACE's entire net gain over MV falls on questions where this evidence-only model also leaves the majority, so the two signals are complementary.

3. When TRACE and MV disagree, TRACE is right far more often. Over all 9,116 questions (G2), TRACE is right where MV is wrong 413 times, against 157 the other way (WebQA 14B Base 144 / 61, SFT 96 / 36, long-horizon 173 / 60), with more wins than losses on every dataset and pool. It leaves the majority answer on 20.3% of questions, where its accuracy is 29.2% against 16.2% for MV. The score follows the evidence, not the answer string: in a HotpotQA example (Appendix J.5), five rollouts give the same correct answer and so share the same vote feature; the two that retrieved the supporting passage score 0.84 and 0.77, the three that did not score −0.76 or lower.

4. The two remaining observations come from answer collapse and from duplicated links.

  • RL pools (Table E.2). High overlap there comes with answer collapse: in 57–61% of RL questions all rollouts return the same answer (2–9% for Base), and every selector then coincides with MV. Where pools are diverse the gain is large: +1.3, +3.8, +3.9 and +3.4 points for 2, 3, 4–5 and 6+ distinct answers on Base/SFT pools.
  • Doc + content. In browsing graphs every identical-content pair is also a same-document pair (383K pairs by both counts, Appendix A.3), so content links connect no new occurrences. They give repeated copies of a passage a one-hop message of their own, which carries no new evidence and competes with the Doc message that brings in different passages of the same source.

Thanks for you rebuttal I have no other comments.

Comments on the paper

1 comment
AuthorsOpenReview ↗

General Response to All Reviewers

We thank the reviewers for their careful reading. We ran every requested control. In short: vote counts or evidence overlap alone give no gain over majority voting (MV); the gain appears only when text content is propagated across rollouts; and it is statistically significant. All numbers use K = 16 and the pools and checkpoints of the submission. "WebQA" is the Qwen2.5-14B Base pool (EM); "Long" is the six-pool long-horizon average. Changes in the revised PDF are in blue.

G1. Where the gain comes from (Table 4, Appendix I)

SelectorUsesWebQALong
MVvotes42.675.5
Learned scorer, vote feature onlyvotes42.675.5
Overlap-weighted voting (no learning)votes, overlap42.075.6
LLM index selection (Qwen3-32B)candidates42.773.7
Structure only (random embeddings)graph, votes43.175.8
Supervised MLP, no graphtext, votes43.675.7
Random rewiring of cross-rollout linkstext, votes44.376.4
TRACE w/o vote featuregraph, text43.176.1
Full TRACEgraph, text, votes45.278.6
  • Votes alone add nothing. A scorer trained on the vote feature reproduces MV question by question.
  • Evidence overlap does not explain the gain. Without learning it stays within −0.6 to +0.3 of MV.
  • Structure or content alone is not enough: +0.5 / +0.3 with random embeddings, at most +1.1 / +0.2 with text but no graph. The full +2.6 / +3.1 needs content propagated through the graph.
  • Evidence and votes are complementary. Retrained without the vote feature, TRACE never sees how many rollouts agree, yet is slightly above MV from evidence alone while answering differently on 39% of WebQA questions. Full TRACE combines both: all of its net gain (+83 questions) falls where the evidence-only model also leaves the majority (148 correct vs. 65 for MV).

G2. What TRACE does differently from MV (Appendix J)

  • Over 9,116 questions (WebQA 14B Base/SFT, six long-horizon pools), TRACE leaves the majority answer on 20.3%. It is right and MV wrong 413 times, the reverse 157 times, with more wins than losses on every WebQA dataset and long-horizon pool.
  • In 154 questions TRACE returns a correct answer with at most two votes, fewer than the MV answer. No voting rule recovers these.
  • When exactly one of 16 rollouts is correct, MV selects it in 8.2% (WebQA) / 20.0% (Long) of cases, TRACE in 23.2% / 38.8%.

G3. Significance (Appendix K). Paired tests (exact McNemar, 10,000-sample bootstrap): the gain over MV is significant on all four Base/SFT WebQA pools (+1.9 to +3.0, p ≤ 2e-7) and on the long-horizon average (+3.07, 95% CI [+1.64, +4.54]).

G4. Revised PDF. Observation 6 and Table 4 (controls); Appendices I–K (controls, question-level analysis with examples, significance); Table 5 and Appendix H (LLM calls, tokens, training cost); Section 4.1 (training data); corrected figures and Table F.1.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.